chore: enable simplify in ruff lint (#2085)

* fix return type

* improve apply with vertorize op

* Update src/lerobot/datasets/aggregate.py

Co-authored-by: Michel Aractingi <michel.aractingi@huggingface.co>

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -25,7 +25,7 @@ body:
     id: system-info
     attributes:
       label: System Info
-      description: If needed, you can share your lerobot configuration with us by running `python -m lerobot.scripts.display_sys_info` and copy-pasting its outputs below
+      description: Please share your LeRobot configuration by running `lerobot-info` (if installed) or `python -m lerobot.scripts.display_sys_info` (if not installed) and pasting the output below.
       render: Shell
       placeholder: lerobot version, OS, python version, numpy version, torch version, and lerobot's configuration
     validations:

.github/workflows/stale.yml

@@ -0,0 +1,68 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This workflow handles closing stale issues and PRs.
+name: Stale
+on:
+  # Allows running this workflow manually from the Actions tab
+  workflow_dispatch:
+
+  # Runs at 02:00
+  schedule:
+    - cron: "0 2 * * *"
+
+env:
+  CLOSE_ISSUE_MESSAGE: >
+    This issue was closed because it has been stalled for 14 days with no activity.
+    Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
+  CLOSE_PR_MESSAGE: >
+    This PR was closed because it has been stalled for 14 days with no activity.
+    Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
+  WARN_ISSUE_MESSAGE: >
+    This issue has been automatically marked as stale because it has not had
+    recent activity (6 months). It will be closed if no further activity occurs.
+    Thank you for your contributions.
+  WARN_PR_MESSAGE: >
+    This PR has been automatically marked as stale because it has not had
+    recent activity (6 months). It will be closed if no further activity occurs.
+    Thank you for your contributions.
+
+jobs:
+  # This job runs the actions/stale action to close stale issues and PRs.
+  stale:
+    name: Close Stale Issues and PRs
+    runs-on: ubuntu-latest
+    permissions:
+      actions: write
+      contents: write # only for delete-branch option
+      issues: write
+      pull-requests: write
+    steps:
+      - uses: actions/stale@v10
+        with:
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
+          stale-issue-label: stale
+          stale-pr-label: stale
+          exempt-issue-labels: never-stale
+          exempt-pr-labels: never-stale
+          days-before-issue-stale: 180 # TODO(Steven): Will modify this to 90 after initial cleanup
+          days-before-issue-close: 14
+          days-before-pr-stale: 180
+          days-before-pr-close: 14
+          delete-branch: true
+          close-issue-message: ${{ env.CLOSE_ISSUE_MESSAGE }}
+          close-pr-message: ${{ env.CLOSE_PR_MESSAGE }}
+          stale-issue-message: ${{ env.WARN_ISSUE_MESSAGE }}
+          stale-pr-message: ${{ env.WARN_PR_MESSAGE }}
+          operations-per-run: 500

.gitignore

@@ -173,3 +173,7 @@ outputs/
 
 # Dev folders
 .cache/*
+*.stl
+*.urdf
+*.xml
+*.part

README.md

@@ -202,7 +202,7 @@ Check out [example 1](https://github.com/huggingface/lerobot/blob/main/examples/
 You can also locally visualize episodes from a dataset on the hub by executing our script from the command line:
 
 ```bash
-python -m lerobot.scripts.visualize_dataset \
+lerobot-dataset-viz \
     --repo-id lerobot/pusht \
     --episode-index 0
 ```
@@ -210,7 +210,7 @@ python -m lerobot.scripts.visualize_dataset \
 or from a dataset in a local folder with the `root` option and the `--local-files-only` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
 
 ```bash
-python -m lerobot.scripts.visualize_dataset \
+lerobot-dataset-viz \
     --repo-id lerobot/pusht \
     --root ./my_local_data_dir \
     --local-files-only 1 \
@@ -221,19 +221,19 @@ It will open `rerun.io` and display the camera streams, robot states and actions
 
 https://github-production-user-asset-6210df.s3.amazonaws.com/4681518/328035972-fd46b787-b532-47e2-bb6f-fd536a55a7ed.mov?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20240505%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20240505T172924Z&X-Amz-Expires=300&X-Amz-Signature=d680b26c532eeaf80740f08af3320d22ad0b8a4e4da1bcc4f33142c15b509eda&X-Amz-SignedHeaders=host&actor_id=24889239&key_id=0&repo_id=748713144
 
-Our script can also visualize datasets stored on a distant server. See `python -m lerobot.scripts.visualize_dataset --help` for more instructions.
+Our script can also visualize datasets stored on a distant server. See `lerobot-dataset-viz --help` for more instructions.
 
 ### The `LeRobotDataset` format
 
 A dataset in `LeRobotDataset` format is very simple to use. It can be loaded from a repository on the Hugging Face hub or a local folder simply with e.g. `dataset = LeRobotDataset("lerobot/aloha_static_coffee")` and can be indexed into like any Hugging Face and PyTorch dataset. For instance `dataset[0]` will retrieve a single temporal frame from the dataset containing observation(s) and an action as PyTorch tensors ready to be fed to a model.
 
-A specificity of `LeRobotDataset` is that, rather than retrieving a single frame by its index, we can retrieve several frames based on their temporal relationship with the indexed frame, by setting `delta_timestamps` to a list of relative times with respect to the indexed frame. For example, with `delta_timestamps = {"observation.image": [-1, -0.5, -0.2, 0]}` one can retrieve, for a given index, 4 frames: 3 "previous" frames 1 second, 0.5 seconds, and 0.2 seconds before the indexed frame, and the indexed frame itself (corresponding to the 0 entry). See example [1_load_lerobot_dataset.py](https://github.com/huggingface/lerobot/blob/main/examples/1_load_lerobot_dataset.py) for more details on `delta_timestamps`.
+A specificity of `LeRobotDataset` is that, rather than retrieving a single frame by its index, we can retrieve several frames based on their temporal relationship with the indexed frame, by setting `delta_timestamps` to a list of relative times with respect to the indexed frame. For example, with `delta_timestamps = {"observation.image": [-1, -0.5, -0.2, 0]}` one can retrieve, for a given index, 4 frames: 3 "previous" frames 1 second, 0.5 seconds, and 0.2 seconds before the indexed frame, and the indexed frame itself (corresponding to the 0 entry). See example [1_load_lerobot_dataset.py](https://github.com/huggingface/lerobot/blob/main/examples/dataset/load_lerobot_dataset.py) for more details on `delta_timestamps`.
 
 Under the hood, the `LeRobotDataset` format makes use of several ways to serialize data which can be useful to understand if you plan to work more closely with this format. We tried to make a flexible yet simple dataset format that would cover most type of features and specificities present in reinforcement learning and robotics, in simulation and in real-world, with a focus on cameras and robot states but easily extended to other types of sensory inputs as long as they can be represented by a tensor.
 
 Here are the important details and internal structure organization of a typical `LeRobotDataset` instantiated with `dataset = LeRobotDataset("lerobot/aloha_static_coffee")`. The exact features will change from dataset to dataset but not the main aspects:
 
-````
+```
 dataset attributes:
   ├ hf_dataset: a Hugging Face dataset (backed by Arrow/parquet). Typical features example:
   │  ├ observation.images.cam_high (VideoFrame):
@@ -269,7 +269,7 @@ dataset attributes:
   ├ root (Path): local directory where the dataset is stored
   ├ image_transforms (Callable): optional image transformations to apply to visual modalities
   └ delta_timestamps (dict): optional delta timestamps for temporal queries
-decoding videos (e.g., 'pyav', 'torchcodec')
+```
 
 A `LeRobotDataset` is serialised using several widespread file formats for each of its parts, namely:
 
@@ -279,42 +279,6 @@ A `LeRobotDataset` is serialised using several widespread file formats for each
 
 Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can specify its location with the `root` argument if it's not in the default `~/.cache/huggingface/lerobot` location.
 
-### Evaluate a pretrained policy
-
-Check out [example 2](https://github.com/huggingface/lerobot/blob/main/examples/2_evaluate_pretrained_policy.py) that illustrates how to download a pretrained policy from Hugging Face hub, and run an evaluation on its corresponding environment.
-
-We also provide a more capable script to parallelize the evaluation over multiple environments during the same rollout. Here is an example with a pretrained model hosted on [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht):
-
-```bash
-lerobot-eval \
-    --policy.path=lerobot/diffusion_pusht \
-    --env.type=pusht \
-    --eval.batch_size=10 \
-    --eval.n_episodes=10 \
-    --policy.use_amp=false \
-    --policy.device=cuda
-````
-
-Note: After training your own policy, you can re-evaluate the checkpoints with:
-
-```bash
-lerobot-eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
-```
-
-See `lerobot-eval --help` for more instructions.
-
-### Train your own policy
-
-Check out [example 3](https://github.com/huggingface/lerobot/blob/main/examples/3_train_policy.py) that illustrates how to train a model using our core library in python, and [example 4](https://github.com/huggingface/lerobot/blob/main/examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
-
-To use wandb for logging training and evaluation curves, make sure you've run `wandb login` as a one-time setup step. Then, when running the training command above, enable WandB in the configuration by adding `--wandb.enable=true`.
-
-A link to the wandb logs for the run will also show up in yellow in your terminal. Here is an example of what they look like in your browser. Please also check [here](https://github.com/huggingface/lerobot/blob/main/examples/4_train_policy_with_script.md#typical-logs-and-metrics) for the explanation of some commonly used metrics in logs.
-
-\<img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/wandb.png" alt="WandB logs example"\>
-
-Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `lerobot-eval --help` for more instructions.
-
 #### Reproduce state-of-the-art (SOTA)
 
 We provide some pretrained policies on our [hub page](https://huggingface.co/lerobot) that can achieve state-of-the-art performances.
@@ -373,3 +337,7 @@ If you want, you can cite this work with:
 ## Star History
 
 [![Star History Chart](https://api.star-history.com/svg?repos=huggingface/lerobot&type=Timeline)](https://star-history.com/#huggingface/lerobot&Timeline)
+
+```
+
+```

benchmarks/video/run_video_benchmark.py

@@ -35,12 +35,13 @@ import torch
 from skimage.metrics import mean_squared_error, peak_signal_noise_ratio, structural_similarity
 from tqdm import tqdm
 
+from benchmarks.video.benchmark import TimeBenchmark
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.video_utils import (
     decode_video_frames_torchvision,
     encode_video_frames,
 )
-from lerobot.utils.benchmark import TimeBenchmark
+from lerobot.utils.constants import OBS_IMAGE
 
 BASE_ENCODING = OrderedDict(
     [
@@ -117,7 +118,7 @@ def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
     hf_dataset = dataset.hf_dataset.with_format(None)
 
     # We only save images from the first camera
-    img_keys = [key for key in hf_dataset.features if key.startswith("observation.image")]
+    img_keys = [key for key in hf_dataset.features if key.startswith(OBS_IMAGE)]
     imgs_dataset = hf_dataset.select_columns(img_keys[0])
 
     for i, item in enumerate(

docker/Dockerfile.internal

@@ -39,6 +39,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     software-properties-common build-essential git curl \
     libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
     libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
+    cmake pkg-config ninja-build \
     && add-apt-repository -y ppa:deadsnakes/ppa \
     && apt-get update \
     && apt-get install -y --no-install-recommends \

docker/Dockerfile.user

@@ -31,6 +31,7 @@ ENV DEBIAN_FRONTEND=noninteractive \
 RUN apt-get update && apt-get install -y --no-install-recommends \
     build-essential git curl libglib2.0-0 libegl1-mesa-dev ffmpeg \
     libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
+    cmake pkg-config ninja-build \
     && curl -LsSf https://astral.sh/uv/install.sh | sh \
     && mv /root/.local/bin/uv /usr/local/bin/uv \
     && useradd --create-home --shell /bin/bash user_lerobot \

docs/source/_toctree.yml

@@ -29,10 +29,21 @@
 - sections:
   - local: smolvla
     title: Finetune SmolVLA
+  - local: libero
+    title: Using Libero
   title: "Policies"
+
+- sections:
+  - local: introduction_processors
+    title: Introduction to Robot Processors
+  - local: debug_processor_pipeline
+    title: Debug your processor pipeline
+  - local: implement_your_own_processor
+    title: Implement your own processor
+  - local: processors_robots_teleop
+    title: Processors for Robots and Teleoperators
+  title: "Robot Processors"
 - sections:
-  - local: hope_jr
-    title: Hope Jr
   - local: so101
     title: SO-101
   - local: so100
@@ -41,9 +52,15 @@
     title: Koch v1.1
   - local: lekiwi
     title: LeKiwi
+  - local: hope_jr
+    title: Hope Jr
   - local: reachy2
     title: Reachy 2
   title: "Robots"
+- sections:
+  - local: phone_teleop
+    title: Phone
+  title: "Teleoperators"
 - sections:
   - local: notebooks
     title: Notebooks

docs/source/async.mdx

@@ -31,7 +31,7 @@ Then, spin up a policy server (in one terminal, or in a separate machine) specif
 You can spin up a policy server running:
 
 ```shell
-python src/lerobot/scripts/server/policy_server.py \
+python src/lerobot/async_inference/policy_server.py \
     --host=127.0.0.1 \
     --port=8080 \
 ```
@@ -39,7 +39,7 @@ python src/lerobot/scripts/server/policy_server.py \
 This will start a policy server listening on `127.0.0.1:8080` (`localhost`, port 8080). At this stage, the policy server is empty, as all information related to which policy to run and with which parameters are specified during the first handshake with the client. Spin up a client with:
 
 ```shell
-python src/lerobot/scripts/server/robot_client.py \
+python src/lerobot/async_inference/robot_client.py \
     --server_address=127.0.0.1:8080 \ # SERVER: the host address and port of the policy server
     --robot.type=so100_follower \ # ROBOT: your robot type
     --robot.port=/dev/tty.usbmodem585A0076841 \ # ROBOT: your robot port
@@ -122,8 +122,8 @@ python -m lerobot.scripts.server.policy_server \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.scripts.server.configs import PolicyServerConfig
-from lerobot.scripts.server.policy_server import serve
+from lerobot.async_inference.configs import PolicyServerConfig
+from lerobot.async_inference.policy_server import serve
 
 config = PolicyServerConfig(
     host="localhost",
@@ -148,7 +148,7 @@ The `RobotClient` streams observations to the `PolicyServer`, and receives actio
 <hfoptions id="start_robot_client">
 <hfoption id="Command">
 ```bash
-python src/lerobot/scripts/server/robot_client.py \
+python src/lerobot/async_inference/robot_client.py \
     --server_address=127.0.0.1:8080 \ # SERVER: the host address and port of the policy server
     --robot.type=so100_follower \ # ROBOT: your robot type
     --robot.port=/dev/tty.usbmodem585A0076841 \ # ROBOT: your robot port
@@ -171,9 +171,9 @@ python src/lerobot/scripts/server/robot_client.py \
 import threading
 from lerobot.robots.so100_follower import SO100FollowerConfig
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.scripts.server.configs import RobotClientConfig
-from lerobot.scripts.server.robot_client import RobotClient
-from lerobot.scripts.server.helpers import visualize_action_queue_size
+from lerobot.async_inference.configs import RobotClientConfig
+from lerobot.async_inference.robot_client import RobotClient
+from lerobot.async_inference.helpers import visualize_action_queue_size
 
 # 1. Create the robot instance
 """Check out the cameras available in your setup by running `python lerobot/find_cameras.py`"""

docs/source/backwardcomp.mdx

@@ -1,5 +1,61 @@
 # Backward compatibility
 
+## Policy Normalization Migration (PR #1452)
+
+**Breaking Change**: LeRobot policies no longer have built-in normalization layers embedded in their weights. Normalization is now handled by external `PolicyProcessorPipeline` components.
+
+### What changed?
+
+|                            | Before PR #1452                                  | After PR #1452                                               |
+| -------------------------- | ------------------------------------------------ | ------------------------------------------------------------ |
+| **Normalization Location** | Embedded in model weights (`normalize_inputs.*`) | External `PolicyProcessorPipeline` components                |
+| **Model State Dict**       | Contains normalization statistics                | **Clean weights only** - no normalization parameters         |
+| **Usage**                  | `policy(batch)` handles everything               | `preprocessor(batch)` → `policy(...)` → `postprocessor(...)` |
+
+### Impact on existing models
+
+- Models trained **before** PR #1452 have normalization embedded in their weights
+- These models need migration to work with the new `PolicyProcessorPipeline` system
+- The migration extracts normalization statistics and creates separate processor pipelines
+
+### Migrating old models
+
+Use the migration script to convert models with embedded normalization:
+
+```shell
+python src/lerobot/processor/migrate_policy_normalization.py \
+    --pretrained-path lerobot/act_aloha_sim_transfer_cube_human \
+    --push-to-hub \
+    --branch migrated
+```
+
+The script:
+
+1. **Extracts** normalization statistics from model weights
+2. **Creates** external preprocessor and postprocessor pipelines
+3. **Removes** normalization layers from model weights
+4. **Saves** clean model + processor pipelines
+5. **Pushes** to Hub with automatic PR creation
+
+### Using migrated models
+
+```python
+# New usage pattern (after migration)
+from lerobot.policies.factory import make_policy, make_pre_post_processors
+
+# Load model and processors separately
+policy = make_policy(config, ds_meta=dataset.meta)
+preprocessor, postprocessor = make_pre_post_processors(
+    policy_cfg=config,
+    dataset_stats=dataset.meta.stats
+)
+
+# Process data through pipeline
+processed_batch = preprocessor(raw_batch)
+action = policy.select_action(processed_batch)
+final_action = postprocessor(action)
+```
+
 ## Hardware API redesign
 
 PR [#777](https://github.com/huggingface/lerobot/pull/777) improves the LeRobot calibration but is **not backward-compatible**. Below is a overview of what changed and how you can continue to work with datasets created before this pull request.

docs/source/debug_processor_pipeline.mdx

@@ -0,0 +1,299 @@
+# Debug Your Processor Pipeline
+
+Processor pipelines can be complex, especially when chaining multiple transformation steps.
+Unlike simple function calls, pipelines lack natural observability, you can't easily see what happens
+between each step or where things go wrong.
+This guide provides debugging tools and techniques specifically designed to address these challenges
+and help you understand data flow through your pipelines.
+
+We'll explore three complementary debugging approaches: **hooks** for runtime monitoring, **step-through debugging** for detailed inspection, and **feature validation** for catching structural mismatches. Each serves a different purpose and together they provide complete visibility into your pipeline's behavior.
+
+## Understanding Hooks
+
+Hooks are functions that get called at specific points during pipeline execution.
+They provide a way to inspect, monitor, or modify data without changing your pipeline code.
+Think of them as "event listeners" for your pipeline.
+
+### What is a Hook?
+
+A hook is a callback function that gets automatically invoked at specific moments during pipeline execution.
+The concept comes from event-driven programming, imagine you could "hook into" the pipeline's execution flow to observe or react to what's happening.
+
+Think of hooks like inserting checkpoints into your pipeline. Every time the pipeline reaches one of these checkpoints, it pauses briefly to call your hook function, giving you a chance to inspect the current state, log information, and validate data.
+
+A hook is simply a function that accepts two parameters:
+
+- `step_idx: int` - The index of the current processing step (0, 1, 2, etc.)
+- `transition: EnvTransition` - The data transition at that point in the pipeline
+
+The beauty of hooks is their non-invasive nature: you can add monitoring, validation, or debugging logic without changing a single line of your pipeline code. The pipeline remains clean and focused on its core logic, while hooks handle the cross-cutting concerns like logging, monitoring, and debugging.
+
+### Before vs After Hooks
+
+The pipeline supports two types of hooks:
+
+- **Before hooks** (`register_before_step_hook`) - Called before each step executes
+- **After hooks** (`register_after_step_hook`) - Called after each step completes
+
+```python
+def before_hook(step_idx: int, transition: EnvTransition):
+    """Called before step processes the transition."""
+    print(f"About to execute step {step_idx}")
+    # Useful for: logging, validation, setup
+
+def after_hook(step_idx: int, transition: EnvTransition):
+    """Called after step has processed the transition."""
+    print(f"Completed step {step_idx}")
+    # Useful for: monitoring results, cleanup, debugging
+
+processor.register_before_step_hook(before_hook)
+processor.register_after_step_hook(after_hook)
+```
+
+### Implementing a NaN Detection Hook
+
+Here's a practical example of a hook that detects NaN values:
+
+```python
+def check_nans(step_idx: int, transition: EnvTransition):
+    """Check for NaN values in observations."""
+    obs = transition.get(TransitionKey.OBSERVATION)
+    if obs:
+        for key, value in obs.items():
+            if isinstance(value, torch.Tensor) and torch.isnan(value).any():
+                print(f"NaN detected in {key} at step {step_idx}")
+
+# Register the hook to run after each step
+processor.register_after_step_hook(check_nans)
+
+# Process your data - the hook will be called automatically
+output = processor(input_data)
+
+# Remove the hook when done debugging
+processor.unregister_after_step_hook(check_nans)
+```
+
+### How Hooks Work Internally
+
+Understanding the internal mechanism helps you use hooks more effectively. The pipeline maintains two separate lists: one for before-step hooks and another for after-step hooks. When you register a hook, it's simply appended to the appropriate list.
+
+During execution, the pipeline follows a strict sequence: for each processing step, it first calls all before-hooks in registration order, then executes the actual step transformation, and finally calls all after-hooks in registration order. This creates a predictable, sandwich-like structure around each step.
+
+The key insight is that hooks don't change the core pipeline logic—they're purely additive. The pipeline's `_forward` method orchestrates this dance between hooks and processing steps, ensuring that your debugging or monitoring code runs at exactly the right moments without interfering with the main data flow.
+
+Here's a simplified view of how the pipeline executes hooks:
+
+```python
+class DataProcessorPipeline:
+    def __init__(self):
+        self.steps = [...]
+        self.before_step_hooks = []  # List of before hooks
+        self.after_step_hooks = []   # List of after hooks
+
+    def _forward(self, transition):
+        """Internal method that processes the transition through all steps."""
+        for step_idx, processor_step in enumerate(self.steps):
+            # 1. Call all BEFORE hooks
+            for hook in self.before_step_hooks:
+                hook(step_idx, transition)
+
+            # 2. Execute the actual processing step
+            transition = processor_step(transition)
+
+            # 3. Call all AFTER hooks
+            for hook in self.after_step_hooks:
+                hook(step_idx, transition)
+
+        return transition
+
+    def register_before_step_hook(self, hook_fn):
+        self.before_step_hooks.append(hook_fn)
+
+    def register_after_step_hook(self, hook_fn):
+        self.after_step_hooks.append(hook_fn)
+```
+
+### Execution Flow
+
+The execution flow looks like this:
+
+```
+Input → Before Hook → Step 0 → After Hook → Before Hook → Step 1 → After Hook → ... → Output
+```
+
+For example, with 3 steps and both hook types:
+
+```python
+def timing_before(step_idx, transition):
+    print(f"⏱️  Starting step {step_idx}")
+
+def validation_after(step_idx, transition):
+    print(f"✅ Completed step {step_idx}")
+
+processor.register_before_step_hook(timing_before)
+processor.register_after_step_hook(validation_after)
+
+# This will output:
+# ⏱️  Starting step 0
+# ✅ Completed step 0
+# ⏱️  Starting step 1
+# ✅ Completed step 1
+# ⏱️  Starting step 2
+# ✅ Completed step 2
+```
+
+### Multiple Hooks
+
+You can register multiple hooks of the same type - they execute in the order registered:
+
+```python
+def log_shapes(step_idx: int, transition: EnvTransition):
+    obs = transition.get(TransitionKey.OBSERVATION)
+    if obs:
+        print(f"Step {step_idx} observation shapes:")
+        for key, value in obs.items():
+            if isinstance(value, torch.Tensor):
+                print(f"  {key}: {value.shape}")
+
+processor.register_after_step_hook(check_nans)      # Executes first
+processor.register_after_step_hook(log_shapes)     # Executes second
+
+# Both hooks will be called after each step in registration order
+output = processor(input_data)
+```
+
+While hooks are excellent for monitoring specific issues (like NaN detection) or gathering metrics during normal pipeline execution, sometimes you need to dive deeper. When you want to understand exactly what happens at each step or debug complex transformation logic, step-through debugging provides the detailed inspection you need.
+
+## Step-Through Debugging
+
+Step-through debugging is like having a slow-motion replay for your pipeline. Instead of watching your data get transformed in one quick blur from input to output, you can pause and examine what happens after each individual step.
+
+This approach is particularly valuable when you're trying to understand a complex pipeline, debug unexpected behavior, or verify that each transformation is working as expected. Unlike hooks, which are great for automated monitoring, step-through debugging gives you manual, interactive control over the inspection process.
+
+The `step_through()` method is a generator that yields the transition state after each processing step, allowing you to inspect intermediate results. Think of it as creating a series of snapshots of your data as it flows through the pipeline—each snapshot shows you exactly what your data looks like after one more transformation has been applied.
+
+### How Step-Through Works
+
+The `step_through()` method fundamentally changes how the pipeline executes. Instead of running all steps in sequence and only returning the final result, it transforms the pipeline into an iterator that yields intermediate results.
+
+Here's what happens internally: the method starts by converting your input data into the pipeline's internal transition format, then yields this initial state. Next, it applies the first processing step and yields the result. Then it applies the second step to that result and yields again, and so on. Each `yield` gives you a complete snapshot of the transition at that point.
+
+This generator pattern is powerful because it's lazy—the pipeline only computes the next step when you ask for it. This means you can stop at any point, inspect the current state thoroughly, and decide whether to continue. You're not forced to run the entire pipeline just to debug one problematic step.
+
+Instead of running the entire pipeline and only seeing the final result, `step_through()` pauses after each step and gives you the intermediate transition:
+
+```python
+# This creates a generator that yields intermediate states
+for i, intermediate_result in enumerate(processor.step_through(input_data)):
+    print(f"=== After step {i} ===")
+
+    # Inspect the observation at this stage
+    obs = intermediate_result.get(TransitionKey.OBSERVATION)
+    if obs:
+        for key, value in obs.items():
+            if isinstance(value, torch.Tensor):
+                print(f"{key}: shape={value.shape}, dtype={value.dtype}")
+```
+
+### Interactive Debugging with Breakpoints
+
+You can add breakpoints in the step-through loop to interactively debug:
+
+```python
+# Step through the pipeline with debugging
+for i, intermediate in enumerate(processor.step_through(data)):
+    print(f"Step {i}: {processor.steps[i].__class__.__name__}")
+
+    # Set a breakpoint to inspect the current state
+    breakpoint()  # Debugger will pause here
+
+    # You can now inspect 'intermediate' in the debugger:
+    # - Check tensor shapes and values
+    # - Verify expected transformations
+    # - Look for unexpected changes
+```
+
+During the debugger session, you can:
+
+- Examine `intermediate[TransitionKey.OBSERVATION]` to see observation data
+- Check `intermediate[TransitionKey.ACTION]` for action transformations
+- Inspect any part of the transition to understand what each step does
+
+Step-through debugging is perfect for understanding the _data_ transformations, but what about the _structure_ of that data? While hooks and step-through help you debug runtime behavior, you also need to ensure your pipeline produces data in the format expected by downstream components. This is where feature contract validation comes in.
+
+## Validating Feature Contracts
+
+Feature contracts define what data structure your pipeline expects as input and produces as output.
+Validating these contracts helps catch mismatches early.
+
+### Understanding Feature Contracts
+
+Each processor step has a `transform_features()` method that describes how it changes the data structure:
+
+```python
+# Get the expected output features from your pipeline
+initial_features = {
+    PipelineFeatureType.OBSERVATION: {
+        "observation.state": PolicyFeature(type=FeatureType.STATE, shape=(7,)),
+        "observation.image": PolicyFeature(type=FeatureType.IMAGE, shape=(3, 224, 224))
+    },
+    PipelineFeatureType.ACTION: {
+        "action": PolicyFeature(type=FeatureType.ACTION, shape=(4,))
+    }
+}
+
+# Check what your pipeline will output
+output_features = processor.transform_features(initial_features)
+
+print("Input features:")
+for feature_type, features in initial_features.items():
+    print(f"  {feature_type}:")
+    for key, feature in features.items():
+        print(f"    {key}: {feature.type.value}, shape={feature.shape}")
+
+print("\nOutput features:")
+for feature_type, features in output_features.items():
+    print(f"  {feature_type}:")
+    for key, feature in features.items():
+        print(f"    {key}: {feature.type.value}, shape={feature.shape}")
+```
+
+### Verifying Expected Features
+
+Check that your pipeline produces the features you expect:
+
+```python
+# Define what features you expect the pipeline to produce
+expected_keys = ["observation.state", "observation.image", "action"]
+
+print("Validating feature contract...")
+for expected_key in expected_keys:
+    found = False
+    for feature_type, features in output_features.items():
+        if expected_key in features:
+            feature = features[expected_key]
+            print(f"✅ {expected_key}: {feature.type.value}, shape={feature.shape}")
+            found = True
+            break
+
+    if not found:
+        print(f"❌ Missing expected feature: {expected_key}")
+```
+
+This validation helps ensure your pipeline will work correctly with downstream components that expect specific data structures.
+
+## Summary
+
+Now that you understand the three debugging approaches, you can tackle any pipeline issue systematically:
+
+1. **Hooks** - For runtime monitoring and validation without modifying pipeline code
+2. **Step-through** - For inspecting intermediate states and understanding transformations
+3. **Feature validation** - For ensuring data structure contracts are met
+
+**When to use each approach:**
+
+- Start with **step-through debugging** when you need to understand what your pipeline does or when something unexpected happens
+- Add **hooks** for continuous monitoring during development and production to catch issues automatically
+- Use **feature validation** before deployment to ensure your pipeline works with downstream components
+
+These three tools work together to give you the complete observability that complex pipelines naturally lack. With hooks watching for issues, step-through helping you understand behavior, and feature validation ensuring compatibility, you'll be able to debug any pipeline confidently and efficiently.

docs/source/hilserl.mdx

@@ -4,7 +4,13 @@ In this tutorial you will go through the full Human-in-the-Loop Sample-Efficient
 
 HIL-SERL is a sample-efficient reinforcement learning algorithm that combines human demonstrations with online learning and human interventions. The approach starts from a small set of human demonstrations, uses them to train a reward classifier, and then employs an actor-learner architecture where humans can intervene during policy execution to guide exploration and correct unsafe behaviors. In this tutorial, you'll use a gamepad to provide interventions and control the robot during the learning process.
 
-It combines three key ingredients: 1. **Offline demonstrations & reward classifier:** a handful of human-teleop episodes plus a vision-based success detector give the policy a shaped starting point. 2. **On-robot actor / learner loop with human interventions:** a distributed Soft Actor Critic (SAC) learner updates the policy while an actor explores on the physical robot; the human can jump in at any time to correct dangerous or unproductive behaviour. 3. **Safety & efficiency tools:** joint/end-effector (EE) bounds, crop region of interest (ROI) preprocessing and WandB monitoring keep the data useful and the hardware safe.
+It combines three key ingredients:
+
+1. **Offline demonstrations & reward classifier:** a handful of human-teleop episodes plus a vision-based success detector give the policy a shaped starting point.
+
+2. **On-robot actor / learner loop with human interventions:** a distributed Soft Actor Critic (SAC) learner updates the policy while an actor explores on the physical robot; the human can jump in at any time to correct dangerous or unproductive behaviour.
+
+3. **Safety & efficiency tools:** joint/end-effector (EE) bounds, crop region of interest (ROI) preprocessing and WandB monitoring keep the data useful and the hardware safe.
 
 Together these elements let HIL-SERL reach near-perfect task success and faster cycle times than imitation-only baselines.
 
@@ -56,49 +62,258 @@ pip install -e ".[hilserl]"
 
 ### Understanding Configuration
 
-The training process begins with proper configuration for the HILSerl environment. The configuration class of interest is `HILSerlRobotEnvConfig` in `lerobot/envs/configs.py`. Which is defined as:
+The training process begins with proper configuration for the HILSerl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
 
 <!-- prettier-ignore-start -->
 ```python
+class GymManipulatorConfig:
+    env: HILSerlRobotEnvConfig    # Environment configuration (nested)
+    dataset: DatasetConfig    # Dataset recording/replay configuration (nested)
+    mode: str | None = None    # "record", "replay", or None (for training)
+    device: str = "cpu"    # Compute device
+
 class HILSerlRobotEnvConfig(EnvConfig):
     robot: RobotConfig | None = None    # Main robot agent (defined in `lerobot/robots`)
-    teleop: TeleoperatorConfig | None = None    # Teleoperator agent, e.g., gamepad or leader arm, (defined in `lerobot/teleoperators`)
-    wrapper: EnvTransformConfig | None = None    # Environment wrapper settings; check `lerobot/scripts/server/gym_manipulator.py`
-    fps: int = 10    # Control frequency
+    teleop: TeleoperatorConfig | None = None    # Teleoperator agent, e.g., gamepad or leader arm
+    processor: HILSerlProcessorConfig    # Processing pipeline configuration (nested)
     name: str = "real_robot"    # Environment name
-    mode: str = None    # "record", "replay", or None (for training)
-    repo_id: str | None = None    # LeRobot dataset repository ID
-    dataset_root: str | None = None    # Local dataset root (optional)
-    task: str = ""    # Task identifier
-    num_episodes: int = 10    # Number of episodes for recording
-    episode: int = 0    # episode index for replay
-    device: str = "cuda"    # Compute device
-    push_to_hub: bool = True    # Whether to push the recorded datasets to Hub
-    pretrained_policy_name_or_path: str | None = None    # For policy loading
-    reward_classifier_pretrained_path: str | None = None    # For reward model
-    number_of_steps_after_success: int = 0    # For reward classifier, collect more positive examples after a success to train a classifier
+    task: str | None = None    # Task identifier
+    fps: int = 10    # Control frequency
+
+# Nested processor configuration
+class HILSerlProcessorConfig:
+    control_mode: str = "gamepad"    # Control mode
+    observation: ObservationConfig | None = None    # Observation processing settings
+    image_preprocessing: ImagePreprocessingConfig | None = None    # Image crop/resize settings
+    gripper: GripperConfig | None = None    # Gripper control and penalty settings
+    reset: ResetConfig | None = None    # Environment reset and timing settings
+    inverse_kinematics: InverseKinematicsConfig | None = None    # IK processing settings
+    reward_classifier: RewardClassifierConfig | None = None    # Reward classifier settings
+    max_gripper_pos: float | None = 100.0    # Maximum gripper position
+
+# Sub-configuration classes
+class ObservationConfig:
+    add_joint_velocity_to_observation: bool = False    # Add joint velocities to state
+    add_current_to_observation: bool = False    # Add motor currents to state
+    display_cameras: bool = False    # Display camera feeds during execution
+
+class ImagePreprocessingConfig:
+    crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None    # Image cropping parameters
+    resize_size: tuple[int, int] | None = None    # Target image size
+
+class GripperConfig:
+    use_gripper: bool = True    # Enable gripper control
+    gripper_penalty: float = 0.0    # Penalty for inappropriate gripper usage
+
+class ResetConfig:
+    fixed_reset_joint_positions: Any | None = None    # Joint positions for reset
+    reset_time_s: float = 5.0    # Time to wait during reset
+    control_time_s: float = 20.0    # Maximum episode duration
+    terminate_on_success: bool = True    # Whether to terminate episodes on success detection
+
+class InverseKinematicsConfig:
+    urdf_path: str | None = None    # Path to robot URDF file
+    target_frame_name: str | None = None    # End-effector frame name
+    end_effector_bounds: dict[str, list[float]] | None = None    # EE workspace bounds
+    end_effector_step_sizes: dict[str, float] | None = None    # EE step sizes per axis
+
+class RewardClassifierConfig:
+    pretrained_path: str | None = None    # Path to pretrained reward classifier
+    success_threshold: float = 0.5    # Success detection threshold
+    success_reward: float = 1.0    # Reward value for successful episodes
+
+# Dataset configuration
+class DatasetConfig:
+    repo_id: str    # LeRobot dataset repository ID
+    task: str    # Task identifier
+    root: str | None = None    # Local dataset root directory
+    num_episodes_to_record: int = 5    # Number of episodes for recording
+    replay_episode: int | None = None    # Episode index for replay
+    push_to_hub: bool = False    # Whether to push datasets to Hub
 ```
 <!-- prettier-ignore-end -->
 
+### Processor Pipeline Architecture
+
+HIL-SERL uses a modular processor pipeline architecture that processes robot observations and actions through a series of composable steps. The pipeline is divided into two main components:
+
+#### Environment Processor Pipeline
+
+The environment processor (`env_processor`) handles incoming observations and environment state:
+
+1. **VanillaObservationProcessorStep**: Converts raw robot observations into standardized format
+2. **JointVelocityProcessorStep** (optional): Adds joint velocity information to observations
+3. **MotorCurrentProcessorStep** (optional): Adds motor current readings to observations
+4. **ForwardKinematicsJointsToEE** (optional): Computes end-effector pose from joint positions
+5. **ImageCropResizeProcessorStep** (optional): Crops and resizes camera images
+6. **TimeLimitProcessorStep** (optional): Enforces episode time limits
+7. **GripperPenaltyProcessorStep** (optional): Applies penalties for inappropriate gripper usage
+8. **RewardClassifierProcessorStep** (optional): Automated reward detection using vision models
+9. **AddBatchDimensionProcessorStep**: Converts data to batch format for neural network processing
+10. **DeviceProcessorStep**: Moves data to the specified compute device (CPU/GPU)
+
+#### Action Processor Pipeline
+
+The action processor (`action_processor`) handles outgoing actions and human interventions:
+
+1. **AddTeleopActionAsComplimentaryDataStep**: Captures teleoperator actions for logging
+2. **AddTeleopEventsAsInfoStep**: Records intervention events and episode control signals
+3. **InterventionActionProcessorStep**: Handles human interventions and episode termination
+4. **Inverse Kinematics Pipeline** (when enabled):
+   - **MapDeltaActionToRobotActionStep**: Converts delta actions to robot action format
+   - **EEReferenceAndDelta**: Computes end-effector reference and delta movements
+   - **EEBoundsAndSafety**: Enforces workspace safety bounds
+   - **InverseKinematicsEEToJoints**: Converts end-effector actions to joint targets
+   - **GripperVelocityToJoint**: Handles gripper control commands
+
+#### Configuration Examples
+
+**Basic Observation Processing**:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_joint_velocity_to_observation": true,
+        "add_current_to_observation": false,
+        "display_cameras": false
+      }
+    }
+  }
+}
+```
+
+**Image Processing**:
+
+```json
+{
+  "env": {
+    "processor": {
+      "image_preprocessing": {
+        "crop_params_dict": {
+          "observation.images.front": [180, 250, 120, 150],
+          "observation.images.side": [180, 207, 180, 200]
+        },
+        "resize_size": [128, 128]
+      }
+    }
+  }
+}
+```
+
+**Inverse Kinematics Setup**:
+
+```json
+{
+  "env": {
+    "processor": {
+      "inverse_kinematics": {
+        "urdf_path": "path/to/robot.urdf",
+        "target_frame_name": "end_effector",
+        "end_effector_bounds": {
+          "min": [0.16, -0.08, 0.03],
+          "max": [0.24, 0.2, 0.1]
+        },
+        "end_effector_step_sizes": {
+          "x": 0.02,
+          "y": 0.02,
+          "z": 0.02
+        }
+      }
+    }
+  }
+}
+```
+
+### Advanced Observation Processing
+
+The HIL-SERL framework supports additional observation processing features that can improve policy learning:
+
+#### Joint Velocity Processing
+
+Enable joint velocity estimation to provide the policy with motion information:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_joint_velocity_to_observation": true
+      }
+    }
+  }
+}
+```
+
+This processor:
+
+- Estimates joint velocities using finite differences between consecutive joint position readings
+- Adds velocity information to the observation state vector
+- Useful for policies that need motion awareness for dynamic tasks
+
+#### Motor Current Processing
+
+Monitor motor currents to detect contact forces and load conditions:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_current_to_observation": true
+      }
+    }
+  }
+}
+```
+
+This processor:
+
+- Reads motor current values from the robot's control system
+- Adds current measurements to the observation state vector
+- Helps detect contact events, object weights, and mechanical resistance
+- Useful for contact-rich manipulation tasks
+
+#### Combined Observation Processing
+
+You can enable multiple observation processing features simultaneously:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_joint_velocity_to_observation": true,
+        "add_current_to_observation": true,
+        "display_cameras": false
+      }
+    }
+  }
+}
+```
+
+**Note**: Enabling additional observation features increases the state space dimensionality, which may require adjusting your policy network architecture and potentially collecting more training data.
+
 ### Finding Robot Workspace Bounds
 
 Before collecting demonstrations, you need to determine the appropriate operational bounds for your robot.
 
 This helps simplify the problem of learning on the real robot in two ways: 1) by limiting the robot's operational space to a specific region that solves the task and avoids unnecessary or unsafe exploration, and 2) by allowing training in end-effector space rather than joint space. Empirically, learning in joint space for reinforcement learning in manipulation is often a harder problem - some tasks are nearly impossible to learn in joint space but become learnable when the action space is transformed to end-effector coordinates.
 
-**Using find_joint_limits.py**
+**Using lerobot-find-joint-limits**
 
 This script helps you find the safe operational bounds for your robot's end-effector. Given that you have a follower and leader arm, you can use the script to find the bounds for the follower arm that will be applied during training.
 Bounding the action space will reduce the redundant exploration of the agent and guarantees safety.
 
 ```bash
-python -m lerobot.scripts.find_joint_limits \
-    --robot.type=so100_follower \
-    --robot.port=/dev/tty.usbmodem58760431541 \
-    --robot.id=black \
-    --teleop.type=so100_leader \
-    --teleop.port=/dev/tty.usbmodem58760431551 \
-    --teleop.id=blue
+lerobot-find-joint-limits \
+  --robot.type=so100_follower \
+  --robot.port=/dev/tty.usbmodem58760431541 \
+  --robot.id=black \
+  --teleop.type=so100_leader \
+  --teleop.port=/dev/tty.usbmodem58760431551 \
+  --teleop.id=blue
 ```
 
 **Workflow**
@@ -128,24 +343,58 @@ With the bounds defined, you can safely collect demonstrations for training. Tra
 
 **Setting Up Record Mode**
 
-Create a configuration file for recording demonstrations (or edit an existing one like [env_config_so100.json](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_so100.json)):
+Create a configuration file for recording demonstrations (or edit an existing one like [env_config.json](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/env_config.json)):
 
-1. Set `mode` to `"record"`
-2. Specify a unique `repo_id` for your dataset (e.g., "username/task_name")
-3. Set `num_episodes` to the number of demonstrations you want to collect
-4. Set `crop_params_dict` to `null` initially (we'll determine crops later)
-5. Configure `robot`, `cameras`, and other hardware settings
+1. Set `mode` to `"record"` at the root level
+2. Specify a unique `repo_id` for your dataset in the `dataset` section (e.g., "username/task_name")
+3. Set `num_episodes_to_record` in the `dataset` section to the number of demonstrations you want to collect
+4. Set `env.processor.image_preprocessing.crop_params_dict` to `{}` initially (we'll determine crops later)
+5. Configure `env.robot`, `env.teleop`, and other hardware settings in the `env` section
 
 Example configuration section:
 
 ```json
-"mode": "record",
-"repo_id": "username/pick_lift_cube",
-"dataset_root": null,
-"task": "pick_and_lift",
-"num_episodes": 15,
-"episode": 0,
-"push_to_hub": true
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "real_robot",
+    "fps": 10,
+    "processor": {
+      "control_mode": "gamepad",
+      "observation": {
+        "display_cameras": false
+      },
+      "image_preprocessing": {
+        "crop_params_dict": {},
+        "resize_size": [128, 128]
+      },
+      "gripper": {
+        "use_gripper": true,
+        "gripper_penalty": 0.0
+      },
+      "reset": {
+        "reset_time_s": 5.0,
+        "control_time_s": 20.0
+      }
+    },
+    "robot": {
+      // ... robot configuration ...
+    },
+    "teleop": {
+      // ... teleoperator configuration ...
+    }
+  },
+  "dataset": {
+    "repo_id": "username/pick_lift_cube",
+    "root": null,
+    "task": "pick_and_lift",
+    "num_episodes_to_record": 15,
+    "replay_episode": 0,
+    "push_to_hub": true
+  },
+  "mode": "record",
+  "device": "cpu"
+}
 ```
 
 ### Using a Teleoperation Device
@@ -191,10 +440,20 @@ The gamepad provides a very convenient way to control the robot and the episode
 To setup the gamepad, you need to set the `control_mode` to `"gamepad"` and define the `teleop` section in the configuration file.
 
 ```json
+{
+  "env": {
     "teleop": {
-        "type": "gamepad",
-        "use_gripper": true
+      "type": "gamepad",
+      "use_gripper": true
     },
+    "processor": {
+      "control_mode": "gamepad",
+      "gripper": {
+        "use_gripper": true
+      }
+    }
+  }
+}
 ```
 
 <p align="center">
@@ -216,11 +475,21 @@ The SO101 leader arm has reduced gears that allows it to move and track the foll
 To setup the SO101 leader, you need to set the `control_mode` to `"leader"` and define the `teleop` section in the configuration file.
 
 ```json
+{
+  "env": {
     "teleop": {
-        "type": "so101_leader",
-        "port": "/dev/tty.usbmodem585A0077921", # check your port number
-        "use_degrees": true
+      "type": "so101_leader",
+      "port": "/dev/tty.usbmodem585A0077921",
+      "use_degrees": true
     },
+    "processor": {
+      "control_mode": "leader",
+      "gripper": {
+        "use_gripper": true
+      }
+    }
+  }
+}
 ```
 
 In order to annotate the success/failure of the episode, **you will need** to use a keyboard to press `s` for success, `esc` for failure.
@@ -246,12 +515,12 @@ During the online training, press `space` to take over the policy and `space` ag
 Start the recording process, an example of the config file can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_so100.json):
 
 ```bash
-python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/env_config_so100.json
+python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config_so100.json
 ```
 
 During recording:
 
-1. The robot will reset to the initial position defined in the configuration file `fixed_reset_joint_positions`
+1. The robot will reset to the initial position defined in the configuration file `env.processor.reset.fixed_reset_joint_positions`
 2. Complete the task successfully
 3. The episode ends with a reward of 1 when you press the "success" button
 4. If the time limit is reached, or the fail button is pressed, the episode ends with a reward of 0
@@ -277,7 +546,7 @@ Note: If you already know the crop parameters, you can skip this step and just s
 Use the `crop_dataset_roi.py` script to interactively select regions of interest in your camera images:
 
 ```bash
-python -m lerobot.scripts.rl.crop_dataset_roi --repo-id username/pick_lift_cube
+python -m lerobot.rl.crop_dataset_roi --repo-id username/pick_lift_cube
 ```
 
 1. For each camera view, the script will display the first frame
@@ -310,11 +579,19 @@ observation.images.front: [180, 250, 120, 150]
 Add these crop parameters to your training configuration:
 
 ```json
-"crop_params_dict": {
-    "observation.images.side": [180, 207, 180, 200],
-    "observation.images.front": [180, 250, 120, 150]
-},
-"resize_size": [128, 128]
+{
+  "env": {
+    "processor": {
+      "image_preprocessing": {
+        "crop_params_dict": {
+          "observation.images.side": [180, 207, 180, 200],
+          "observation.images.front": [180, 250, 120, 150]
+        },
+        "resize_size": [128, 128]
+      }
+    }
+  }
+}
 ```
 
 **Recommended image resolution**
@@ -338,31 +615,57 @@ Before training, you need to collect a dataset with labeled examples. The `recor
 To collect a dataset, you need to modify some parameters in the environment configuration based on HILSerlRobotEnvConfig.
 
 ```bash
-python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/reward_classifier_train_config.json
+python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/reward_classifier_train_config.json
 ```
 
 **Key Parameters for Data Collection**
 
-- **mode**: set it to `"record"` to collect a dataset
-- **repo_id**: `"hf_username/dataset_name"`, name of the dataset and repo on the hub
-- **num_episodes**: Number of episodes to record
-- **number_of_steps_after_success**: Number of additional frames to record after a success (reward=1) is detected
-- **fps**: Number of frames per second to record
-- **push_to_hub**: Whether to push the dataset to the hub
+- **mode**: set it to `"record"` to collect a dataset (at root level)
+- **dataset.repo_id**: `"hf_username/dataset_name"`, name of the dataset and repo on the hub
+- **dataset.num_episodes_to_record**: Number of episodes to record
+- **env.processor.reset.terminate_on_success**: Whether to automatically terminate episodes when success is detected (default: `true`)
+- **env.fps**: Number of frames per second to record
+- **dataset.push_to_hub**: Whether to push the dataset to the hub
 
-The `number_of_steps_after_success` parameter is crucial as it allows you to collect more positive examples. When a success is detected, the system will continue recording for the specified number of steps while maintaining the reward=1 label. Otherwise, there won't be enough states in the dataset labeled to 1 to train a good classifier.
+The `env.processor.reset.terminate_on_success` parameter allows you to control episode termination behavior. When set to `false`, episodes will continue even after success is detected, allowing you to collect more positive examples with the reward=1 label. This is crucial for training reward classifiers as it provides more success state examples in your dataset. When set to `true` (default), episodes terminate immediately upon success detection.
+
+**Important**: For reward classifier training, set `terminate_on_success: false` to collect sufficient positive examples. For regular HIL-SERL training, keep it as `true` to enable automatic episode termination when the task is completed successfully.
 
 Example configuration section for data collection:
 
 ```json
 {
+  "env": {
+    "type": "gym_manipulator",
+    "name": "real_robot",
+    "fps": 10,
+    "processor": {
+      "reset": {
+        "reset_time_s": 5.0,
+        "control_time_s": 20.0,
+        "terminate_on_success": false
+      },
+      "gripper": {
+        "use_gripper": true
+      }
+    },
+    "robot": {
+      // ... robot configuration ...
+    },
+    "teleop": {
+      // ... teleoperator configuration ...
+    }
+  },
+  "dataset": {
+    "repo_id": "hf_username/dataset_name",
+    "dataset_root": "data/your_dataset",
+    "task": "reward_classifier_task",
+    "num_episodes_to_record": 20,
+    "replay_episode": null,
+    "push_to_hub": true
+  },
   "mode": "record",
-  "repo_id": "hf_username/dataset_name",
-  "dataset_root": "data/your_dataset",
-  "num_episodes": 20,
-  "push_to_hub": true,
-  "fps": 10,
-  "number_of_steps_after_success": 15
+  "device": "cpu"
 }
 ```
 
@@ -421,9 +724,17 @@ To use your trained reward classifier, configure the `HILSerlRobotEnvConfig` to
 
 <!-- prettier-ignore-start -->
 ```python
-env_config = HILSerlRobotEnvConfig(
-    reward_classifier_pretrained_path="path_to_your_pretrained_trained_model",
-    # Other environment parameters
+config = GymManipulatorConfig(
+    env=HILSerlRobotEnvConfig(
+        processor=HILSerlProcessorConfig(
+            reward_classifier=RewardClassifierConfig(
+                pretrained_path="path_to_your_pretrained_trained_model"
+            )
+        ),
+        # Other environment parameters
+    ),
+    dataset=DatasetConfig(...),
+    mode=None  # For training
 )
 ```
 <!-- prettier-ignore-end -->
@@ -432,14 +743,25 @@ or set the argument in the json config file.
 
 ```json
 {
-  "reward_classifier_pretrained_path": "path_to_your_pretrained_model"
+  "env": {
+    "processor": {
+      "reward_classifier": {
+        "pretrained_path": "path_to_your_pretrained_model",
+        "success_threshold": 0.7,
+        "success_reward": 1.0
+      },
+      "reset": {
+        "terminate_on_success": true
+      }
+    }
+  }
 }
 ```
 
 Run `gym_manipulator.py` to test the model.
 
 ```bash
-python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/env_config.json
+python -m lerobot.rl.gym_manipulator --config_path path/to/env_config.json
 ```
 
 The reward classifier will automatically provide rewards based on the visual input from the robot's cameras.
@@ -447,12 +769,12 @@ The reward classifier will automatically provide rewards based on the visual inp
 **Example Workflow for training the reward classifier**
 
 1. **Create the configuration files**:
-   Create the necessary json configuration files for the reward classifier and the environment. Check the examples [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/tree/main).
+   Create the necessary json configuration files for the reward classifier and the environment. Check the examples [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/reward_classifier/config.json).
 
 2. **Collect a dataset**:
 
    ```bash
-   python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
+   python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
    ```
 
 3. **Train the classifier**:
@@ -463,7 +785,7 @@ The reward classifier will automatically provide rewards based on the visual inp
 
 4. **Test the classifier**:
    ```bash
-   python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
+   python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
    ```
 
 ### Training with Actor-Learner
@@ -472,7 +794,7 @@ The LeRobot system uses a distributed actor-learner architecture for training. T
 
 **Configuration Setup**
 
-Create a training configuration file (example available [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/train_config_hilserl_so100.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/configs/train.py`.
+Create a training configuration file (example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/train_config.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/configs/train.py`.
 
 1. Configure the policy settings (`type="sac"`, `device`, etc.)
 2. Set `dataset` to your cropped dataset
@@ -485,7 +807,7 @@ Create a training configuration file (example available [here](https://huggingfa
 First, start the learner server process:
 
 ```bash
-python -m lerobot.scripts.rl.learner --config_path src/lerobot/configs/train_config_hilserl_so100.json
+python -m lerobot.rl.learner --config_path src/lerobot/configs/train_config_hilserl_so100.json
 ```
 
 The learner:
@@ -500,7 +822,7 @@ The learner:
 In a separate terminal, start the actor process with the same configuration:
 
 ```bash
-python -m lerobot.scripts.rl.actor --config_path src/lerobot/configs/train_config_hilserl_so100.json
+python -m lerobot.rl.actor --config_path src/lerobot/configs/train_config_hilserl_so100.json
 ```
 
 The actor:

docs/source/hilserl_sim.mdx

@@ -26,15 +26,18 @@ pip install -e ".[hilserl]"
 
 ## Configuration
 
-To use `gym_hil` with LeRobot, you need to create a configuration file. An example is provided [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/gym_hil_env.json). Key configuration sections include:
+To use `gym_hil` with LeRobot, you need to create a configuration file. An example is provided [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/gym_hil/env_config.json). Key configuration sections include:
 
 ### Environment Type and Task
 
 ```json
 {
-  "type": "hil",
-  "name": "franka_sim",
-  "task": "PandaPickCubeGamepad-v0",
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0",
+    "fps": 10
+  },
   "device": "cuda"
 }
 ```
@@ -45,28 +48,40 @@ Available tasks:
 - `PandaPickCubeGamepad-v0`: With gamepad control
 - `PandaPickCubeKeyboard-v0`: With keyboard control
 
-### Gym Wrappers Configuration
+### Processor Configuration
 
 ```json
-"wrapper": {
-    "gripper_penalty": -0.02,
-    "control_time_s": 15.0,
-    "use_gripper": true,
-    "fixed_reset_joint_positions": [0.0, 0.195, 0.0, -2.43, 0.0, 2.62, 0.785],
-    "end_effector_step_sizes": {
-        "x": 0.025,
-        "y": 0.025,
-        "z": 0.025
-    },
-    "control_mode": "gamepad"
+{
+  "env": {
+    "processor": {
+      "control_mode": "gamepad",
+      "gripper": {
+        "use_gripper": true,
+        "gripper_penalty": -0.02
+      },
+      "reset": {
+        "control_time_s": 15.0,
+        "fixed_reset_joint_positions": [
+          0.0, 0.195, 0.0, -2.43, 0.0, 2.62, 0.785
+        ]
+      },
+      "inverse_kinematics": {
+        "end_effector_step_sizes": {
+          "x": 0.025,
+          "y": 0.025,
+          "z": 0.025
+        }
+      }
     }
+  }
+}
 ```
 
 Important parameters:
 
-- `gripper_penalty`: Penalty for excessive gripper movement
-- `use_gripper`: Whether to enable gripper control
-- `end_effector_step_sizes`: Size of the steps in the x,y,z axes of the end-effector
+- `gripper.gripper_penalty`: Penalty for excessive gripper movement
+- `gripper.use_gripper`: Whether to enable gripper control
+- `inverse_kinematics.end_effector_step_sizes`: Size of the steps in the x,y,z axes of the end-effector
 - `control_mode`: Set to `"gamepad"` to use a gamepad controller
 
 ## Running with HIL RL of LeRobot
@@ -75,39 +90,50 @@ Important parameters:
 
 To run the environment, set mode to null:
 
-<!-- prettier-ignore-start -->
-```python
-python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/gym_hil_env.json
+```bash
+python -m lerobot.rl.gym_manipulator --config_path path/to/gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 ### Recording a Dataset
 
 To collect a dataset, set the mode to `record` whilst defining the repo_id and number of episodes to record:
 
-<!-- prettier-ignore-start -->
-```python
-python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/gym_hil_env.json
+```json
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0"
+  },
+  "dataset": {
+    "repo_id": "username/sim_dataset",
+    "root": null,
+    "task": "pick_cube",
+    "num_episodes_to_record": 10,
+    "replay_episode": null,
+    "push_to_hub": true
+  },
+  "mode": "record"
+}
+```
+
+```bash
+python -m lerobot.rl.gym_manipulator --config_path path/to/gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 ### Training a Policy
 
-To train a policy, checkout the configuration example available [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/train_gym_hil_env.json) and run the actor and learner servers:
+To train a policy, checkout the configuration example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/gym_hil/train_config.json) and run the actor and learner servers:
 
-<!-- prettier-ignore-start -->
-```python
-python -m lerobot.scripts.rl.actor --config_path path/to/train_gym_hil_env.json
+```bash
+python -m lerobot.rl.actor --config_path path/to/train_gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 In a different terminal, run the learner server:
 
-<!-- prettier-ignore-start -->
-```python
-python -m lerobot.scripts.rl.learner --config_path path/to/train_gym_hil_env.json
+```bash
+python -m lerobot.rl.learner --config_path path/to/train_gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 The simulation environment provides a safe and repeatable way to develop and test your Human-In-the-Loop reinforcement learning components before deploying to real robots.
 

docs/source/il_robots.mdx

@@ -200,7 +200,7 @@ from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderCo
 from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import _init_rerun
+from lerobot.utils.visualization_utils import init_rerun
 from lerobot.record import record_loop
 
 NUM_EPISODES = 5
@@ -237,7 +237,7 @@ dataset = LeRobotDataset.create(
 
 # Initialize the keyboard listener and rerun visualization
 _, events = init_keyboard_listener()
-_init_rerun(session_name="recording")
+init_rerun(session_name="recording")
 
 # Connect the robot and teleoperator
 robot.connect()
@@ -517,13 +517,16 @@ from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerCon
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import _init_rerun
+from lerobot.utils.visualization_utils import init_rerun
 from lerobot.record import record_loop
+from lerobot.policies.factory import make_processor
 
 NUM_EPISODES = 5
 FPS = 30
 EPISODE_TIME_SEC = 60
 TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
 
 # Create the robot configuration
 camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
@@ -535,7 +538,7 @@ robot_config = SO100FollowerConfig(
 robot = SO100Follower(robot_config)
 
 # Initialize the policy
-policy = ACTPolicy.from_pretrained("<hf_username>/<my_policy_repo_id>")
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
 # Configure the dataset features
 action_features = hw_to_dataset_features(robot.action_features, "action")
@@ -544,7 +547,7 @@ dataset_features = {**action_features, **obs_features}
 
 # Create the dataset
 dataset = LeRobotDataset.create(
-    repo_id="<hf_username>/eval_<dataset_repo_id>",
+    repo_id=HF_DATASET_ID,
     fps=FPS,
     features=dataset_features,
     robot_type=robot.name,
@@ -554,11 +557,17 @@ dataset = LeRobotDataset.create(
 
 # Initialize the keyboard listener and rerun visualization
 _, events = init_keyboard_listener()
-_init_rerun(session_name="recording")
+init_rerun(session_name="recording")
 
 # Connect the robot
 robot.connect()
 
+preprocessor, postprocessor = make_processor(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+)
+
 for episode_idx in range(NUM_EPISODES):
     log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
@@ -568,6 +577,8 @@ for episode_idx in range(NUM_EPISODES):
         events=events,
         fps=FPS,
         policy=policy,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
         dataset=dataset,
         control_time_s=EPISODE_TIME_SEC,
         single_task=TASK_DESCRIPTION,

docs/source/il_sim.mdx

@@ -22,13 +22,38 @@ pip install -e ".[hilserl]"
 
 ## Teleoperate and Record a Dataset
 
-To use `gym_hil` with LeRobot, you need to use a configuration file. An example config file can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_gym_hil_il.json).
+To use `gym_hil` with LeRobot, you need to use a configuration file. An example config file can be found [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/sim_il/env_config.json).
 
-To teleoperate and collect a dataset, we need to modify this config file and you should add your `repo_id` here: `"repo_id": "il_gym",` and `"num_episodes": 30,` and make sure you set `mode` to `record`, "mode": "record".
+To teleoperate and collect a dataset, we need to modify this config file. Here's an example configuration for imitation learning data collection:
 
-If you do not have a Nvidia GPU also change `"device": "cuda"` parameter in the config file (for example to `mps` for MacOS).
+```json
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0",
+    "fps": 10
+  },
+  "dataset": {
+    "repo_id": "your_username/il_gym",
+    "root": null,
+    "task": "pick_cube",
+    "num_episodes_to_record": 30,
+    "replay_episode": null,
+    "push_to_hub": true
+  },
+  "mode": "record",
+  "device": "cuda"
+}
+```
 
-By default the config file assumes you use a controller. To use your keyboard please change the envoirment specified at `"task"` in the config file and set it to `"PandaPickCubeKeyboard-v0"`.
+Key configuration points:
+
+- Set your `repo_id` in the `dataset` section: `"repo_id": "your_username/il_gym"`
+- Set `num_episodes_to_record: 30` to collect 30 demonstration episodes
+- Ensure `mode` is set to `"record"`
+- If you don't have an NVIDIA GPU, change `"device": "cuda"` to `"mps"` for macOS or `"cpu"`
+- To use keyboard instead of gamepad, change `"task"` to `"PandaPickCubeKeyboard-v0"`
 
 Then we can run this command to start:
 
@@ -36,14 +61,14 @@ Then we can run this command to start:
 <hfoption id="Linux">
 
 ```bash
-python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
+python -m lerobot.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
 ```
 
 </hfoption>
 <hfoption id="MacOS">
 
 ```bash
-mjpython -m lerobot.scripts.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
+mjpython -m lerobot.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
 ```
 
 </hfoption>
@@ -140,9 +165,32 @@ huggingface-cli upload ${HF_USER}/il_sim_test${CKPT} \
 
 ## Evaluate your policy in Sim
 
-To evaluate your policy we have to use the config file that can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/eval_config_gym_hil.json).
+To evaluate your policy we have to use a configuration file. An example can be found [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/sim_il/eval_config.json).
 
-Make sure to replace the `repo_id` with the dataset you trained on, for example `pepijn223/il_sim_dataset` and replace the `pretrained_policy_name_or_path` with your model id, for example `pepijn223/il_sim_model`
+Here's an example evaluation configuration:
+
+```json
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0",
+    "fps": 10
+  },
+  "dataset": {
+    "repo_id": "your_username/il_sim_dataset",
+    "dataset_root": null,
+    "task": "pick_cube"
+  },
+  "pretrained_policy_name_or_path": "your_username/il_sim_model",
+  "device": "cuda"
+}
+```
+
+Make sure to replace:
+
+- `repo_id` with the dataset you trained on (e.g., `your_username/il_sim_dataset`)
+- `pretrained_policy_name_or_path` with your model ID (e.g., `your_username/il_sim_model`)
 
 Then you can run this command to visualize your trained policy
 
@@ -150,14 +198,14 @@ Then you can run this command to visualize your trained policy
 <hfoption id="Linux">
 
 ```bash
-python -m lerobot.scripts.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
+python -m lerobot.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
 ```
 
 </hfoption>
 <hfoption id="MacOS">
 
 ```bash
-mjpython -m lerobot.scripts.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
+mjpython -m lerobot.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
 ```
 
 </hfoption>

docs/source/implement_your_own_processor.mdx

@@ -0,0 +1,273 @@
+# Implement your own Robot Processor
+
+In this tutorial, you'll learn how to implement your own Robot Processor.
+It begins by exploring the need for a custom processor, then uses the `NormalizerProcessorStep` as the running example to explain how to implement, configure, and serialize a processor. Finally, it lists all helper processors that ship with LeRobot.
+
+## Why would you need a custom processor?
+
+In most cases, when reading raw data from sensors or when models output actions, you need to process this data to make it compatible with your target system. For example, a common need is normalizing data ranges to make them suitable for neural networks.
+
+LeRobot's `NormalizerProcessorStep` handles this crucial task:
+
+```python
+# Input: raw joint positions in [0, 180] degrees
+raw_action = torch.tensor([90.0, 45.0, 135.0])
+
+# After processing: normalized to [-1, 1] range for model training
+normalizer = NormalizerProcessorStep(features=features, norm_map=norm_map, stats=dataset_stats)
+normalized_result = normalizer(transition)
+# ...
+```
+
+Other common processing needs include:
+
+- **Device placement**: Moving tensors between CPU/GPU and converting data types
+- **Format conversion**: Transforming between different data structures
+- **Batching**: Adding/removing batch dimensions for model compatibility
+- **Safety constraints**: Applying limits to robot commands
+
+```python
+# Example pipeline combining multiple processors
+pipeline = PolicyProcessorPipeline([
+    RenameObservationsProcessorStep(rename_map={}),
+    AddBatchDimensionProcessorStep(),
+    NormalizerProcessorStep(features=features, stats=stats),
+    DeviceProcessorStep(device="cuda"),
+    # ...
+])
+```
+
+LeRobot provides a pipeline mechanism to implement sequences of processing steps for both input data and output actions, making it easy to compose these transformations in the right order for optimal performance.
+
+## How to implement your own processor?
+
+We'll use the `NormalizerProcessorStep` as our main example because it demonstrates essential processor patterns including state management, configuration serialization, and tensor handling that you'll commonly need.
+
+Prepare the sequence of processing steps necessary for your problem. A processor step is a class that implements the following methods:
+
+- `__call__`: implements the processing step for the input transition.
+- `get_config`: gets the configuration of the processor step.
+- `state_dict`: gets the state of the processor step.
+- `load_state_dict`: loads the state of the processor step.
+- `reset`: resets the state of the processor step.
+- `feature_contract`: displays the modification to the feature space during the processor step.
+
+### Implement the `__call__` method
+
+The `__call__` method is the core of your processor step. It takes an `EnvTransition` and returns a modified `EnvTransition`. Here's how the `NormalizerProcessorStep` works:
+
+```python
+@dataclass
+@ProcessorStepRegistry.register("normalizer_processor")
+class NormalizerProcessorStep(ProcessorStep):
+    """Normalize observations/actions using dataset statistics."""
+
+    features: dict[str, PolicyFeature]
+    norm_map: dict[FeatureType, NormalizationMode]
+    stats: dict[str, dict[str, Any]] | None = None
+    eps: float = 1e-8
+    _tensor_stats: dict = field(default_factory=dict, init=False, repr=False)
+
+    def __post_init__(self):
+        """Convert stats to tensors for efficient computation."""
+        self.stats = self.stats or {}
+        self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=torch.float32)
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        new_transition = transition.copy()
+        # Normalize observations
+        # ...
+        # Normalize action
+        # ...
+        return new_transition
+
+```
+
+See the full implementation in `src/lerobot/processor/normalize_processor.py` for complete details.
+
+**Key principles:**
+
+- **Always use `transition.copy()`** to avoid side effects
+- **Handle both observations and actions** consistently
+- **Separate config from state**: `get_config()` returns JSON-serializable params, `state_dict()` returns tensors
+- **Convert stats to tensors** in `__post_init__()` for efficient computation
+
+### Configuration and State Management
+
+Processors support serialization through three methods that separate configuration from tensor state. The `NormalizerProcessorStep` demonstrates this perfectly - it carries dataset statistics (tensors) in its state, and hyperparameters in its config:
+
+```python
+# Continuing the NormalizerProcessorStep example...
+
+def get_config(self) -> dict[str, Any]:
+    """JSON-serializable configuration (no tensors)."""
+    return {
+        "eps": self.eps,
+        "features": {k: {"type": v.type.value, "shape": v.shape} for k, v in self.features.items()},
+        "norm_map": {ft.value: nm.value for ft, nm in self.norm_map.items()},
+        # ...
+    }
+
+def state_dict(self) -> dict[str, torch.Tensor]:
+    """Tensor state only (e.g., dataset statistics)."""
+    flat: dict[str, torch.Tensor] = {}
+    for key, sub in self._tensor_stats.items():
+        for stat_name, tensor in sub.items():
+            flat[f"{key}.{stat_name}"] = tensor.cpu()  # Always save to CPU
+    return flat
+
+def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
+    """Restore tensor state at runtime."""
+    self._tensor_stats.clear()
+    for flat_key, tensor in state.items():
+        key, stat_name = flat_key.rsplit(".", 1)
+        # Load to processor's configured device
+        self._tensor_stats.setdefault(key, {})[stat_name] = tensor.to(
+            dtype=torch.float32, device=self.device
+        )
+        # ...
+```
+
+**Usage:**
+
+```python
+# Save (e.g., inside a policy)
+config = normalizer.get_config()
+tensors = normalizer.state_dict()
+
+# Restore (e.g., loading a pretrained policy)
+new_normalizer = NormalizerProcessorStep(**config)
+new_normalizer.load_state_dict(tensors)
+# Now new_normalizer has the same stats and configuration
+```
+
+### Transform features
+
+The `transform_features` method defines how your processor transforms feature names and shapes. This is crucial for policy configuration and debugging.
+
+For `NormalizerProcessorStep`, features are typically preserved unchanged since normalization doesn't alter keys or shapes:
+
+```python
+def transform_features(self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+    """Normalization preserves all feature definitions."""
+    return features  # No changes to feature structure
+    # ...
+```
+
+When your processor renames or reshapes data, implement this method to reflect the mapping for downstream components. For example, a simple rename processor:
+
+```python
+def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+    # Simple renaming
+    if "pixels" in features:
+        features["observation.image"] = features.pop("pixels")
+
+    # Pattern-based renaming
+    for key in list(features.keys()):
+        if key.startswith("env_state."):
+            suffix = key[len("env_state."):]
+            features[f"observation.{suffix}"] = features.pop(key)
+            # ...
+
+    return features
+```
+
+**Key principles:**
+
+- Use `features.pop(old_key)` to remove and get the old feature
+- Use `features[new_key] = old_feature` to add the renamed feature
+- Always return the modified features dictionary
+- Document transformations clearly in the docstring
+
+### Using overrides
+
+You can override step parameters at load-time using `overrides`. This is handy for non-serializable objects or site-specific settings. It works both in policy factories and with `DataProcessorPipeline.from_pretrained(...)`.
+
+**Foundational model adaptation**: This is particularly useful when working with foundational pretrained policies where you rarely have access to the original training statistics. You can inject your own dataset statistics to adapt the normalizer to your specific robot or environment data.
+
+Example: during policy evaluation on the robot, override the device and rename map.
+Use this to run a policy trained on CUDA on a CPU-only robot, or to remap camera keys when the robot uses different names than the dataset.
+
+Direct usage with `from_pretrained`:
+
+```python
+from lerobot.processor import RobotProcessorPipeline
+
+# Load a foundational policy trained on diverse robot data
+# but adapt normalization to your specific robot/environment
+new_stats = LeRobotDataset(repo_id="username/my-dataset").meta.stats
+processor = RobotProcessorPipeline.from_pretrained(
+    "huggingface/foundational-robot-policy",  # Pretrained foundation model
+    overrides={
+        "normalizer_processor": {"stats": new_stats},     # Inject your robot's statistics
+        "device_processor": {"device": "cuda:0"},         # registry name for registered steps
+        "rename_processor": {"rename_map": robot_key_map}, # Map your robot's observation keys
+        # ...
+    },
+)
+```
+
+## Best Practices
+
+Based on analysis of all LeRobot processor implementations, here are the key patterns and practices:
+
+### 1. **Safe Data Handling**
+
+Always create copies of input data to avoid unintended side effects. Use `transition.copy()` and `observation.copy()` rather than modifying data in-place. This prevents your processor from accidentally affecting other components in the pipeline.
+
+Check for required data before processing and handle missing data gracefully. If your processor expects certain keys (like `"pixels"` for image processing), validate their presence first. For optional data, use safe access patterns like `transition.get()` and handle `None` values appropriately.
+
+When data validation fails, provide clear, actionable error messages that help users understand what went wrong and how to fix it.
+
+### 2. **Choose Appropriate Base Classes**
+
+LeRobot provides specialized base classes that reduce boilerplate code and ensure consistency. Use `ObservationProcessorStep` when you only need to modify observations, `ActionProcessorStep` for action-only processing, and `RobotActionProcessorStep` specifically for dictionary-based robot actions.
+
+Only inherit directly from `ProcessorStep` when you need full control over the entire transition or when processing multiple transition components simultaneously. The specialized base classes handle the transition management for you and provide type safety.
+
+### 3. **Registration and Naming**
+
+Register your processors with descriptive, namespaced names using `@ProcessorStepRegistry.register()`. Use organization prefixes like `"robotics_lab/safety_clipper"` or `"acme_corp/vision_enhancer"` to avoid naming conflicts. Avoid generic names like `"processor"` or `"step"` that could clash with other implementations.
+
+Good registration makes your processors discoverable and enables clean serialization/deserialization when saving and loading pipelines.
+
+### 4. **State Management Patterns**
+
+Distinguish between configuration parameters (JSON-serializable values) and internal state (tensors, buffers). Use dataclass fields with `init=False, repr=False` for internal state that shouldn't appear in the constructor or string representation.
+
+Implement the `reset()` method to clear internal state between episodes. This is crucial for stateful processors that accumulate data over time, like moving averages or temporal filters.
+
+Remember that `get_config()` should only return JSON-serializable configuration, while `state_dict()` handles tensor state separately.
+
+### 5. **Input Validation and Error Handling**
+
+Validate input types and shapes before processing. Check tensor properties like `dtype` and dimensions to ensure compatibility with your algorithms. For robot actions, verify that required pose components or joint values are present and within expected ranges.
+
+Use early returns for edge cases where no processing is needed. Provide clear, descriptive error messages that include the expected vs. actual data types or shapes. This makes debugging much easier for users.
+
+### 6. **Device and Dtype Awareness**
+
+Design your processors to automatically adapt to the device and dtype of input tensors. Internal tensors (like normalization statistics) should match the input tensor's device and dtype to ensure compatibility with multi-GPU training, mixed precision, and distributed setups.
+
+Implement a `to()` method that moves your processor's internal state to the specified device. Check device/dtype compatibility at runtime and automatically migrate internal state when needed. This pattern enables seamless operation across different hardware configurations without manual intervention.
+
+## Conclusion
+
+You now have all the tools to implement custom processors in LeRobot! The key steps are:
+
+1. **Define your processor** as a dataclass with the required methods (`__call__`, `get_config`, `state_dict`, `load_state_dict`, `reset`, `transform_features`)
+2. **Register it** using `@ProcessorStepRegistry.register("name")` for discoverability
+3. **Integrate it** into a `DataProcessorPipeline` with other processing steps
+4. **Use base classes** like `ObservationProcessorStep` when possible to reduce boilerplate
+5. **Implement device/dtype awareness** to support multi-GPU and mixed precision setups
+
+The processor system is designed to be modular and composable, allowing you to build complex data processing pipelines from simple, focused components. Whether you're preprocessing sensor data for training or post-processing model outputs for robot execution, custom processors give you the flexibility to handle any data transformation your robotics application requires.
+
+Key principles for robust processors:
+
+- **Device/dtype adaptation**: Internal tensors should match input tensors
+- **Clear error messages**: Help users understand what went wrong
+- **Base class usage**: Leverage specialized base classes to reduce boilerplate
+- **Feature contracts**: Declare data structure changes with `transform_features()`
+
+Start simple, test thoroughly, and ensure your processors work seamlessly across different hardware configurations!

docs/source/introduction_processors.mdx

@@ -0,0 +1,314 @@
+# Introduction to Processors
+
+In robotics, there's a fundamental mismatch between the data that robots and humans produce and what machine learning models expect.
+Robots output raw sensor data like camera images and joint positions that need normalization, batching, and device placement before models can process them.
+Language instructions from humans must be tokenized into numerical representations, and different robots use different coordinate systems that need standardization.
+
+The challenge extends to model outputs as well.
+Models might output end-effector positions while robots need joint-space commands, or teleoperators produce relative movements while robots expect absolute commands.
+Model predictions are often normalized and need conversion back to real-world scales.
+
+Cross-domain translation adds another layer of complexity.
+Training data from one robot setup needs adaptation for deployment on different hardware, models trained with specific camera configurations must work with new arrangements, and datasets with different naming conventions need harmonization.
+
+**That's where processors come in.** They serve as universal translators that bridge these gaps, ensuring seamless data flow from sensors to models to actuators.
+Processors handle all the preprocessing and postprocessing steps needed to convert raw environment data into model-ready inputs and vice versa.
+
+This means that your favorite policy can be used like this:
+
+```python
+import torch
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies.your_policy import YourPolicy
+from lerobot.processor.pipeline import RobotProcessorPipeline, PolicyProcessorPipeline
+dataset = LeRobotDataset("hf_user/dataset", episodes=[0])
+sample = dataset[10]
+
+model = YourPolicy.from_pretrained(
+    "hf_user/model",
+)
+model.eval()
+model.to("cuda")
+preprocessor, postprocessor = make_pre_post_processors(model.config, pretrained_path="hf_user/model", dataset_stats=dataset.meta.stats)
+
+preprocessed_sample = preprocessor(sample)
+action = model.select_action(preprocessed_sample)
+postprocessed_action = postprocessor(action)
+```
+
+## What are Processors?
+
+In robotics, data comes in many forms: images from cameras, joint positions from sensors, text instructions from users, and more. Each type of data requires specific transformations before a model can use it effectively. Models need this data to be:
+
+- **Normalized**: Scaled to appropriate ranges for neural network processing
+- **Batched**: Organized with proper dimensions for batch processing
+- **Tokenized**: Text converted to numerical representations
+- **Device-placed**: Moved to the right hardware (CPU/GPU)
+- **Type-converted**: Cast to appropriate data types
+
+Processors handle these transformations through composable, reusable steps that can be chained together into pipelines. Think of them as a modular assembly line where each station performs a specific transformation on your data.
+
+## Core Concepts
+
+### EnvTransition: The Universal Data Container
+
+The `EnvTransition` is the fundamental data structure that flows through all processors.
+It's a typed dictionary that represents a complete robot-environment interaction:
+
+- **OBSERVATION**: All sensor data (images, states, proprioception)
+- **ACTION**: The action to execute or that was executed
+- **REWARD**: Reinforcement learning signal
+- **DONE/TRUNCATED**: Episode boundary indicators
+- **INFO**: Arbitrary metadata
+- **COMPLEMENTARY_DATA**: Task descriptions, indices, padding flags, inter-step data
+
+### ProcessorStep: The Building Block
+
+A `ProcessorStep` is a single transformation unit that processes transitions. It's an abstract base class with two required methods:
+
+```python
+from lerobot.processor import ProcessorStep, EnvTransition
+
+class MyProcessorStep(ProcessorStep):
+    """Example processor step - inherit and implement abstract methods."""
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """Transform the transition - REQUIRED abstract method."""
+        # Your processing logic here
+        return transition
+
+    def transform_features(self, features):
+        """Declare how this step transforms feature shapes/types - REQUIRED abstract method."""
+        return features  # Most processors return features unchanged
+```
+
+`__call__` is the core of your processor step. It takes an `EnvTransition` and returns a modified `EnvTransition`.
+
+`transform_features` is used to declare how this step transforms feature shapes/types.
+
+### DataProcessorPipeline: The Generic Orchestrator
+
+The `DataProcessorPipeline[TInput, TOutput]` chains multiple `ProcessorStep` instances together:
+
+```python
+from lerobot.processor import RobotProcessorPipeline, PolicyProcessorPipeline
+
+# For robot hardware (unbatched data)
+robot_processor = RobotProcessorPipeline[RobotAction, RobotAction](
+    steps=[step1, step2, step3],
+    name="robot_pipeline"
+)
+
+# For model training/inference (batched data)
+policy_processor = PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+    steps=[step1, step2, step3],
+    name="policy_pipeline"
+)
+```
+
+## RobotProcessorPipeline vs PolicyProcessorPipeline
+
+The key distinction is in the data structures they handle:
+
+| Aspect          | RobotProcessorPipeline                       | PolicyProcessorPipeline                  |
+| --------------- | -------------------------------------------- | ---------------------------------------- |
+| **Input**       | `dict[str, Any]` - Individual robot values   | `dict[str, Any]` - Batched tensors       |
+| **Output**      | `dict[str, Any]` - Individual robot commands | `torch.Tensor` - Policy predictions      |
+| **Use Case**    | Real-time robot control                      | Model training/inference                 |
+| **Data Format** | Unbatched, heterogeneous                     | Batched, homogeneous                     |
+| **Examples**    | `{"joint_1": 0.5}`                           | `{"observation.state": tensor([[0.5]])}` |
+
+**Use `RobotProcessorPipeline`** for robot hardware interfaces:
+
+```python
+# Robot data structures: dict[str, Any] for observations and actions
+robot_obs: dict[str, Any] = {
+    "joint_1": 0.5,           # Individual joint values
+    "joint_2": -0.3,
+    "camera_0": image_array   # Raw camera data
+}
+
+robot_action: dict[str, Any] = {
+    "joint_1": 0.2,          # Target joint positions
+    "joint_2": 0.1,
+    "gripper": 0.8
+}
+```
+
+**Use `PolicyProcessorPipeline`** for model training and batch processing:
+
+```python
+# Policy data structures: batch dicts and tensors
+policy_batch: dict[str, Any] = {
+    "observation.state": torch.tensor([[0.5, -0.3]]),      # Batched states
+    "observation.images.camera0": torch.tensor(...),        # Batched images
+    "action": torch.tensor([[0.2, 0.1, 0.8]])              # Batched actions
+}
+
+policy_action: torch.Tensor = torch.tensor([[0.2, 0.1, 0.8]])  # Model output tensor
+```
+
+## Converter Functions
+
+LeRobot provides converter functions to bridge different data formats in `lerobot.processor.converters`. These functions handle the crucial translations between robot hardware data structures, policy model formats, and the internal `EnvTransition` representation that flows through processor pipelines.
+
+| Category                       | Function                      | Description                     |
+| ------------------------------ | ----------------------------- | ------------------------------- |
+| **Robot Hardware Converters**  | `robot_action_to_transition`  | Robot dict → EnvTransition      |
+|                                | `observation_to_transition`   | Robot obs → EnvTransition       |
+|                                | `transition_to_robot_action`  | EnvTransition → Robot dict      |
+| **Policy/Training Converters** | `batch_to_transition`         | Batch dict → EnvTransition      |
+|                                | `transition_to_batch`         | EnvTransition → Batch dict      |
+|                                | `policy_action_to_transition` | Policy tensor → EnvTransition   |
+|                                | `transition_to_policy_action` | EnvTransition → Policy tensor   |
+| **Utilities**                  | `create_transition`           | Build transitions with defaults |
+|                                | `identity_transition`         | Pass-through converter          |
+
+The key insight is that **robot hardware converters** work with individual values and dictionaries, while **policy/training converters** work with batched tensors and model outputs. The converter functions automatically handle the structural differences, so your processor steps can focus on the core transformations without worrying about data format compatibility.
+
+## Processor Examples
+
+The following examples demonstrate real-world processor configurations for policy training and inference.
+
+Here is an example processor for policy training and inference:
+
+```python
+# Training data preprocessing (optimized order for GPU performance)
+training_preprocessor = PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+    steps=[
+        RenameObservationsProcessorStep(rename_map={}),     # Standardize keys
+        AddBatchDimensionProcessorStep(),                   # Add batch dims
+        TokenizerProcessorStep(tokenizer_name="...", ...),  # Tokenize language
+        DeviceProcessorStep(device="cuda"),                 # Move to GPU first
+        NormalizerProcessorStep(features=..., stats=...),   # Normalize on GPU
+    ]
+)
+
+# Model output postprocessing
+training_postprocessor = PolicyProcessorPipeline[torch.Tensor, torch.Tensor](
+    steps=[
+        DeviceProcessorStep(device="cpu"),                  # Move to CPU
+        UnnormalizerProcessorStep(features=..., stats=...), # Denormalize
+    ]
+    to_transition=policy_action_to_transition,
+    to_output=transition_to_policy_action,
+)
+```
+
+### An interaction between a robot and a policy with processors
+
+The most common real-world scenario combines both pipeline types robot hardware generates observations that need policy processing, and policy outputs need robot-compatible postprocessing:
+
+```python
+# Real deployment: Robot sensors → Model → Robot commands
+with torch.no_grad():
+    while not done:
+        raw_obs = robot.get_observation()  # dict[str, Any]
+
+        # Add your robot observation to policy observation processor
+
+        policy_input = policy_preprocessor(raw_obs)  # Batched dict
+
+        policy_output = policy.select_action(policy_input)  # Policy tensor
+
+        policy_action = policy_postprocessor(policy_output)
+
+        # Add your robot action to policy action processor
+
+        robot.send_action(policy_action)
+```
+
+## Feature Contracts: Shape and Type Transformation
+
+Processors don't just transform data - they can also **change the data structure itself**. The `transform_features()` method declares these changes, which is crucial for dataset recording and policy creation.
+
+### Why Feature Contracts Matter
+
+When building datasets or policies, LeRobot needs to know:
+
+- **What data fields will exist** after processing
+- **What shapes and types** each field will have
+- **How to configure models** for the expected data structure
+
+```python
+# Example: A processor that adds velocity to observations
+class VelocityProcessor(ObservationProcessorStep):
+    def observation(self, obs):
+        new_obs = obs.copy()
+        if "observation.state" in obs:
+            # concatenate computed velocity field to the state
+            new_obs["observation.state"] = self._compute_velocity(obs["observation.state"])
+        return new_obs
+
+    def transform_features(self, features):
+        """Declare the new velocity field we're adding."""
+        state_feature = features[PipelineFeatureType.OBSERVATION].get("observation.state")
+        if state_feature:
+            double_shape = (state_feature.shape[0] * 2,) if state_feature.shape else (2,)
+            features[PipelineFeatureType.OBSERVATION]["observation.state"] = PolicyFeature(
+                type=FeatureType.STATE, shape=double_shape
+            )
+        return features
+```
+
+### Feature Specification Functions
+
+`create_initial_features()` and `aggregate_pipeline_dataset_features()` solve a critical dataset creation problem: determining the exact final data structure before any data is processed.
+Since processor pipelines can add new features (like velocity fields), change tensor shapes (like cropping images), or rename keys, datasets need to know the complete output specification upfront to allocate proper storage and define schemas.
+These functions work together by starting with robot hardware specifications (`create_initial_features()`) then simulating the entire pipeline transformation (`aggregate_pipeline_dataset_features()`) to compute the final feature dictionary that gets passed to `LeRobotDataset.create()`, ensuring perfect alignment between what processors output and what datasets expect to store.
+
+```python
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features
+
+# Start with robot's raw features
+initial_features = create_initial_features(
+    observation=robot.observation_features,  # {"joint_1.pos": float, "camera_0": (480,640,3)}
+    action=robot.action_features            # {"joint_1.pos": float, "gripper.pos": float}
+)
+
+# Apply processor pipeline to compute final features
+final_features = aggregate_pipeline_dataset_features(
+    pipeline=my_processor_pipeline,
+    initial_features=initial_features,
+    use_videos=True
+)
+
+# Use for dataset creation
+dataset = LeRobotDataset.create(
+    repo_id="my_dataset",
+    features=final_features,  # Knows exactly what data to expect
+    ...
+)
+```
+
+## Common Processor Steps
+
+LeRobot provides many registered processor steps. Here are the most commonly used core processors:
+
+### Essential Processors
+
+- **`normalizer_processor`**: Normalize observations/actions using dataset statistics (mean/std or min/max)
+- **`device_processor`**: Move tensors to CPU/GPU with optional dtype conversion
+- **`to_batch_processor`**: Add batch dimensions to transitions for model compatibility
+- **`rename_observations_processor`**: Rename observation keys using mapping dictionaries
+- **`tokenizer_processor`**: Tokenize natural language task descriptions into tokens and attention masks
+
+### Next Steps
+
+- **[Implement Your Own Processor](implement_your_own_processor.mdx)** - Create custom processor steps
+- **[Debug Your Pipeline](debug_processor_pipeline.mdx)** - Troubleshoot and optimize pipelines
+- **[Processors for Robots and Teleoperators](processors_robots_teleop.mdx)** - Real-world integration patterns
+
+## Summary
+
+Processors solve the data translation problem in robotics by providing:
+
+- **Modular transformations**: Composable, reusable processing steps
+- **Type safety**: Generic pipelines with compile-time checking
+- **Performance optimization**: GPU-accelerated operations
+- **Robot/Policy distinction**: Separate pipelines for different data structures
+- **Comprehensive ecosystem**: 30+ registered processors for common tasks
+
+The key insight: `RobotProcessorPipeline` handles unbatched robot hardware data, while `PolicyProcessorPipeline` handles batched model data. Choose the right tool for your data structure!

docs/source/koch.mdx

@@ -277,7 +277,7 @@ leader.disconnect()
 </hfoption>
 </hfoptions>
 
-Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./getting_started_real_world_robot)
+Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./il_robots)
 
 > [!TIP]
 > If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).

docs/source/lekiwi.mdx

@@ -323,7 +323,7 @@ To replay an episode run the API example below, make sure to change `remote_ip`,
 python examples/lekiwi/replay.py
 ```
 
-Congrats 🎉, your robot is all set to learn a task on its own. Start training it by the training part of this tutorial: [Getting started with real-world robots](./getting_started_real_world_robot)
+Congrats 🎉, your robot is all set to learn a task on its own. Start training it by the training part of this tutorial: [Getting started with real-world robots](./il_robots)
 
 ## Evaluate your policy
 

docs/source/lerobot-dataset-v3.mdx

@@ -8,6 +8,7 @@ This docs will guide you to:
 - Record a dataset and push it to the Hub
 - Load datasets for training with `LeRobotDataset`
 - Stream datasets without downloading using `StreamingLeRobotDataset`
+- Apply image transforms for data augmentation during training
 - Migrate existing `v2.1` datasets to `v3.0`
 
 ## What’s new in `v3`
@@ -150,6 +151,117 @@ dataset = StreamingLeRobotDataset(repo_id)  # streams directly from the Hub
   </figure>
 </div>
 
+## Image transforms
+
+Image transforms are data augmentations applied to camera frames during training to improve model robustness and generalization. LeRobot supports various transforms including brightness, contrast, saturation, hue, and sharpness adjustments.
+
+### Using transforms during dataset creation/recording
+
+Currently, transforms are applied during **training time only**, not during recording. When you create or record a dataset, the raw images are stored without transforms. This allows you to experiment with different augmentations later without re-recording data.
+
+### Adding transforms to existing datasets (API)
+
+Use the `image_transforms` parameter when loading a dataset for training:
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.transforms import ImageTransforms, ImageTransformsConfig, ImageTransformConfig
+
+# Option 1: Use default transform configuration (disabled by default)
+transforms_config = ImageTransformsConfig(
+    enable=True,  # Enable transforms
+    max_num_transforms=3,  # Apply up to 3 transforms per frame
+    random_order=False,  # Apply in standard order
+)
+transforms = ImageTransforms(transforms_config)
+
+dataset = LeRobotDataset(
+    repo_id="your-username/your-dataset",
+    image_transforms=transforms
+)
+
+# Option 2: Create custom transform configuration
+custom_transforms_config = ImageTransformsConfig(
+    enable=True,
+    max_num_transforms=2,
+    random_order=True,
+    tfs={
+        "brightness": ImageTransformConfig(
+            weight=1.0,
+            type="ColorJitter",
+            kwargs={"brightness": (0.7, 1.3)}  # Adjust brightness range
+        ),
+        "contrast": ImageTransformConfig(
+            weight=2.0,  # Higher weight = more likely to be selected
+            type="ColorJitter",
+            kwargs={"contrast": (0.8, 1.2)}
+        ),
+        "sharpness": ImageTransformConfig(
+            weight=0.5,  # Lower weight = less likely to be selected
+            type="SharpnessJitter",
+            kwargs={"sharpness": (0.3, 2.0)}
+        ),
+    }
+)
+
+dataset = LeRobotDataset(
+    repo_id="your-username/your-dataset",
+    image_transforms=ImageTransforms(custom_transforms_config)
+)
+
+# Option 3: Use pure torchvision transforms
+from torchvision.transforms import v2
+
+torchvision_transforms = v2.Compose([
+    v2.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),
+    v2.GaussianBlur(kernel_size=3, sigma=(0.1, 2.0)),
+])
+
+dataset = LeRobotDataset(
+    repo_id="your-username/your-dataset",
+    image_transforms=torchvision_transforms
+)
+```
+
+### Available transform types
+
+LeRobot provides several transform types:
+
+- **`ColorJitter`**: Adjusts brightness, contrast, saturation, and hue
+- **`SharpnessJitter`**: Randomly adjusts image sharpness
+- **`Identity`**: No transformation (useful for testing)
+
+You can also use any `torchvision.transforms.v2` transform by passing it directly to the `image_transforms` parameter.
+
+### Configuration options
+
+- **`enable`**: Enable/disable transforms (default: `False`)
+- **`max_num_transforms`**: Maximum number of transforms applied per frame (default: `3`)
+- **`random_order`**: Apply transforms in random order vs. standard order (default: `False`)
+- **`weight`**: Sampling probability for each transform (higher = more likely, if sum of weights is not 1, they will be normalized)
+- **`kwargs`**: Transform-specific parameters (e.g., brightness range)
+
+### Visualizing transforms
+
+Use the visualization script to preview how transforms affect your data:
+
+```bash
+lerobot-imgtransform-viz \
+  --repo-id=your-username/your-dataset \
+  --output-dir=./transform_examples \
+  --n-examples=5
+```
+
+This saves example images showing the effect of each transform, helping you tune parameters.
+
+### Best practices
+
+- **Start conservative**: Begin with small ranges (e.g., brightness 0.9-1.1) and increase gradually
+- **Test first**: Use the visualization script to ensure transforms look reasonable
+- **Monitor training**: Strong augmentations can hurt performance if too aggressive
+- **Match your domain**: If your robot operates in varying lighting, use brightness/contrast transforms
+- **Combine wisely**: Using too many transforms simultaneously can make training unstable
+
 ## Migrate `v2.1` → `v3.0`
 
 A converter aggregates per‑episode files into larger shards and writes episode offsets/metadata. Convert your dataset using the instructions below.

docs/source/libero.mdx

@@ -0,0 +1,127 @@
+# LIBERO
+
+**LIBERO** is a benchmark designed to study **lifelong robot learning**. The idea is that robots won’t just be pretrained once in a factory, they’ll need to keep learning and adapting with their human users over time. This ongoing adaptation is called **lifelong learning in decision making (LLDM)**, and it’s a key step toward building robots that become truly personalized helpers.
+
+- 📄 [LIBERO paper](https://arxiv.org/abs/2306.03310)
+- 💻 [Original LIBERO repo](https://github.com/Lifelong-Robot-Learning/LIBERO)
+
+To make progress on this challenge, LIBERO provides a set of standardized tasks that focus on **knowledge transfer**: how well a robot can apply what it has already learned to new situations. By evaluating on LIBERO, different algorithms can be compared fairly and researchers can build on each other’s work.
+
+LIBERO includes **five task suites**:
+
+- **LIBERO-Spatial (`libero_spatial`)** – tasks that require reasoning about spatial relations.
+- **LIBERO-Object (`libero_object`)** – tasks centered on manipulating different objects.
+- **LIBERO-Goal (`libero_goal`)** – goal-conditioned tasks where the robot must adapt to changing targets.
+- **LIBERO-90 (`libero_90`)** – 90 short-horizon tasks from the LIBERO-100 collection.
+- **LIBERO-Long (`libero_10`)** – 10 long-horizon tasks from the LIBERO-100 collection.
+
+Together, these suites cover **130 tasks**, ranging from simple object manipulations to complex multi-step scenarios. LIBERO is meant to grow over time, and to serve as a shared benchmark where the community can test and improve lifelong learning algorithms.
+
+![An overview of the LIBERO benchmark](https://libero-project.github.io/assets/img/libero/fig1.png)
+
+## Evaluating with LIBERO
+
+At **LeRobot**, we ported [LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO) into our framework and used it mainly to **evaluate [SmolVLA](https://huggingface.co/docs/lerobot/en/smolvla)**, our lightweight Vision-Language-Action model.
+
+LIBERO is now part of our **multi-eval supported simulation**, meaning you can benchmark your policies either on a **single suite of tasks** or across **multiple suites at once** with just a flag.
+
+To Install LIBERO, after following LeRobot official instructions, just do:
+`pip install -e ".[libero]"`
+
+### Single-suite evaluation
+
+Evaluate a policy on one LIBERO suite:
+
+```bash
+lerobot-eval \
+  --policy.path="your-policy-id" \
+  --env.type=libero \
+  --env.task=libero_object \
+  --eval.batch_size=2 \
+  --eval.n_episodes=3
+```
+
+- `--env.task` picks the suite (`libero_object`, `libero_spatial`, etc.).
+- `--eval.batch_size` controls how many environments run in parallel.
+- `--eval.n_episodes` sets how many episodes to run in total.
+
+---
+
+### Multi-suite evaluation
+
+Benchmark a policy across multiple suites at once:
+
+```bash
+lerobot-eval \
+  --policy.path="your-policy-id" \
+  --env.type=libero \
+  --env.task=libero_object,libero_spatial \
+  --eval.batch_size=1 \
+  --eval.n_episodes=2
+```
+
+- Pass a comma-separated list to `--env.task` for multi-suite evaluation.
+
+### Policy inputs and outputs
+
+When using LIBERO through LeRobot, policies interact with the environment via **observations** and **actions**:
+
+- **Observations**
+  - `observation.state` – proprioceptive features (agent state).
+  - `observation.images.image` – main camera view (`agentview_image`).
+  - `observation.images.image2` – wrist camera view (`robot0_eye_in_hand_image`).
+
+  ⚠️ **Note:** LeRobot enforces the `.images.*` prefix for any multi-modal visual features. Always ensure that your policy config `input_features` use the same naming keys, and that your dataset metadata keys follow this convention during evaluation.
+  If your data contains different keys, you must rename the observations to match what the policy expects, since naming keys are encoded inside the normalization statistics layer.
+  This will be fixed with the upcoming Pipeline PR.
+
+- **Actions**
+  - Continuous control values in a `Box(-1, 1, shape=(7,))` space.
+
+We also provide a notebook for quick testing:
+Training with LIBERO
+
+## Training with LIBERO
+
+When training on LIBERO tasks, make sure your dataset parquet and metadata keys follow the LeRobot convention.
+
+The environment expects:
+
+- `observation.state` → 8-dim agent state
+- `observation.images.image` → main camera (`agentview_image`)
+- `observation.images.image2` → wrist camera (`robot0_eye_in_hand_image`)
+
+⚠️ Cleaning the dataset upfront is **cleaner and more efficient** than remapping keys inside the code.
+To avoid potential mismatches and key errors, we provide a **preprocessed LIBERO dataset** that is fully compatible with the current LeRobot codebase and requires no additional manipulation:
+👉 [HuggingFaceVLA/libero](https://huggingface.co/datasets/HuggingFaceVLA/libero)
+
+For reference, here is the **original dataset** published by Physical Intelligence:
+👉 [physical-intelligence/libero](https://huggingface.co/datasets/physical-intelligence/libero)
+
+---
+
+### Example training command
+
+```bash
+lerobot-train \
+  --policy.type=smolvla \
+  --policy.repo_id=${HF_USER}/libero-test \
+  --policy.load_vlm_weights=true \
+  --dataset.repo_id=HuggingFaceVLA/libero \
+  --env.type=libero \
+  --env.task=libero_10 \
+  --output_dir=./outputs/ \
+  --steps=100000 \
+  --batch_size=4 \
+  --eval.batch_size=1 \
+  --eval.n_episodes=1 \
+  --eval_freq=1000 \
+```
+
+---
+
+### Note on rendering
+
+LeRobot uses MuJoCo for simulation. You need to set the rendering backend before training or evaluation:
+
+- `export MUJOCO_GL=egl` → for headless servers (e.g. HPC, cloud)

docs/source/phone_teleop.mdx

@@ -0,0 +1,191 @@
+# Phone
+
+Use your phone (iOS or Android) to control your robot.
+
+**In this guide you'll learn:**
+
+- How to connect an iOS/Android phone
+- How phone pose is mapped to robot end‑effector (EE) targets
+- How to tweak safety limits, gripper control, and IK settings
+
+To use phone to control your robot, install the relevant dependencies with:
+
+```bash
+pip install lerobot[phone]
+```
+
+## Get started
+
+### Supported platforms
+
+- iOS: Uses the HEBI Mobile I/O app (ARKit pose + buttons). Download the app first, open it and the examples will discover it on your network and stream the phone pose and inputs.
+- Android: Uses the `teleop` package (WebXR). When you start the Python process, it prints a local URL. Open the link on your phone, tap Start, then use Move to stream pose.
+
+Links:
+
+- Android WebXR library: [`teleop` on PyPI](https://pypi.org/project/teleop/)
+- iOS app: [HEBI Mobile I/O](https://docs.hebi.us/tools.html#mobile-io)
+
+### Phone orientation and controls
+
+- Orientation: hold the phone with the screen facing up and the top edge pointing in the same direction as the robot gripper. This ensures calibration aligns the phone’s frame with the robot frame so motion feels natural, see the image below for reference.
+- Enable/disable:
+  - iOS: Hold `B1` to enable teleoperation, release to stop. The first press captures a reference pose.
+  - Android: Press and hold the `Move` button, release to stop. The first press captures a reference pose.
+- Gripper control:
+  - iOS: Analog input `A3` controls the gripper as velocity input.
+  - Android: Buttons `A` and `B` act like increment/decrement (A opens, B closes). You can tune velocity in the `GripperVelocityToJoint` step.
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/phone_teleop.webp" alt="Phone teleop orientation" title="Phone teleop orientation" width="40%">
+
+### Step 1: Choose the platform
+
+Modify the examples to use `PhoneOS.IOS` or `PhoneOS.ANDROID` in `PhoneConfig`. The API is identical across platforms, only the input source differs. All examples are under `examples/` and have `phone_so100_*.py` variants.
+
+Teleoperation example:
+
+```36:43:examples/phone_so100_teleop.py
+from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+
+teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
+teleop_device = Phone(teleop_config)
+```
+
+### Step 2: Connect and calibrate
+
+When `Phone(teleop_config)` is created and `connect()` is called, calibration is prompted automatically. Hold the phone in the orientation described above, then:
+
+- iOS: press and hold `B1` to capture the reference pose.
+- Android: press `Move` button on the WebXR page to capture the reference pose.
+
+Why calibrate? We capture the current pose so subsequent poses are expressed in a robot aligned frame. When you again press the button to enable control, the position is recaptured to avoid drift when your phone is repositioned while it was disabled.
+
+### Step 3: Run an example
+
+Run on of the examples scripts to teleoperate, record a dataset, replay a dataset or evaluate a policy.
+
+All scripts assume you configured your robot (e.g., SO-100 follower) and set the correct serial port.
+
+Additionally you need to **copy the urdf of the robot to the examples folder**. For the examples in this tutorial (Using SO100/SO101) it is highly recommended to use the urdf in the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf)
+
+- Run this example to teleoperate:
+
+  ```bash
+  python examples/phone_to_so100/teleoperate.py
+  ```
+
+After running the example:
+
+- Android: after starting the script, open the printed local URL on your phone, tap Start, then press and hold Move.
+- iOS: open HEBI Mobile I/O first; B1 enables motion. A3 controls the gripper.
+
+Additionally you can customize mapping or safety limits by editing the processor steps shown in the examples. You can also remap inputs (e.g., use a different analog input) or adapt the pipeline to other robots (e.g., LeKiwi) by modifying the input and kinematics steps. More about this in the [Processors for Robots and Teleoperators](./processors_robots_teleop.mdx) guide.
+
+- Run this example to record a dataset, which saves absolute end effector observations and actions:
+
+  ```bash
+  python examples/phone_to_so100/record.py
+  ```
+
+- Run this example to replay recorded episodes:
+
+  ```bash
+  python examples/phone_to_so100/replay.py
+  ```
+
+- Run this example to evaluate a pretrained policy:
+
+  ```bash
+  python examples/phone_to_so100/evaluate.py
+  ```
+
+### Important pipeline steps and options
+
+- Kinematics are used in multiple steps. We use [Placo](https://github.com/Rhoban/placo) which is a wrapper around Pinocchio for handling our kinematics. We construct the kinematics object by passing the robot's URDF and target frame. We set `target_frame_name` to the gripper frame.
+
+  ```examples/phone_to_so100/teleoperate.py
+  kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+  )
+
+  ```
+
+- The `MapPhoneActionToRobotAction` step converts the calibrated phone pose and inputs into target deltas and gripper commands, below is shown what the step outputs.
+
+  ```src/lerobot/teleoperators/phone/phone_processor.py
+  action["enabled"] = enabled
+        action["target_x"] = -pos[1] if enabled else 0.0
+        action["target_y"] = pos[0] if enabled else 0.0
+        action["target_z"] = pos[2] if enabled else 0.0
+        action["target_wx"] = rotvec[1] if enabled else 0.0
+        action["target_wy"] = rotvec[0] if enabled else 0.0
+        action["target_wz"] = -rotvec[2] if enabled else 0.0
+        action["gripper_vel"] = gripper_vel  # Still send gripper action when disabled
+  ```
+
+- The `EEReferenceAndDelta` step converts target deltas to an absolute desired EE pose, storing a reference on enable, the `end_effector_step_sizes` are the step sizes for the EE pose and can be modified to change the motion speed.
+
+  ```examples/phone_to_so100/teleoperate.py
+  EEReferenceAndDelta(
+      kinematics=kinematics_solver,
+      end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+      motor_names=list(robot.bus.motors.keys()),
+      use_latched_reference=True,
+  ),
+  ```
+
+- The `EEBoundsAndSafety` step clamps EE motion to a workspace and checks for large ee step jumps to ensure safety. The `end_effector_bounds` are the bounds for the EE pose and can be modified to change the workspace. The `max_ee_step_m` are the step limits for the EE pose and can be modified to change the safety limits.
+
+  ```examples/phone_to_so100/teleoperate.py
+  EEBoundsAndSafety(
+      end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+      max_ee_step_m=0.10,
+  )
+  ```
+
+- The `GripperVelocityToJoint` step turns a velocity‑like gripper input into absolute gripper position using the current measured state. The `speed_factor` is the factor by which the velocity is multiplied.
+
+  ```examples/phone_to_so100/teleoperate.py
+  GripperVelocityToJoint(speed_factor=20.0)
+  ```
+
+#### Different IK initial guesses
+
+We use different IK initial guesses in the kinematic steps. As initial guess either the current measured joints or the previous IK solution is used.
+
+- Closed loop (used in record/eval): sets `initial_guess_current_joints=True` so IK starts from the measured joints each frame.
+
+  ```examples/phone_to_so100/record.py
+  InverseKinematicsEEToJoints(
+      kinematics=kinematics_solver,
+      motor_names=list(robot.bus.motors.keys()),
+      initial_guess_current_joints=True,  # closed loop
+  )
+  ```
+
+- Open loop (used in replay): sets `initial_guess_current_joints=False` so IK continues from the previous IK solution rather than the measured state. This preserves action stability when we replay without feedback.
+
+  ```examples/phone_to_so100/replay.py
+  InverseKinematicsEEToJoints(
+      kinematics=kinematics_solver,
+      motor_names=list(robot.bus.motors.keys()),
+      initial_guess_current_joints=False,  # open loop
+  )
+  ```
+
+### Pipeline steps explained
+
+- MapPhoneActionToRobotAction: converts calibrated phone pose and inputs into target deltas and a gripper command. Motion is gated by an enable signal (B1 on iOS, Move on Android).
+- EEReferenceAndDelta: latches a reference EE pose on enable and combines it with target deltas to produce an absolute desired EE pose each frame. When disabled, it keeps sending the last commanded pose.
+- EEBoundsAndSafety: clamps the EE pose to a workspace and rate‑limits jumps for safety. Also declares `action.ee.*` features.
+- InverseKinematicsEEToJoints: turns an EE pose into joint positions with IK. `initial_guess_current_joints=True` is recommended for closed‑loop control; set `False` for open‑loop replay for stability.
+- GripperVelocityToJoint: integrates a velocity‑like gripper input into an absolute gripper position using the current measured state.
+- ForwardKinematicsJointsToEE: computes `observation.state.ee.*` from observed joints for logging and training on EE state.
+
+### Troubleshooting
+
+- iOS not discovered: ensure HEBI Mobile I/O is open and your laptop/phone are on the same network.
+- Android URL not reachable: check local you used `https` instead of `http`, use the exact IP printed by the script and allow your browser to enter and ignore the certificate issue.
+- Motion feels inverted: adjust the sign flips in `MapPhoneActionToRobotAction` or swap axes to match your setup.

docs/source/processors_robots_teleop.mdx

@@ -0,0 +1,151 @@
+# Processors for Robots and Teleoperators
+
+This guide shows how to build and modify processing pipelines that connect teleoperators (e.g., phone) to robots and datasets. Pipelines standardize conversions between different action/observation spaces so you can swap teleops and robots without rewriting glue code.
+
+We use the Phone to SO‑100 follower examples for concreteness, but the same patterns apply to other robots.
+
+**What you'll learn**
+
+- Absolute vs. relative EE control: What each means, trade‑offs, and how to choose for your task.
+- Three-pipeline pattern: How to map teleop actions → dataset actions → robot commands, and robot observations → dataset observations.
+- Adapters (`to_transition` / `to_output`): How these convert raw dicts to `EnvTransition` and back to reduce boilerplate.
+- Dataset feature contracts: How steps declare features via `transform_features(...)`, and how to aggregate/merge them for recording.
+- Choosing a representation: When to store joints, absolute EE poses, or relative EE deltas—and how that affects training.
+- Pipeline customization guidance: How to swap robots/URDFs safely and tune bounds, step sizes, and options like IK initialization.
+
+### Absolute vs relative EE control
+
+The examples in this guide use absolute end effector (EE) poses because they are easy to reason about. In practice, relative EE deltas or joint position are often preferred as learning features.
+
+With processors, you choose the learning features you want to use for your policy. This could be joints positions/velocities, absolute EE, or relative EE positions. You can also choose to store other features, such as joint torques, motor currents, etc.
+
+## Three pipelines
+
+We often compose three pipelines. Depending on your setup, some can be empty if action and observation spaces already match.
+Each of these pipelines handle different conversions between different action and observation spaces. Below is a quick explanation of each pipeline.
+
+1. Pipeline 1: Teleop action space → dataset action space (phone pose → EE targets)
+2. Pipeline 2: Dataset action space → robot command space (EE targets → joints)
+3. Pipeline 3: Robot observation space → dataset observation space (joints → EE pose)
+
+Below is an example of the three pipelines that we use in the phone to SO-100 follower examples:
+
+```69:90:examples/phone_so100_record.py
+phone_to_robot_ee_pose_processor = RobotProcessorPipeline[RobotAction, RobotAction]( # teleop -> dataset action
+    steps=[
+        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+        EEReferenceAndDelta(
+            kinematics=kinematics_solver, end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5}, motor_names=list(robot.bus.motors.keys()),
+        ),
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]}, max_ee_step_m=0.20,
+        ),
+        GripperVelocityToJoint(),
+    ],
+    to_transition=robot_action_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+robot_ee_to_joints_processor = RobotProcessorPipeline[RobotAction, RobotAction]( # dataset action -> robot
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()), initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+robot_joints_to_ee_pose = RobotProcessorPipeline[RobotObservation, RobotObservation]( # robot obs -> dataset obs
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=observation_to_transition,
+    to_output=transition_to_observation,
+)
+```
+
+## Why to_transition / to_output
+
+To convert from robot/teleoperator to pipeline and back, we use the `to_transition` and `to_output` pipeline adapters.
+They standardize conversions to reduce boilerplate code, and form the bridge between the robot and teleoperators raw dictionaries and the pipeline’s `EnvTransition` format.
+In the phone to SO-100 follower examples we use the following adapters:
+
+- `robot_action_to_transition`: transforms the teleop action dict to a pipeline transition.
+- `transition_to_robot_action`: transforms the pipeline transition to a robot action dict.
+- `observation_to_transition`: transforms the robot observation dict to a pipeline transition.
+- `transition_to_observation`: transforms the pipeline transition to a observation dict.
+
+Checkout [src/lerobot/processor/converters.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/converters.py) for more details.
+
+## Dataset feature contracts
+
+Dataset features are determined by the keys saved in the dataset. Each step can declare what features it modifies in a contract called `transform_features(...)`. Once you build a processor, the processor can then aggregate all of these features with `aggregate_pipeline_dataset_features()` and merge multiple feature dicts with `combine_feature_dicts(...)`.
+
+Below is and example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:
+
+```src/lerobot/robots/so100_follower/robot_kinematic_processor.py
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        # We only use the ee pose in the dataset, so we don't need the joint positions
+        for n in self.motor_names:
+            features[PipelineFeatureType.ACTION].pop(f"{n}.pos", None)
+        # We specify the dataset features of this step that we want to be stored in the dataset
+        for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]:
+            features[PipelineFeatureType.ACTION][f"ee.{k}"] = PolicyFeature(
+                type=FeatureType.STATE, shape=(1,)
+            )
+        return features
+```
+
+Here we declare what PolicyFeatures we modify in this step, so we know what features we can expect when we run the processor. These features can then be aggregated and used to create the dataset features.
+
+Below is an example of how we aggregate and merge features in the phone to SO-100 record example:
+
+```121:145:examples/phone_so100_record.py
+features=combine_feature_dicts(
+        # Run the feature contract of the pipelines
+        # This tells you how the features would look like after the pipeline steps
+        aggregate_pipeline_dataset_features(
+            pipeline=phone_to_robot_ee_pose_processor,
+            initial_features=create_initial_features(action=phone.action_features), # <- Action features we can expect, these come from our teleop device (phone) and action processor
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_joints_to_ee_pose,
+            initial_features=create_initial_features(observation=robot.observation_features), # <- Observation features we can expect, these come from our robot and observation processor
+            use_videos=True,
+            patterns=["observation.state.ee"], # <- Here you could optionally filter the features we want to store in the dataset, with a specific pattern
+
+        ),
+    ),
+```
+
+How it works:
+
+- `aggregate_pipeline_dataset_features(...)`: applies `transform_features` across the pipeline and filters by patterns (images included when `use_videos=True`, and state features included when `patterns` is specified).
+- `combine_feature_dicts(...)`: combine multiple feature dicts.
+- Recording with `record_loop(...)` uses `build_dataset_frame(...)` to build frames consistent with `dataset.features` before we call `add_frame(...)` to add the frame to the dataset.
+
+## Guidance when customizing robot pipelines
+
+You can store any of the following features as your action/observation space:
+
+- Joint positions
+- Absolute EE poses
+- Relative EE deltas
+- Other features: joint velocity, torques, etc.
+
+Pick what you want to use for your policy action and observation space and configure/modify the pipelines and steps accordingly.
+
+### Different robots
+
+- You can easily reuse pipelines, for example to use another robot with phone teleop, modify the examples and swap the robot `RobotKinematics` (URDF) and `motor_names` to use your own robot with Phone teleop. Additionally you should ensure `target_frame_name` points to your gripper/wrist.
+
+### Safety first
+
+- When changing pipelines, start with tight bounds, implement safety steps when working with real robots.
+- Its advised to start with simulation first and then move to real robots.
+
+Thats it! We hope this guide helps you get started with customizing your robot pipelines, If you run into any issues at any point, jump into our [Discord community](https://discord.com/invite/s3KuuzsPFb) for support.

docs/source/smolvla.mdx

@@ -29,7 +29,7 @@ SmolVLA is Hugging Face’s lightweight foundation model for robotics. Designed
 ## Collect a dataset
 
 SmolVLA is a base model, so fine-tuning on your own data is required for optimal performance in your setup.
-We recommend recording ~50 episodes of your task as a starting point. Follow our guide to get started: [Recording a Dataset](https://huggingface.co/docs/lerobot/getting_started_real_world_robot#record-a-dataset)
+We recommend recording ~50 episodes of your task as a starting point. Follow our guide to get started: [Recording a Dataset](./il_robots)
 
 <Tip>
 
@@ -93,7 +93,7 @@ lerobot-train --help
 
 ## Evaluate the finetuned model and run it in real-time
 
-Similarly for when recording an episode, it is recommended that you are logged in to the HuggingFace Hub. You can follow the corresponding steps: [Record a dataset](./getting_started_real_world_robot#record-a-dataset).
+Similarly for when recording an episode, it is recommended that you are logged in to the HuggingFace Hub. You can follow the corresponding steps: [Record a dataset](./il_robots).
 Once you are logged in, you can run inference in your setup by doing:
 
 ```bash

docs/source/so100.mdx

@@ -634,7 +634,7 @@ leader.disconnect()
 </hfoption>
 </hfoptions>
 
-Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./getting_started_real_world_robot)
+Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./il_robots)
 
 > [!TIP]
 > If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).

docs/source/so101.mdx

@@ -430,7 +430,7 @@ leader.disconnect()
 </hfoption>
 </hfoptions>
 
-Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./getting_started_real_world_robot)
+Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./il_robots)
 
 > [!TIP]
 > If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).

examples/2_evaluate_pretrained_policy.py

@@ -1,139 +0,0 @@
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-This script demonstrates how to evaluate a pretrained policy from the HuggingFace Hub or from your local
-training outputs directory. In the latter case, you might want to run examples/3_train_policy.py first.
-
-It requires the installation of the 'gym_pusht' simulation environment. Install it by running:
-```bash
-pip install -e ".[pusht]"
-```
-"""
-
-from pathlib import Path
-
-import gym_pusht  # noqa: F401
-import gymnasium as gym
-import imageio
-import numpy
-import torch
-
-from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
-
-# Create a directory to store the video of the evaluation
-output_directory = Path("outputs/eval/example_pusht_diffusion")
-output_directory.mkdir(parents=True, exist_ok=True)
-
-# Select your device
-device = "cuda"
-
-# Provide the [hugging face repo id](https://huggingface.co/lerobot/diffusion_pusht):
-pretrained_policy_path = "lerobot/diffusion_pusht"
-# OR a path to a local outputs/train folder.
-# pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
-
-policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
-
-# Initialize evaluation environment to render two observation types:
-# an image of the scene and state/position of the agent. The environment
-# also automatically stops running after 300 interactions/steps.
-env = gym.make(
-    "gym_pusht/PushT-v0",
-    obs_type="pixels_agent_pos",
-    max_episode_steps=300,
-)
-
-# We can verify that the shapes of the features expected by the policy match the ones from the observations
-# produced by the environment
-print(policy.config.input_features)
-print(env.observation_space)
-
-# Similarly, we can check that the actions produced by the policy will match the actions expected by the
-# environment
-print(policy.config.output_features)
-print(env.action_space)
-
-# Reset the policy and environments to prepare for rollout
-policy.reset()
-numpy_observation, info = env.reset(seed=42)
-
-# Prepare to collect every rewards and all the frames of the episode,
-# from initial state to final state.
-rewards = []
-frames = []
-
-# Render frame of the initial state
-frames.append(env.render())
-
-step = 0
-done = False
-while not done:
-    # Prepare observation for the policy running in Pytorch
-    state = torch.from_numpy(numpy_observation["agent_pos"])
-    image = torch.from_numpy(numpy_observation["pixels"])
-
-    # Convert to float32 with image from channel first in [0,255]
-    # to channel last in [0,1]
-    state = state.to(torch.float32)
-    image = image.to(torch.float32) / 255
-    image = image.permute(2, 0, 1)
-
-    # Send data tensors from CPU to GPU
-    state = state.to(device, non_blocking=True)
-    image = image.to(device, non_blocking=True)
-
-    # Add extra (empty) batch dimension, required to forward the policy
-    state = state.unsqueeze(0)
-    image = image.unsqueeze(0)
-
-    # Create the policy input dictionary
-    observation = {
-        "observation.state": state,
-        "observation.image": image,
-    }
-
-    # Predict the next action with respect to the current observation
-    with torch.inference_mode():
-        action = policy.select_action(observation)
-
-    # Prepare the action for the environment
-    numpy_action = action.squeeze(0).to("cpu").numpy()
-
-    # Step through the environment and receive a new observation
-    numpy_observation, reward, terminated, truncated, info = env.step(numpy_action)
-    print(f"{step=} {reward=} {terminated=}")
-
-    # Keep track of all the rewards and frames
-    rewards.append(reward)
-    frames.append(env.render())
-
-    # The rollout is considered done when the success state is reached (i.e. terminated is True),
-    # or the maximum number of iterations is reached (i.e. truncated is True)
-    done = terminated | truncated | done
-    step += 1
-
-if terminated:
-    print("Success!")
-else:
-    print("Failure!")
-
-# Get the speed of environment (i.e. its number of frames per second).
-fps = env.metadata["render_fps"]
-
-# Encode all frames into a mp4 video.
-video_path = output_directory / "rollout.mp4"
-imageio.mimsave(str(video_path), numpy.stack(frames), fps=fps)
-
-print(f"Video of the evaluation is available in '{video_path}'.")

examples/4_train_policy_with_script.md

@@ -1,311 +0,0 @@
-This tutorial will explain the training script, how to use it, and particularly how to configure everything needed for the training run.
-
-> **Note:** The following assumes you're running these commands on a machine equipped with a cuda GPU. If you don't have one (or if you're using a Mac), you can add `--policy.device=cpu` (`--policy.device=mps` respectively). However, be advised that the code executes much slower on cpu.
-
-## The training script
-
-LeRobot offers a training script at [`lerobot/scripts/train.py`](../src/lerobot/scripts/train.py). At a high level it does the following:
-
-- Initialize/load a configuration for the following steps using.
-- Instantiates a dataset.
-- (Optional) Instantiates a simulation environment corresponding to that dataset.
-- Instantiates a policy.
-- Runs a standard training loop with forward pass, backward pass, optimization step, and occasional logging, evaluation (of the policy on the environment), and checkpointing.
-
-## Overview of the configuration system
-
-In the training script, the main function `train` expects a `TrainPipelineConfig` object:
-
-<!-- prettier-ignore-start -->
-```python
-# train.py
-@parser.wrap()
-def train(cfg: TrainPipelineConfig):
-```
-<!-- prettier-ignore-end -->
-
-You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../src/lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)
-
-When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated to this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)
-
-Let's have a look at a simplified example. Amongst other attributes, the training config has the following attributes:
-
-<!-- prettier-ignore-start -->
-```python
-@dataclass
-class TrainPipelineConfig:
-    dataset: DatasetConfig
-    env: envs.EnvConfig | None = None
-    policy: PreTrainedConfig | None = None
-```
-<!-- prettier-ignore-end -->
-
-in which `DatasetConfig` for example is defined as such:
-
-<!-- prettier-ignore-start -->
-```python
-@dataclass
-class DatasetConfig:
-    repo_id: str
-    episodes: list[int] | None = None
-    video_backend: str = "pyav"
-```
-<!-- prettier-ignore-end -->
-
-This creates a hierarchical relationship where, for example assuming we have a `cfg` instance of `TrainPipelineConfig`, we can access the `repo_id` value with `cfg.dataset.repo_id`.
-From the command line, we can specify this value by using a very similar syntax `--dataset.repo_id=repo/id`.
-
-By default, every field takes its default value specified in the dataclass. If a field doesn't have a default value, it needs to be specified either from the command line or from a config file – which path is also given in the command line (more in this below). In the example above, the `dataset` field doesn't have a default value which means it must be specified.
-
-## Specifying values from the CLI
-
-Let's say that we want to train [Diffusion Policy](../src/lerobot/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
-
-```bash
-lerobot-train \
-    --dataset.repo_id=lerobot/pusht \
-    --policy.type=diffusion \
-    --env.type=pusht
-```
-
-Let's break this down:
-
-- To specify the dataset, we just need to specify its `repo_id` on the hub which is the only required argument in the `DatasetConfig`. The rest of the fields have default values and in this case we are fine with those so we can just add the option `--dataset.repo_id=lerobot/pusht`.
-- To specify the policy, we can just select diffusion policy using `--policy` appended with `.type`. Here, `.type` is a special argument which allows us to select config classes inheriting from `draccus.ChoiceRegistry` and that have been decorated with the `register_subclass()` method. To have a better explanation of this feature, have a look at this [Draccus demo](https://github.com/dlwh/draccus?tab=readme-ov-file#more-flexible-configuration-with-choice-types). In our code, we use this mechanism mainly to select policies, environments, robots, and some other components like optimizers. The policies available to select are located in [lerobot/policies](../src/lerobot/policies)
-- Similarly, we select the environment with `--env.type=pusht`. The different environment configs are available in [`lerobot/envs/configs.py`](../src/lerobot/envs/configs.py)
-
-Let's see another example. Let's say you've been training [ACT](../src/lerobot/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
-
-```bash
-lerobot-train \
-    --policy.type=act \
-    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
-    --env.type=aloha \
-    --output_dir=outputs/train/act_aloha_insertion
-```
-
-> Notice we added `--output_dir` to explicitly tell where to write outputs from this run (checkpoints, training state, configs etc.). This is not mandatory and if you don't specify it, a default directory will be created from the current date and time, env.type and policy.type. This will typically look like `outputs/train/2025-01-24/16-10-05_aloha_act`.
-
-We now want to train a different policy for aloha on another task. We'll change the dataset and use [lerobot/aloha_sim_transfer_cube_human](https://huggingface.co/datasets/lerobot/aloha_sim_transfer_cube_human) instead. Of course, we also need to change the task of the environment as well to match this other task.
-Looking at the [`AlohaEnv`](../src/lerobot/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
-
-```bash
-lerobot-train \
-    --policy.type=act \
-    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
-    --env.type=aloha \
-    --env.task=AlohaTransferCube-v0 \
-    --output_dir=outputs/train/act_aloha_transfer
-```
-
-## Loading from a config file
-
-Now, let's assume that we want to reproduce the run just above. That run has produced a `train_config.json` file in its checkpoints, which serializes the `TrainPipelineConfig` instance it used:
-
-```json
-{
-    "dataset": {
-        "repo_id": "lerobot/aloha_sim_transfer_cube_human",
-        "episodes": null,
-        ...
-    },
-    "env": {
-        "type": "aloha",
-        "task": "AlohaTransferCube-v0",
-        "fps": 50,
-        ...
-    },
-    "policy": {
-        "type": "act",
-        "n_obs_steps": 1,
-        ...
-    },
-    ...
-}
-```
-
-We can then simply load the config values from this file using:
-
-```bash
-lerobot-train \
-    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
-    --output_dir=outputs/train/act_aloha_transfer_2
-```
-
-`--config_path` is also a special argument which allows to initialize the config from a local config file. It can point to a directory that contains `train_config.json` or to the config file itself directly.
-
-Similarly to Hydra, we can still override some parameters in the CLI if we want to, e.g.:
-
-```bash
-lerobot-train \
-    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
-    --output_dir=outputs/train/act_aloha_transfer_2
-    --policy.n_action_steps=80
-```
-
-> Note: While `--output_dir` is not required in general, in this case we need to specify it since it will otherwise take the value from the `train_config.json` (which is `outputs/train/act_aloha_transfer`). In order to prevent accidental deletion of previous run checkpoints, we raise an error if you're trying to write in an existing directory. This is not the case when resuming a run, which is what you'll learn next.
-
-`--config_path` can also accept the repo_id of a repo on the hub that contains a `train_config.json` file, e.g. running:
-
-```bash
-lerobot-train --config_path=lerobot/diffusion_pusht
-```
-
-will start a training run with the same configuration used for training [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)
-
-## Resume training
-
-Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to do that here.
-
-Let's reuse the command from the previous run and add a few more options:
-
-```bash
-lerobot-train \
-    --policy.type=act \
-    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
-    --env.type=aloha \
-    --env.task=AlohaTransferCube-v0 \
-    --log_freq=25 \
-    --save_freq=100 \
-    --output_dir=outputs/train/run_resumption
-```
-
-Here we've taken care to set up the log frequency and checkpointing frequency to low numbers so we can showcase resumption. You should be able to see some logging and have a first checkpoint within 1 minute (depending on hardware). Wait for the first checkpoint to happen, you should see a line that looks like this in your terminal:
-
-```
-INFO 2025-01-24 16:10:56 ts/train.py:263 Checkpoint policy after step 100
-```
-
-Now let's simulate a crash by killing the process (hit `ctrl`+`c`). We can then simply resume this run from the last checkpoint available with:
-
-```bash
-lerobot-train \
-    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
-    --resume=true
-```
-
-You should see from the logging that your training picks up from where it left off.
-
-Another reason for which you might want to resume a run is simply to extend training and add more training steps. The number of training steps is set by the option `--steps`, which is 100 000 by default.
-You could double the number of steps of the previous run with:
-
-```bash
-lerobot-train \
-    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
-    --resume=true \
-    --steps=200000
-```
-
-## Outputs of a run
-
-In the output directory, there will be a folder called `checkpoints` with the following structure:
-
-```bash
-outputs/train/run_resumption/checkpoints
-├── 000100  # checkpoint_dir for training step 100
-│   ├── pretrained_model/
-│   │   ├── config.json  # policy config
-│   │   ├── model.safetensors  # policy weights
-│   │   └── train_config.json  # train config
-│   └── training_state/
-│       ├── optimizer_param_groups.json  #  optimizer param groups
-│       ├── optimizer_state.safetensors  # optimizer state
-│       ├── rng_state.safetensors  # rng states
-│       ├── scheduler_state.json  # scheduler state
-│       └── training_step.json  # training step
-├── 000200
-└── last -> 000200  # symlink to the last available checkpoint
-```
-
-## Fine-tuning a pre-trained policy
-
-In addition to the features currently in Draccus, we've added a special `.path` argument for the policy, which allows to load a policy as you would with `PreTrainedPolicy.from_pretrained()`. In that case, `path` can be a local directory that contains a checkpoint or a repo_id pointing to a pretrained policy on the hub.
-
-For example, we could fine-tune a [policy pre-trained on the aloha transfer task](https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human) on the aloha insertion task. We can achieve this with:
-
-```bash
-lerobot-train \
-    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \
-    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
-    --env.type=aloha \
-    --env.task=AlohaInsertion-v0
-```
-
-When doing so, keep in mind that the features of the fine-tuning dataset would have to match the input/output features of the pretrained policy.
-
-## Typical logs and metrics
-
-When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you configured your run correctly. The final configuration will also be saved with the checkpoint.
-
-After that, you will see training log like this one:
-
-```
-INFO 2024-08-14 13:35:12 ts/train.py:192 step:0 smpl:64 ep:1 epch:0.00 loss:1.112 grdn:15.387 lr:2.0e-07 updt_s:1.738 data_s:4.774
-```
-
-or evaluation log:
-
-```
-INFO 2024-08-14 13:38:45 ts/train.py:226 step:100 smpl:6K ep:52 epch:0.25 ∑rwrd:20.693 success:0.0% eval_s:120.266
-```
-
-These logs will also be saved in wandb if `wandb.enable` is set to `true`. Here are the meaning of some abbreviations:
-
-- `smpl`: number of samples seen during training.
-- `ep`: number of episodes seen during training. An episode contains multiple samples in a complete manipulation task.
-- `epch`: number of time all unique samples are seen (epoch).
-- `grdn`: gradient norm.
-- `∑rwrd`: compute the sum of rewards in every evaluation episode and then take an average of them.
-- `success`: average success rate of eval episodes. Reward and success are usually different except for the sparsing reward setting, where reward=1 only when the task is completed successfully.
-- `eval_s`: time to evaluate the policy in the environment, in second.
-- `updt_s`: time to update the network parameters, in second.
-- `data_s`: time to load a batch of data, in second.
-
-Some metrics are useful for initial performance profiling. For example, if you find the current GPU utilization is low via the `nvidia-smi` command and `data_s` sometimes is too high, you may need to modify batch size or number of dataloading workers to accelerate dataloading. We also recommend [pytorch profiler](https://github.com/huggingface/lerobot?tab=readme-ov-file#improve-your-code-with-profiling) for detailed performance probing.
-
-## In short
-
-We'll summarize here the main use cases to remember from this tutorial.
-
-#### Train a policy from scratch – CLI
-
-```bash
-lerobot-train \
-    --policy.type=act \  # <- select 'act' policy
-    --env.type=pusht \  # <- select 'pusht' environment
-    --dataset.repo_id=lerobot/pusht  # <- train on this dataset
-```
-
-#### Train a policy from scratch - config file + CLI
-
-```bash
-lerobot-train \
-    --config_path=path/to/pretrained_model \  # <- can also be a repo_id
-    --policy.n_action_steps=80  # <- you may still override values
-```
-
-#### Resume/continue a training run
-
-```bash
-lerobot-train \
-    --config_path=checkpoint/pretrained_model/ \
-    --resume=true \
-    --steps=200000  # <- you can change some training parameters
-```
-
-#### Fine-tuning
-
-```bash
-lerobot-train \
-    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \  # <- can also be a local path to a checkpoint
-    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
-    --env.type=aloha \
-    --env.task=AlohaInsertion-v0
-```
-
----
-
-Now that you know the basics of how to train a policy, you might want to know how to apply this knowledge to actual robots, or how to record your own datasets and train policies on your specific task?
-If that's the case, head over to the next tutorial [`7_get_started_with_real_robot.md`](./7_get_started_with_real_robot.md).
-
-Or in the meantime, happy training! 🤗

examples/backward_compatibility/replay.py

@@ -44,6 +44,7 @@ from lerobot.robots import (  # noqa: F401
     so100_follower,
     so101_follower,
 )
+from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import busy_wait
 from lerobot.utils.utils import (
     init_logging,
@@ -78,16 +79,16 @@ def replay(cfg: ReplayConfig):
 
     robot = make_robot_from_config(cfg.robot)
     dataset = LeRobotDataset(cfg.dataset.repo_id, root=cfg.dataset.root, episodes=[cfg.dataset.episode])
-    actions = dataset.hf_dataset.select_columns("action")
+    actions = dataset.hf_dataset.select_columns(ACTION)
     robot.connect()
 
     log_say("Replaying episode", cfg.play_sounds, blocking=True)
     for idx in range(dataset.num_frames):
         start_episode_t = time.perf_counter()
 
-        action_array = actions[idx]["action"]
+        action_array = actions[idx][ACTION]
         action = {}
-        for i, name in enumerate(dataset.features["action"]["names"]):
+        for i, name in enumerate(dataset.features[ACTION]["names"]):
             key = f"{name.removeprefix('main_')}.pos"
             action[key] = action_array[i].item()
 

examples/dataset/load_lerobot_dataset.py

@@ -136,7 +136,7 @@ print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
 # PyTorch datasets.
 dataloader = torch.utils.data.DataLoader(
     dataset,
-    num_workers=0,
+    num_workers=4,
     batch_size=32,
     shuffle=True,
 )

examples/dataset/use_dataset_image_transforms.py

@@ -0,0 +1,177 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This example demonstrates how to use image transforms with LeRobot datasets for data augmentation during training.
+
+Image transforms are applied to camera frames to improve model robustness and generalization. They are applied
+at training time only, not during dataset recording, allowing you to experiment with different augmentations
+without re-recording data.
+"""
+
+import torch
+from torchvision.transforms import v2
+from torchvision.transforms.functional import to_pil_image
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.transforms import ImageTransformConfig, ImageTransforms, ImageTransformsConfig
+
+
+def save_image(tensor, filename):
+    """Helper function to save a tensor as an image file."""
+    if tensor.dim() == 3:  # [C, H, W]
+        if tensor.max() > 1.0:
+            tensor = tensor / 255.0
+        tensor = torch.clamp(tensor, 0.0, 1.0)
+        pil_image = to_pil_image(tensor)
+        pil_image.save(filename)
+        print(f"Saved: {filename}")
+    else:
+        print(f"Skipped {filename}: unexpected tensor shape {tensor.shape}")
+
+
+def example_1_default_transforms():
+    """Example 1: Use default transform configuration and save original vs transformed images"""
+    print("\n Example 1: Default Transform Configuration with Image Saving")
+
+    repo_id = "pepijn223/record_main_0"  # Example dataset
+
+    try:
+        # Load dataset without transforms (original)
+        dataset_original = LeRobotDataset(repo_id=repo_id)
+
+        # Load dataset with transforms enabled
+        transforms_config = ImageTransformsConfig(
+            enable=True,  # Enable transforms (disabled by default)
+            max_num_transforms=2,  # Apply up to 2 transforms per frame
+            random_order=False,  # Apply in standard order
+        )
+        dataset_with_transforms = LeRobotDataset(
+            repo_id=repo_id, image_transforms=ImageTransforms(transforms_config)
+        )
+
+        # Save original and transformed images for comparison
+        if len(dataset_original) > 0:
+            frame_idx = 0  # Use first frame
+            original_sample = dataset_original[frame_idx]
+            transformed_sample = dataset_with_transforms[frame_idx]
+
+            print(f"Saving comparison images (frame {frame_idx}):")
+
+            for cam_key in dataset_original.meta.camera_keys:
+                if cam_key in original_sample and cam_key in transformed_sample:
+                    cam_name = cam_key.replace(".", "_").replace("/", "_")
+
+                    # Save original and transformed images
+                    save_image(original_sample[cam_key], f"{cam_name}_original.png")
+                    save_image(transformed_sample[cam_key], f"{cam_name}_transformed.png")
+
+    except Exception as e:
+        print(f"Could not load dataset '{repo_id}': {e}")
+
+
+def example_2_custom_transforms():
+    """Example 2: Create custom transform configuration and save examples"""
+    print("\n Example 2: Custom Transform Configuration")
+
+    repo_id = "pepijn223/record_main_0"  # Example dataset
+
+    try:
+        # Create custom transform configuration with strong effects
+        custom_transforms_config = ImageTransformsConfig(
+            enable=True,
+            max_num_transforms=2,  # Apply up to 2 transforms per frame
+            random_order=True,  # Apply transforms in random order
+            tfs={
+                "brightness": ImageTransformConfig(
+                    weight=1.0,
+                    type="ColorJitter",
+                    kwargs={"brightness": (0.5, 1.5)},  # Strong brightness range
+                ),
+                "contrast": ImageTransformConfig(
+                    weight=1.0,  # Higher weight = more likely to be selected
+                    type="ColorJitter",
+                    kwargs={"contrast": (0.6, 1.4)},  # Strong contrast
+                ),
+                "sharpness": ImageTransformConfig(
+                    weight=0.5,  # Lower weight = less likely to be selected
+                    type="SharpnessJitter",
+                    kwargs={"sharpness": (0.2, 2.0)},  # Strong sharpness variation
+                ),
+            },
+        )
+
+        dataset_with_custom_transforms = LeRobotDataset(
+            repo_id=repo_id, image_transforms=ImageTransforms(custom_transforms_config)
+        )
+
+        # Save examples with strong transforms
+        if len(dataset_with_custom_transforms) > 0:
+            sample = dataset_with_custom_transforms[0]
+            print("Saving custom transform examples:")
+
+            for cam_key in dataset_with_custom_transforms.meta.camera_keys:
+                if cam_key in sample:
+                    cam_name = cam_key.replace(".", "_").replace("/", "_")
+                    save_image(sample[cam_key], f"{cam_name}_custom_transforms.png")
+
+    except Exception as e:
+        print(f"Could not load dataset '{repo_id}': {e}")
+
+
+def example_3_torchvision_transforms():
+    """Example 3: Use pure torchvision transforms and save examples"""
+    print("\n Example 3: Pure Torchvision Transforms")
+
+    repo_id = "pepijn223/record_main_0"  # Example dataset
+
+    try:
+        # Create torchvision transform pipeline
+        torchvision_transforms = v2.Compose(
+            [
+                v2.ColorJitter(brightness=0.3, contrast=0.3, saturation=0.3, hue=0.1),
+                v2.GaussianBlur(kernel_size=3, sigma=(0.1, 2.0)),
+                v2.RandomRotation(degrees=10),  # Small rotation
+            ]
+        )
+
+        dataset_with_torchvision = LeRobotDataset(repo_id=repo_id, image_transforms=torchvision_transforms)
+
+        # Save examples with torchvision transforms
+        if len(dataset_with_torchvision) > 0:
+            sample = dataset_with_torchvision[0]
+            print("Saving torchvision transform examples:")
+
+            for cam_key in dataset_with_torchvision.meta.camera_keys:
+                if cam_key in sample:
+                    cam_name = cam_key.replace(".", "_").replace("/", "_")
+                    save_image(sample[cam_key], f"{cam_name}_torchvision.png")
+
+    except Exception as e:
+        print(f"Could not load dataset '{repo_id}': {e}")
+
+
+def main():
+    """Run all examples"""
+    print("LeRobot Dataset Image Transforms Examples")
+
+    example_1_default_transforms()
+    example_2_custom_transforms()
+    example_3_torchvision_transforms()
+
+
+if __name__ == "__main__":
+    main()

examples/lekiwi/evaluate.py

@@ -1,31 +1,54 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.record import record_loop
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
+from lerobot.scripts.lerobot_record import record_loop
+from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import _init_rerun
+from lerobot.utils.visualization_utils import init_rerun
 
 NUM_EPISODES = 2
 FPS = 30
 EPISODE_TIME_SEC = 60
 TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
 
-# Create the robot and teleoperator configurations
+# Create the robot configuration & robot
 robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
+
 robot = LeKiwiClient(robot_config)
 
-policy = ACTPolicy.from_pretrained("<hf_username>/<policy_repo_id>")
+# Create policy
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
 # Configure the dataset features
-action_features = hw_to_dataset_features(robot.action_features, "action")
-obs_features = hw_to_dataset_features(robot.observation_features, "observation")
+action_features = hw_to_dataset_features(robot.action_features, ACTION)
+obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
 dataset_features = {**action_features, **obs_features}
 
 # Create the dataset
 dataset = LeRobotDataset.create(
-    repo_id="<hf_username>/<eval_dataset_repo_id>",
+    repo_id=HF_DATASET_ID,
     fps=FPS,
     features=dataset_features,
     robot_type=robot.name,
@@ -33,33 +56,52 @@ dataset = LeRobotDataset.create(
     image_writer_threads=4,
 )
 
+# Build Policy Processors
+preprocessor, postprocessor = make_pre_post_processors(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+    # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
+    preprocessor_overrides={"device_processor": {"device": str(policy.config.device)}},
+)
+
+# Connect the robot
 # To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
 robot.connect()
 
-_init_rerun(session_name="recording")
+# TODO(Steven): Update this example to use pipelines
+teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
 
+# Initialize the keyboard listener and rerun visualization
 listener, events = init_keyboard_listener()
+init_rerun(session_name="lekiwi_evaluate")
 
 if not robot.is_connected:
     raise ValueError("Robot is not connected!")
 
+print("Starting evaluate loop...")
 recorded_episodes = 0
 while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
     log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
 
-    # Run the policy inference loop
+    # Main record loop
     record_loop(
         robot=robot,
         events=events,
         fps=FPS,
         policy=policy,
+        preprocessor=preprocessor,  # Pass the pre and post policy processors
+        postprocessor=postprocessor,
         dataset=dataset,
         control_time_s=EPISODE_TIME_SEC,
         single_task=TASK_DESCRIPTION,
         display_data=True,
+        teleop_action_processor=teleop_action_processor,
+        robot_action_processor=robot_action_processor,
+        robot_observation_processor=robot_observation_processor,
     )
 
-    # Logic for reset env
+    # Reset the environment if not stopping or re-recording
     if not events["stop_recording"] and (
         (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
     ):
@@ -71,6 +113,9 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
             control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
+            teleop_action_processor=teleop_action_processor,
+            robot_action_processor=robot_action_processor,
+            robot_observation_processor=robot_observation_processor,
         )
 
     if events["rerecord_episode"]:
@@ -80,11 +125,12 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
         dataset.clear_episode_buffer()
         continue
 
+    # Save episode
     dataset.save_episode()
     recorded_episodes += 1
 
-# Upload to hub and clean up
-dataset.push_to_hub()
-
+# Clean up
+log_say("Stop recording")
 robot.disconnect()
 listener.stop()
+dataset.push_to_hub()

examples/lekiwi/record.py

@@ -1,37 +1,60 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
-from lerobot.record import record_loop
+from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
 from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient
+from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.keyboard import KeyboardTeleop, KeyboardTeleopConfig
 from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
+from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import _init_rerun
+from lerobot.utils.visualization_utils import init_rerun
 
-NUM_EPISODES = 3
+NUM_EPISODES = 2
 FPS = 30
 EPISODE_TIME_SEC = 30
 RESET_TIME_SEC = 10
 TASK_DESCRIPTION = "My task description"
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
 
 # Create the robot and teleoperator configurations
 robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
 leader_arm_config = SO100LeaderConfig(port="/dev/tty.usbmodem585A0077581", id="my_awesome_leader_arm")
 keyboard_config = KeyboardTeleopConfig()
 
+# Initialize the robot and teleoperator
 robot = LeKiwiClient(robot_config)
 leader_arm = SO100Leader(leader_arm_config)
 keyboard = KeyboardTeleop(keyboard_config)
 
+# TODO(Steven): Update this example to use pipelines
+teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+
 # Configure the dataset features
-action_features = hw_to_dataset_features(robot.action_features, "action")
-obs_features = hw_to_dataset_features(robot.observation_features, "observation")
+action_features = hw_to_dataset_features(robot.action_features, ACTION)
+obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
 dataset_features = {**action_features, **obs_features}
 
 # Create the dataset
 dataset = LeRobotDataset.create(
-    repo_id="<hf_username>/<dataset_repo_id>",
+    repo_id=HF_REPO_ID,
     fps=FPS,
     features=dataset_features,
     robot_type=robot.name,
@@ -39,23 +62,25 @@ dataset = LeRobotDataset.create(
     image_writer_threads=4,
 )
 
+# Connect the robot and teleoperator
 # To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
 robot.connect()
 leader_arm.connect()
 keyboard.connect()
 
-_init_rerun(session_name="lekiwi_record")
-
+# Initialize the keyboard listener and rerun visualization
 listener, events = init_keyboard_listener()
+init_rerun(session_name="lekiwi_record")
 
 if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
-    raise ValueError("Robot, leader arm of keyboard is not connected!")
+    raise ValueError("Robot or teleop is not connected!")
 
+print("Starting record loop...")
 recorded_episodes = 0
 while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
     log_say(f"Recording episode {recorded_episodes}")
 
-    # Run the record loop
+    # Main record loop
     record_loop(
         robot=robot,
         events=events,
@@ -65,9 +90,12 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
         control_time_s=EPISODE_TIME_SEC,
         single_task=TASK_DESCRIPTION,
         display_data=True,
+        teleop_action_processor=teleop_action_processor,
+        robot_action_processor=robot_action_processor,
+        robot_observation_processor=robot_observation_processor,
     )
 
-    # Logic for reset env
+    # Reset the environment if not stopping or re-recording
     if not events["stop_recording"] and (
         (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
     ):
@@ -80,6 +108,9 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
             control_time_s=RESET_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
+            teleop_action_processor=teleop_action_processor,
+            robot_action_processor=robot_action_processor,
+            robot_observation_processor=robot_observation_processor,
         )
 
     if events["rerecord_episode"]:
@@ -89,13 +120,14 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
         dataset.clear_episode_buffer()
         continue
 
+    # Save episode
     dataset.save_episode()
     recorded_episodes += 1
 
-# Upload to hub and clean up
-dataset.push_to_hub()
-
+# Clean up
+log_say("Stop recording")
 robot.disconnect()
 leader_arm.disconnect()
 keyboard.disconnect()
 listener.stop()
+dataset.push_to_hub()

examples/lekiwi/replay.py

@@ -1,32 +1,60 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import time
 
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
 from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient
+from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import busy_wait
 from lerobot.utils.utils import log_say
 
 EPISODE_IDX = 0
 
+# Initialize the robot config
 robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
+
+# Initialize the robot
 robot = LeKiwiClient(robot_config)
 
+# Fetch the dataset to replay
 dataset = LeRobotDataset("<hf_username>/<dataset_repo_id>", episodes=[EPISODE_IDX])
-actions = dataset.hf_dataset.select_columns("action")
+# Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
+episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
+actions = episode_frames.select_columns(ACTION)
 
+# Connect to the robot
 robot.connect()
 
 if not robot.is_connected:
     raise ValueError("Robot is not connected!")
 
+print("Starting replay loop...")
 log_say(f"Replaying episode {EPISODE_IDX}")
-for idx in range(dataset.num_frames):
+for idx in range(len(episode_frames)):
     t0 = time.perf_counter()
 
+    # Get recorded action from dataset
     action = {
-        name: float(actions[idx]["action"][i]) for i, name in enumerate(dataset.features["action"]["names"])
+        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
     }
-    robot.send_action(action)
+
+    # Send action to robot
+    _ = robot.send_action(action)
 
     busy_wait(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
 

examples/lekiwi/teleoperate.py

@@ -1,10 +1,26 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import time
 
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.teleoperators.keyboard.teleop_keyboard import KeyboardTeleop, KeyboardTeleopConfig
 from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
 from lerobot.utils.robot_utils import busy_wait
-from lerobot.utils.visualization_utils import _init_rerun, log_rerun_data
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 FPS = 30
 
@@ -13,35 +29,44 @@ robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="my_lekiwi")
 teleop_arm_config = SO100LeaderConfig(port="/dev/tty.usbmodem585A0077581", id="my_awesome_leader_arm")
 keyboard_config = KeyboardTeleopConfig(id="my_laptop_keyboard")
 
+# Initialize the robot and teleoperator
 robot = LeKiwiClient(robot_config)
 leader_arm = SO100Leader(teleop_arm_config)
 keyboard = KeyboardTeleop(keyboard_config)
 
+# Connect to the robot and teleoperator
 # To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
 robot.connect()
 leader_arm.connect()
 keyboard.connect()
 
-_init_rerun(session_name="lekiwi_teleop")
+# Init rerun viewer
+init_rerun(session_name="lekiwi_teleop")
 
 if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
-    raise ValueError("Robot, leader arm of keyboard is not connected!")
+    raise ValueError("Robot or teleop is not connected!")
 
+print("Starting teleop loop...")
 while True:
     t0 = time.perf_counter()
 
+    # Get robot observation
     observation = robot.get_observation()
 
+    # Get teleop action
+    # Arm
     arm_action = leader_arm.get_action()
     arm_action = {f"arm_{k}": v for k, v in arm_action.items()}
-
+    # Keyboard
     keyboard_keys = keyboard.get_action()
     base_action = robot._from_keyboard_to_base_action(keyboard_keys)
 
-    log_rerun_data(observation, {**arm_action, **base_action})
-
     action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
 
-    robot.send_action(action)
+    # Send action to robot
+    _ = robot.send_action(action)
+
+    # Visualize
+    log_rerun_data(observation=observation, action=action)
 
     busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))

examples/phone_to_so100/evaluate.py

@@ -0,0 +1,197 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.processor import (
+    RobotAction,
+    RobotObservation,
+    RobotProcessorPipeline,
+    make_default_teleop_action_processor,
+)
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    ForwardKinematicsJointsToEE,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.scripts.lerobot_record import record_loop
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import init_rerun
+
+NUM_EPISODES = 5
+FPS = 30
+EPISODE_TIME_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
+
+# Create the robot configuration & robot
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem58760434471",
+    id="my_awesome_follower_arm",
+    cameras=camera_config,
+    use_degrees=True,
+)
+
+robot = SO100Follower(robot_config)
+
+# Create policy
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert EE action to joints action
+robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Build pipeline to convert joints observation to EE observation
+robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=observation_to_transition,
+    to_output=transition_to_observation,
+)
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_DATASET_ID,
+    fps=FPS,
+    features=combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_joints_to_ee_pose_processor,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=True,
+        ),
+        # User for now should be explicit on the feature keys that were used for record
+        # Alternatively, the user can pass the processor step that has the right features
+        aggregate_pipeline_dataset_features(
+            pipeline=make_default_teleop_action_processor(),
+            initial_features=create_initial_features(
+                action={
+                    f"ee.{k}": PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+                    for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]
+                }
+            ),
+            use_videos=True,
+        ),
+    ),
+    robot_type=robot.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+# Build Policy Processors
+preprocessor, postprocessor = make_pre_post_processors(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+    # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
+    preprocessor_overrides={"device_processor": {"device": str(policy.config.device)}},
+)
+
+# Connect the robot
+robot.connect()
+
+# Initialize the keyboard listener and rerun visualization
+listener, events = init_keyboard_listener()
+init_rerun(session_name="phone_so100_evaluate")
+
+if not robot.is_connected:
+    raise ValueError("Robot is not connected!")
+
+print("Starting evaluate loop...")
+episode_idx = 0
+for episode_idx in range(NUM_EPISODES):
+    log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    # Main record loop
+    record_loop(
+        robot=robot,
+        events=events,
+        fps=FPS,
+        policy=policy,
+        preprocessor=preprocessor,  # Pass the pre and post policy processors
+        postprocessor=postprocessor,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        teleop_action_processor=make_default_teleop_action_processor(),
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose_processor,
+    )
+
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
+        log_say("Reset the environment")
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            control_time_s=EPISODE_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=make_default_teleop_action_processor(),
+            robot_action_processor=robot_ee_to_joints_processor,
+            robot_observation_processor=robot_joints_to_ee_pose_processor,
+        )
+
+    if events["rerecord_episode"]:
+        log_say("Re-record episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
+
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
+
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+listener.stop()
+dataset.push_to_hub()

examples/phone_to_so100/record.py

@@ -0,0 +1,203 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    EEBoundsAndSafety,
+    EEReferenceAndDelta,
+    ForwardKinematicsJointsToEE,
+    GripperVelocityToJoint,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.scripts.lerobot_record import record_loop
+from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
+from lerobot.teleoperators.phone.teleop_phone import Phone
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import init_rerun
+
+NUM_EPISODES = 2
+FPS = 30
+EPISODE_TIME_SEC = 60
+RESET_TIME_SEC = 30
+TASK_DESCRIPTION = "My task description"
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+# Create the robot and teleoperator configurations
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411",
+    id="my_awesome_follower_arm",
+    cameras=camera_config,
+    use_degrees=True,
+)
+teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
+
+# Initialize the robot and teleoperator
+robot = SO100Follower(robot_config)
+phone = Phone(teleop_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert phone action to EE action
+phone_to_robot_ee_pose_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+        EEReferenceAndDelta(
+            kinematics=kinematics_solver,
+            end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+            motor_names=list(robot.bus.motors.keys()),
+            use_latched_reference=True,
+        ),
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.20,
+        ),
+        GripperVelocityToJoint(speed_factor=20.0),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Build pipeline to convert EE action to joints action
+robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Build pipeline to convert joint observation to EE observation
+robot_joints_to_ee_pose = RobotProcessorPipeline[RobotObservation, RobotObservation](
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=observation_to_transition,
+    to_output=transition_to_observation,
+)
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_REPO_ID,
+    fps=FPS,
+    features=combine_feature_dicts(
+        # Run the feature contract of the pipelines
+        # This tells you how the features would look like after the pipeline steps
+        aggregate_pipeline_dataset_features(
+            pipeline=phone_to_robot_ee_pose_processor,
+            initial_features=create_initial_features(action=phone.action_features),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_joints_to_ee_pose,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=True,
+        ),
+    ),
+    robot_type=robot.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+# Connect the robot and teleoperator
+robot.connect()
+phone.connect()
+
+# Initialize the keyboard listener and rerun visualization
+listener, events = init_keyboard_listener()
+init_rerun(session_name="phone_so100_record")
+
+if not robot.is_connected or not phone.is_connected:
+    raise ValueError("Robot or teleop is not connected!")
+
+
+print("Starting record loop. Move your phone to teleoperate the robot...")
+episode_idx = 0
+while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+    log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    # Main record loop
+    record_loop(
+        robot=robot,
+        events=events,
+        fps=FPS,
+        teleop=phone,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        teleop_action_processor=phone_to_robot_ee_pose_processor,
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose,
+    )
+
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+        log_say("Reset the environment")
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            teleop=phone,
+            control_time_s=RESET_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=phone_to_robot_ee_pose_processor,
+            robot_action_processor=robot_ee_to_joints_processor,
+            robot_observation_processor=robot_joints_to_ee_pose,
+        )
+
+    if events["rerecord_episode"]:
+        log_say("Re-recording episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
+
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
+
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+phone.disconnect()
+listener.stop()
+dataset.push_to_hub()

examples/phone_to_so100/replay.py

@@ -0,0 +1,100 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    robot_action_observation_to_transition,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.utils.constants import ACTION
+from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.utils import log_say
+
+EPISODE_IDX = 0
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+# Initialize the robot config
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+)
+
+# Initialize the robot
+robot = SO100Follower(robot_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert EE action to joints action
+robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=False,  # Because replay is open loop
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Fetch the dataset to replay
+dataset = LeRobotDataset(HF_REPO_ID, episodes=[EPISODE_IDX])
+# Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
+episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
+actions = episode_frames.select_columns(ACTION)
+
+# Connect to the robot
+robot.connect()
+
+if not robot.is_connected:
+    raise ValueError("Robot is not connected!")
+
+print("Starting replay loop...")
+log_say(f"Replaying episode {EPISODE_IDX}")
+for idx in range(len(episode_frames)):
+    t0 = time.perf_counter()
+
+    # Get recorded action from dataset
+    ee_action = {
+        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+    }
+
+    # Get robot observation
+    robot_obs = robot.get_observation()
+
+    # Dataset EE -> robot joints
+    joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+
+    # Send action to robot
+    _ = robot.send_action(joint_action)
+
+    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+
+# Clean up
+robot.disconnect()

examples/phone_to_so100/teleoperate.py

@@ -0,0 +1,113 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specif
+
+import time
+
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    robot_action_observation_to_transition,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    EEBoundsAndSafety,
+    EEReferenceAndDelta,
+    GripperVelocityToJoint,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
+from lerobot.teleoperators.phone.teleop_phone import Phone
+from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+
+FPS = 30
+
+# Initialize the robot and teleoperator
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+)
+teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
+
+# Initialize the robot and teleoperator
+robot = SO100Follower(robot_config)
+teleop_device = Phone(teleop_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert phone action to ee pose action to joint action
+phone_to_robot_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+        EEReferenceAndDelta(
+            kinematics=kinematics_solver,
+            end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+            motor_names=list(robot.bus.motors.keys()),
+            use_latched_reference=True,
+        ),
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.10,
+        ),
+        GripperVelocityToJoint(
+            speed_factor=20.0,
+        ),
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Connect to the robot and teleoperator
+robot.connect()
+teleop_device.connect()
+
+# Init rerun viewer
+init_rerun(session_name="phone_so100_teleop")
+
+if not robot.is_connected or not teleop_device.is_connected:
+    raise ValueError("Robot or teleop is not connected!")
+
+print("Starting teleop loop. Move your phone to teleoperate the robot...")
+while True:
+    t0 = time.perf_counter()
+
+    # Get robot observation
+    robot_obs = robot.get_observation()
+
+    # Get teleop action
+    phone_obs = teleop_device.get_action()
+
+    # Phone -> EE pose -> Joints transition
+    joint_action = phone_to_robot_joints_processor((phone_obs, robot_obs))
+
+    # Send action to robot
+    _ = robot.send_action(joint_action)
+
+    # Visualize
+    log_rerun_data(observation=phone_obs, action=joint_action)
+
+    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))

examples/so100_to_so100_EE/evaluate.py

@@ -0,0 +1,198 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.processor import (
+    RobotAction,
+    RobotObservation,
+    RobotProcessorPipeline,
+    make_default_teleop_action_processor,
+)
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    ForwardKinematicsJointsToEE,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.scripts.lerobot_record import record_loop
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import init_rerun
+
+NUM_EPISODES = 5
+FPS = 30
+EPISODE_TIME_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
+
+# Create the robot configuration & robot
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411",
+    id="my_awesome_follower_arm",
+    cameras=camera_config,
+    use_degrees=True,
+)
+
+robot = SO100Follower(robot_config)
+
+# Create policy
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert EE action to joints action
+robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Build pipeline to convert joints observation to EE observation
+robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=observation_to_transition,
+    to_output=transition_to_observation,
+)
+
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_DATASET_ID,
+    fps=FPS,
+    features=combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_joints_to_ee_pose_processor,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=True,
+        ),
+        # User for now should be explicit on the feature keys that were used for record
+        # Alternatively, the user can pass the processor step that has the right features
+        aggregate_pipeline_dataset_features(
+            pipeline=make_default_teleop_action_processor(),
+            initial_features=create_initial_features(
+                action={
+                    f"ee.{k}": PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+                    for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]
+                }
+            ),
+            use_videos=True,
+        ),
+    ),
+    robot_type=robot.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+# Build Policy Processors
+preprocessor, postprocessor = make_pre_post_processors(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+    # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
+    preprocessor_overrides={"device_processor": {"device": str(policy.config.device)}},
+)
+
+# Connect the robot and teleoperator
+robot.connect()
+
+# Initialize the keyboard listener and rerun visualization
+listener, events = init_keyboard_listener()
+init_rerun(session_name="so100_so100_evaluate")
+
+if not robot.is_connected:
+    raise ValueError("Robot is not connected!")
+
+print("Starting evaluate loop...")
+episode_idx = 0
+for episode_idx in range(NUM_EPISODES):
+    log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    # Main record loop
+    record_loop(
+        robot=robot,
+        events=events,
+        fps=FPS,
+        policy=policy,
+        preprocessor=preprocessor,  # Pass the pre and post policy processors
+        postprocessor=postprocessor,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        teleop_action_processor=make_default_teleop_action_processor(),
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose_processor,
+    )
+
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
+        log_say("Reset the environment")
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            control_time_s=EPISODE_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=make_default_teleop_action_processor(),
+            robot_action_processor=robot_ee_to_joints_processor,
+            robot_observation_processor=robot_joints_to_ee_pose_processor,
+        )
+
+    if events["rerecord_episode"]:
+        log_say("Re-record episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
+
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
+
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+listener.stop()
+dataset.push_to_hub()

examples/so100_to_so100_EE/record.py

@@ -0,0 +1,202 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    EEBoundsAndSafety,
+    ForwardKinematicsJointsToEE,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.scripts.lerobot_record import record_loop
+from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
+from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import init_rerun
+
+NUM_EPISODES = 2
+FPS = 30
+EPISODE_TIME_SEC = 60
+RESET_TIME_SEC = 30
+TASK_DESCRIPTION = "My task description"
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+# Create the robot and teleoperator configurations
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+follower_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", cameras=camera_config, use_degrees=True
+)
+leader_config = SO100LeaderConfig(port="/dev/tty.usbmodem5A460819811", id="my_awesome_leader_arm")
+
+# Initialize the robot and teleoperator
+follower = SO100Follower(follower_config)
+leader = SO100Leader(leader_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+follower_kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(follower.bus.motors.keys()),
+)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+leader_kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(leader.bus.motors.keys()),
+)
+
+# Build pipeline to convert follower joints to EE observation
+follower_joints_to_ee = RobotProcessorPipeline[RobotObservation, RobotObservation](
+    steps=[
+        ForwardKinematicsJointsToEE(
+            kinematics=follower_kinematics_solver, motor_names=list(follower.bus.motors.keys())
+        ),
+    ],
+    to_transition=observation_to_transition,
+    to_output=transition_to_observation,
+)
+
+# Build pipeline to convert leader joints to EE action
+leader_joints_to_ee = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        ForwardKinematicsJointsToEE(
+            kinematics=leader_kinematics_solver, motor_names=list(leader.bus.motors.keys())
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Build pipeline to convert EE action to follower joints
+ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    [
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.10,
+        ),
+        InverseKinematicsEEToJoints(
+            kinematics=follower_kinematics_solver,
+            motor_names=list(follower.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_REPO_ID,
+    fps=FPS,
+    features=combine_feature_dicts(
+        # Run the feature contract of the pipelines
+        # This tells you how the features would look like after the pipeline steps
+        aggregate_pipeline_dataset_features(
+            pipeline=leader_joints_to_ee,
+            initial_features=create_initial_features(action=leader.action_features),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=follower_joints_to_ee,
+            initial_features=create_initial_features(observation=follower.observation_features),
+            use_videos=True,
+        ),
+    ),
+    robot_type=follower.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+
+# Connect the robot and teleoperator
+leader.connect()
+follower.connect()
+
+# Initialize the keyboard listener and rerun visualization
+listener, events = init_keyboard_listener()
+init_rerun(session_name="recording_phone")
+
+if not leader.is_connected or not follower.is_connected:
+    raise ValueError("Robot or teleop is not connected!")
+
+print("Starting record loop...")
+episode_idx = 0
+while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+    log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    # Main record loop
+    record_loop(
+        robot=follower,
+        events=events,
+        fps=FPS,
+        teleop=leader,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        teleop_action_processor=leader_joints_to_ee,
+        robot_action_processor=ee_to_follower_joints,
+        robot_observation_processor=follower_joints_to_ee,
+    )
+
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+        log_say("Reset the environment")
+        record_loop(
+            robot=follower,
+            events=events,
+            fps=FPS,
+            teleop=leader,
+            control_time_s=RESET_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=leader_joints_to_ee,
+            robot_action_processor=ee_to_follower_joints,
+            robot_observation_processor=follower_joints_to_ee,
+        )
+
+    if events["rerecord_episode"]:
+        log_say("Re-recording episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
+
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
+
+# Clean up
+log_say("Stop recording")
+leader.disconnect()
+follower.disconnect()
+listener.stop()
+dataset.push_to_hub()

examples/so100_to_so100_EE/replay.py

@@ -0,0 +1,101 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import time
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    robot_action_observation_to_transition,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.utils.constants import ACTION
+from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.utils import log_say
+
+EPISODE_IDX = 0
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+# Initialize the robot config
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+)
+
+# Initialize the robot
+robot = SO100Follower(robot_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert EE action to joints action
+robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=False,  # Because replay is open loop
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Fetch the dataset to replay
+dataset = LeRobotDataset(HF_REPO_ID, episodes=[EPISODE_IDX])
+# Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
+episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
+actions = episode_frames.select_columns(ACTION)
+
+# Connect to the robot
+robot.connect()
+
+if not robot.is_connected:
+    raise ValueError("Robot is not connected!")
+
+print("Starting replay loop...")
+log_say(f"Replaying episode {EPISODE_IDX}")
+for idx in range(len(episode_frames)):
+    t0 = time.perf_counter()
+
+    # Get recorded action from dataset
+    ee_action = {
+        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+    }
+
+    # Get robot observation
+    robot_obs = robot.get_observation()
+
+    # Dataset EE -> robot joints
+    joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+
+    # Send action to robot
+    _ = robot.send_action(joint_action)
+
+    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+
+# Clean up
+robot.disconnect()

examples/so100_to_so100_EE/teleoperate.py

@@ -0,0 +1,121 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    robot_action_observation_to_transition,
+    robot_action_to_transition,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    EEBoundsAndSafety,
+    ForwardKinematicsJointsToEE,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
+from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
+from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+
+FPS = 30
+
+# Initialize the robot and teleoperator config
+follower_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+)
+leader_config = SO100LeaderConfig(port="/dev/tty.usbmodem5A460819811", id="my_awesome_leader_arm")
+
+# Initialize the robot and teleoperator
+follower = SO100Follower(follower_config)
+leader = SO100Leader(leader_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+follower_kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(follower.bus.motors.keys()),
+)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+leader_kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(leader.bus.motors.keys()),
+)
+
+# Build pipeline to convert teleop joints to EE action
+leader_to_ee = RobotProcessorPipeline[RobotAction, RobotAction](
+    steps=[
+        ForwardKinematicsJointsToEE(
+            kinematics=leader_kinematics_solver, motor_names=list(leader.bus.motors.keys())
+        ),
+    ],
+    to_transition=robot_action_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# build pipeline to convert EE action to robot joints
+ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    [
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.10,
+        ),
+        InverseKinematicsEEToJoints(
+            kinematics=follower_kinematics_solver,
+            motor_names=list(follower.bus.motors.keys()),
+            initial_guess_current_joints=False,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Connect to the robot and teleoperator
+follower.connect()
+leader.connect()
+
+# Init rerun viewer
+init_rerun(session_name="so100_so100_EE_teleop")
+
+print("Starting teleop loop...")
+while True:
+    t0 = time.perf_counter()
+
+    # Get robot observation
+    robot_obs = follower.get_observation()
+
+    # Get teleop observation
+    leader_joints_obs = leader.get_action()
+
+    # teleop joints -> teleop EE action
+    leader_ee_act = leader_to_ee(leader_joints_obs)
+
+    # teleop EE -> robot joints
+    follower_joints_act = ee_to_follower_joints((leader_ee_act, robot_obs))
+
+    # Send action to robot
+    _ = follower.send_action(follower_joints_act)
+
+    # Visualize
+    log_rerun_data(observation=leader_ee_act, action=follower_joints_act)
+
+    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))

examples/training/train_policy.py

@@ -12,11 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""This script demonstrates how to train Diffusion Policy on the PushT environment.
-
-Once you have trained a model with this script, you can try to evaluate it on
-examples/2_evaluate_pretrained_policy.py
-"""
+"""This script demonstrates how to train Diffusion Policy on the PushT environment."""
 
 from pathlib import Path
 
@@ -27,6 +23,7 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetad
 from lerobot.datasets.utils import dataset_to_policy_features
 from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
+from lerobot.policies.factory import make_pre_post_processors
 
 
 def main():
@@ -56,9 +53,10 @@ def main():
     cfg = DiffusionConfig(input_features=input_features, output_features=output_features)
 
     # We can now instantiate our policy with this config and the dataset stats.
-    policy = DiffusionPolicy(cfg, dataset_stats=dataset_metadata.stats)
+    policy = DiffusionPolicy(cfg)
     policy.train()
     policy.to(device)
+    preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
 
     # Another policy-dataset interaction is with the delta_timestamps. Each policy expects a given number frames
     # which can differ for inputs, outputs and rewards (if there are some).
@@ -99,7 +97,7 @@ def main():
     done = False
     while not done:
         for batch in dataloader:
-            batch = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in batch.items()}
+            batch = preprocessor(batch)
             loss, _ = policy.forward(batch)
             loss.backward()
             optimizer.step()
@@ -114,6 +112,8 @@ def main():
 
     # Save a policy checkpoint.
     policy.save_pretrained(output_directory)
+    preprocessor.save_pretrained(output_directory)
+    postprocessor.save_pretrained(output_directory)
 
 
 if __name__ == "__main__":

examples/training/train_with_streaming.py

@@ -13,23 +13,20 @@
 # limitations under the License.
 
 """This script demonstrates how to train a Diffusion Policy on the PushT environment,
-using a dataset processed in streaming mode.
-
-Once you have trained a model with this script, you can try to evaluate it on
-examples/2_evaluate_pretrained_policy.py
-"""
+using a dataset processed in streaming mode."""
 
 from pathlib import Path
 
 import torch
 
 from lerobot.configs.types import FeatureType
-from lerobot.constants import ACTION
 from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
 from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
 from lerobot.datasets.utils import dataset_to_policy_features
 from lerobot.policies.act.configuration_act import ACTConfig
 from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.utils.constants import ACTION
 
 
 def main():
@@ -50,9 +47,7 @@ def main():
     training_steps = 10
     log_freq = 1
 
-    dataset_id = (
-        "aractingi/droid_1.0.1"  # 26M frames! Would require 4TB of disk space if installed locally (:
-    )
+    dataset_id = "lerobot/droid_1.0.1"  # 26M frames! Would require 4TB of disk space if installed locally (:
     dataset_metadata = LeRobotDatasetMetadata(dataset_id)
     features = dataset_to_policy_features(dataset_metadata.features)
     output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
@@ -60,9 +55,10 @@ def main():
 
     # We can now instantiate our policy with this config and the dataset stats.
     cfg = ACTConfig(input_features=input_features, output_features=output_features)
-    policy = ACTPolicy(cfg, dataset_stats=dataset_metadata.stats)
+    policy = ACTPolicy(cfg)
     policy.train()
     policy.to(device)
+    preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
 
     # Delta timestamps are used to (1) augment frames used during training and (2) supervise the policy.
     # Here, we use delta-timestamps to only provide ground truth actions for supervision
@@ -89,13 +85,7 @@ def main():
     done = False
     while not done:
         for batch in dataloader:
-            batch = {
-                k: (v.type(torch.float32) if isinstance(v, torch.Tensor) and v.dtype != torch.bool else v)
-                for k, v in batch.items()
-            }
-            batch = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in batch.items()}
-
-            # batch = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in batch.items()}
+            batch = preprocessor(batch)
             loss, _ = policy.forward(batch)
             loss.backward()
             optimizer.step()
@@ -110,6 +100,8 @@ def main():
 
     # Save a policy checkpoint.
     policy.save_pretrained(output_directory)
+    preprocessor.save_pretrained(output_directory)
+    postprocessor.save_pretrained(output_directory)
 
 
 if __name__ == "__main__":

pyproject.toml

@@ -59,7 +59,7 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=2.19.0,<=3.6.0", # TODO: Bumb dependency
+    "datasets>=4.0.0",
     "diffusers>=0.27.2",
     "huggingface-hub[hf-transfer,cli]>=0.34.2",
 
@@ -94,7 +94,7 @@ dependencies = [
 # Common
 pygame-dep = ["pygame>=2.5.1"]
 placo-dep = ["placo>=0.9.6"]
-transformers-dep = ["transformers>=4.50.3,<4.52.0"] # TODO: Bumb dependency
+transformers-dep = ["transformers>=4.52.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"]
 
 # Motors
@@ -111,6 +111,7 @@ intelrealsense = [
     "pyrealsense2>=2.55.1.6486 ; sys_platform != 'darwin'",
     "pyrealsense2-macosx>=2.54 ; sys_platform == 'darwin'",
 ]
+phone = ["hebi-py>=2.8.0", "teleop>=0.1.0"]
 # stretch = [
 #     "hello-robot-stretch-body>=0.7.27 ; sys_platform == 'linux'",
 #     "pyrender @ git+https://github.com/mmatl/pyrender.git ; sys_platform == 'linux'",
@@ -120,7 +121,7 @@ intelrealsense = [
 # Policies
 pi0 = ["lerobot[transformers-dep]"]
 smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14", "accelerate>=1.7.0", "safetensors>=0.4.3"]
-hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.9", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
+hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.11", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
 # Features
 async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3"]
@@ -134,6 +135,8 @@ video_benchmark = ["scikit-image>=0.23.2", "pandas>=2.2.2"]
 aloha = ["gym-aloha>=0.1.1"]
 pusht = ["gym-pusht>=0.1.5", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
 xarm = ["gym-xarm>=0.1.1"]
+libero = ["lerobot[transformers-dep]", "libero @ git+https://github.com/huggingface/lerobot-libero.git@main#egg=libero"]
+
 
 # All
 all = [
@@ -153,19 +156,25 @@ all = [
     "lerobot[video_benchmark]",
     "lerobot[aloha]",
     "lerobot[pusht]",
-    "lerobot[xarm]"
+    "lerobot[xarm]",
+    "lerobot[phone]",
+    "lerobot[libero]",
 ]
 
 [project.scripts]
-lerobot-calibrate="lerobot.calibrate:main"
-lerobot-find-cameras="lerobot.find_cameras:main"
-lerobot-find-port="lerobot.find_port:main"
-lerobot-record="lerobot.record:main"
-lerobot-replay="lerobot.replay:main"
-lerobot-setup-motors="lerobot.setup_motors:main"
-lerobot-teleoperate="lerobot.teleoperate:main"
-lerobot-eval="lerobot.scripts.eval:main"
-lerobot-train="lerobot.scripts.train:main"
+lerobot-calibrate="lerobot.scripts.lerobot_calibrate:main"
+lerobot-find-cameras="lerobot.scripts.lerobot_find_cameras:main"
+lerobot-find-port="lerobot.scripts.lerobot_find_port:main"
+lerobot-record="lerobot.scripts.lerobot_record:main"
+lerobot-replay="lerobot.scripts.lerobot_replay:main"
+lerobot-setup-motors="lerobot.scripts.lerobot_setup_motors:main"
+lerobot-teleoperate="lerobot.scripts.lerobot_teleoperate:main"
+lerobot-eval="lerobot.scripts.lerobot_eval:main"
+lerobot-train="lerobot.scripts.lerobot_train:main"
+lerobot-dataset-viz="lerobot.scripts.lerobot_dataset_viz:main"
+lerobot-info="lerobot.scripts.lerobot_info:main"
+lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
+lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
 
 # ---------------- Tool Configurations ----------------
 [tool.setuptools.packages.find]
@@ -192,7 +201,7 @@ exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]
 # N: pep8-naming
 # TODO: Uncomment rules when ready to use
 select = [
-    "E", "W", "F", "I", "B", "C4", "T20", "N" # "SIM", "A", "S", "D", "RUF", "UP"
+    "E", "W", "F", "I", "B", "C4", "T20", "N", "UP", "SIM" #, "A", "S", "D", "RUF"
 ]
 ignore = [
     "E501", # Line too long
@@ -258,8 +267,83 @@ default.extend-ignore-identifiers-re = [
 # color = true
 # paths = ["src/lerobot"]
 
+# TODO: Enable mypy gradually module by module across multiple PRs
+# Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations
+
 # [tool.mypy]
 # python_version = "3.10"
 # warn_return_any = true
 # warn_unused_configs = true
 # ignore_missing_imports = false
+# strict = true
+# disallow_untyped_defs = true
+# disallow_incomplete_defs = true
+# check_untyped_defs = true
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.utils.*"
+# # include = "src/lerobot/utils/**/*.py"
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.configs.*"
+# # include = "src/lerobot/configs/**/*.py"
+
+# # Data processing modules
+# [[tool.mypy.overrides]]
+# module = "lerobot.processor.*"
+# # include = "src/lerobot/processor/**/*.py"
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.datasets.*"
+# # include = "src/lerobot/datasets/**/*.py"
+
+# # Core machine learning modules
+# [[tool.mypy.overrides]]
+# module = "lerobot.optim.*"
+# # include = "src/lerobot/optim/**/*.py"
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.model.*"
+# # include = "src/lerobot/model/**/*.py"
+
+# # Hardware interfaces
+# [[tool.mypy.overrides]]
+# module = "lerobot.cameras.*"
+# # include = "src/lerobot/cameras/**/*.py"
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.motors.*"
+# # include = "src/lerobot/motors/**/*.py"
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.robots.*"
+# # include = "src/lerobot/robots/**/*.py"
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.teleoperators.*"
+# # include = "src/lerobot/teleoperators/**/*.py"
+
+# # Complex modules (enable these last)
+# [[tool.mypy.overrides]]
+# module = "lerobot.policies.*"
+# # include = "src/lerobot/policies/**/*.py"
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.rl.*"
+# # include = "src/lerobot/rl/**/*.py"
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.envs.*"
+# # include = "src/lerobot/envs/**/*.py"
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.async_inference.*"
+# # include = "src/lerobot/async_inference/**/*.py"
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.transport.*"
+# # include = "src/lerobot/transport/**/*.py"
+
+# [[tool.mypy.overrides]]
+# module = "lerobot.scripts.*"
+# # include = "src/lerobot/scripts/**/*.py"

src/lerobot/async_inference/configs.py

@@ -18,7 +18,8 @@ from dataclasses import dataclass, field
 import torch
 
 from lerobot.robots.config import RobotConfig
-from lerobot.scripts.server.constants import (
+
+from .constants import (
     DEFAULT_FPS,
     DEFAULT_INFERENCE_LATENCY,
     DEFAULT_OBS_QUEUE_TIMEOUT,

src/lerobot/async_inference/helpers.py

@@ -22,16 +22,15 @@ from pathlib import Path
 import torch
 
 from lerobot.configs.types import PolicyFeature
-from lerobot.constants import OBS_IMAGES, OBS_STATE
 from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
 
 # NOTE: Configs need to be loaded for the client to be able to instantiate the policy config
 from lerobot.policies import ACTConfig, DiffusionConfig, PI0Config, SmolVLAConfig, VQBeTConfig  # noqa: F401
 from lerobot.robots.robot import Robot
+from lerobot.utils.constants import OBS_IMAGES, OBS_STATE, OBS_STR
 from lerobot.utils.utils import init_logging
 
 Action = torch.Tensor
-ActionChunk = torch.Tensor
 
 # observation as received from the robot
 RawObservation = dict[str, torch.Tensor]
@@ -46,7 +45,7 @@ Observation = dict[str, torch.Tensor]
 def visualize_action_queue_size(action_queue_size: list[int]) -> None:
     import matplotlib.pyplot as plt
 
-    fig, ax = plt.subplots()
+    _, ax = plt.subplots()
     ax.set_title("Action Queue Size Over Time")
     ax.set_xlabel("Environment steps")
     ax.set_ylabel("Action Queue Size")
@@ -66,7 +65,7 @@ def validate_robot_cameras_for_policy(
 
 
 def map_robot_keys_to_lerobot_features(robot: Robot) -> dict[str, dict]:
-    return hw_to_dataset_features(robot.observation_features, "observation", use_video=False)
+    return hw_to_dataset_features(robot.observation_features, OBS_STR, use_video=False)
 
 
 def is_image_key(k: str) -> bool:
@@ -141,7 +140,7 @@ def make_lerobot_observation(
     lerobot_features: dict[str, dict],
 ) -> LeRobotObservation:
     """Make a lerobot observation from a raw observation."""
-    return build_dataset_frame(lerobot_features, robot_obs, prefix="observation")
+    return build_dataset_frame(lerobot_features, robot_obs, prefix=OBS_STR)
 
 
 def prepare_raw_observation(

src/lerobot/async_inference/policy_server.py

@@ -15,7 +15,7 @@
 """
 Example:
 ```shell
-python src/lerobot/scripts/server/policy_server.py \
+python src/lerobot/async_inference/policy_server.py \
      --host=127.0.0.1 \
      --port=8080 \
      --fps=30 \
@@ -38,9 +38,15 @@ import grpc
 import torch
 
 from lerobot.policies.factory import get_policy_class
-from lerobot.scripts.server.configs import PolicyServerConfig
-from lerobot.scripts.server.constants import SUPPORTED_POLICIES
-from lerobot.scripts.server.helpers import (
+from lerobot.transport import (
+    services_pb2,  # type: ignore
+    services_pb2_grpc,  # type: ignore
+)
+from lerobot.transport.utils import receive_bytes_in_chunks
+
+from .configs import PolicyServerConfig
+from .constants import SUPPORTED_POLICIES
+from .helpers import (
     FPSTracker,
     Observation,
     RemotePolicyConfig,
@@ -50,11 +56,6 @@ from lerobot.scripts.server.helpers import (
     observations_similar,
     raw_observation_to_observation,
 )
-from lerobot.transport import (
-    services_pb2,  # type: ignore
-    services_pb2_grpc,  # type: ignore
-)
-from lerobot.transport.utils import receive_bytes_in_chunks
 
 
 class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):

src/lerobot/async_inference/robot_client.py

@@ -15,7 +15,7 @@
 """
 Example command:
 ```shell
-python src/lerobot/scripts/server/robot_client.py \
+python src/lerobot/async_inference/robot_client.py \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
@@ -57,9 +57,15 @@ from lerobot.robots import (  # noqa: F401
     so100_follower,
     so101_follower,
 )
-from lerobot.scripts.server.configs import RobotClientConfig
-from lerobot.scripts.server.constants import SUPPORTED_ROBOTS
-from lerobot.scripts.server.helpers import (
+from lerobot.transport import (
+    services_pb2,  # type: ignore
+    services_pb2_grpc,  # type: ignore
+)
+from lerobot.transport.utils import grpc_channel_options, send_bytes_in_chunks
+
+from .configs import RobotClientConfig
+from .constants import SUPPORTED_ROBOTS
+from .helpers import (
     Action,
     FPSTracker,
     Observation,
@@ -72,11 +78,6 @@ from lerobot.scripts.server.helpers import (
     validate_robot_cameras_for_policy,
     visualize_action_queue_size,
 )
-from lerobot.transport import (
-    services_pb2,  # type: ignore
-    services_pb2_grpc,  # type: ignore
-)
-from lerobot.transport.utils import grpc_channel_options, send_bytes_in_chunks
 
 
 class RobotClient:

src/lerobot/cameras/opencv/camera_opencv.py

@@ -31,7 +31,7 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
 import cv2
 import numpy as np
 
-from lerobot.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
 from ..utils import get_cv2_backend, get_cv2_rotation

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -31,7 +31,7 @@ import numpy as np
 from reachy2_sdk.media.camera import CameraView
 from reachy2_sdk.media.camera_manager import CameraManager
 
-from lerobot.errors import DeviceNotConnectedError
+from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
 from .configuration_reachy2_camera import ColorMode, Reachy2CameraConfig

src/lerobot/cameras/realsense/camera_realsense.py

@@ -29,7 +29,7 @@ try:
 except Exception as e:
     logging.info(f"Could not import realsense: {e}")
 
-from lerobot.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode

src/lerobot/cameras/utils.py

@@ -15,14 +15,10 @@
 # limitations under the License.
 
 import platform
-from pathlib import Path
-from typing import TypeAlias
 
 from .camera import Camera
 from .configs import CameraConfig, Cv2Rotation
 
-IndexOrPath: TypeAlias = int | Path
-
 
 def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[str, Camera]:
     cameras = {}

src/lerobot/configs/default.py

@@ -16,9 +16,6 @@
 
 from dataclasses import dataclass, field
 
-from lerobot import (
-    policies,  # noqa: F401
-)
 from lerobot.datasets.transforms import ImageTransformsConfig
 from lerobot.datasets.video_utils import get_safe_default_codec
 

src/lerobot/configs/policies.py

@@ -26,10 +26,10 @@ from huggingface_hub import hf_hub_download
 from huggingface_hub.constants import CONFIG_NAME
 from huggingface_hub.errors import HfHubHTTPError
 
-from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
-from lerobot.constants import ACTION, OBS_STATE
+from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.optim.optimizers import OptimizerConfig
 from lerobot.optim.schedulers import LRSchedulerConfig
+from lerobot.utils.constants import ACTION, OBS_STATE
 from lerobot.utils.hub import HubMixin
 from lerobot.utils.utils import auto_select_torch_device, is_amp_available, is_torch_device_available
 
@@ -53,7 +53,6 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
     """
 
     n_obs_steps: int = 1
-    normalization_mapping: dict[str, NormalizationMode] = field(default_factory=dict)
 
     input_features: dict[str, PolicyFeature] = field(default_factory=dict)
     output_features: dict[str, PolicyFeature] = field(default_factory=dict)
@@ -197,11 +196,10 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
             config = json.load(f)
 
         config.pop("type")
-        with tempfile.NamedTemporaryFile("w+") as f:
+        with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as f:
             json.dump(config, f)
             config_file = f.name
-            f.flush()
 
-            cli_overrides = policy_kwargs.pop("cli_overrides", [])
-            with draccus.config_type("json"):
-                return draccus.parse(orig_config.__class__, config_file, args=cli_overrides)
+        cli_overrides = policy_kwargs.pop("cli_overrides", [])
+        with draccus.config_type("json"):
+            return draccus.parse(orig_config.__class__, config_file, args=cli_overrides)

src/lerobot/configs/types.py

@@ -15,7 +15,6 @@
 # https://stackoverflow.com/questions/24481852/serialising-an-enum-member-to-json
 from dataclasses import dataclass
 from enum import Enum
-from typing import Any, Protocol
 
 
 class FeatureType(str, Enum):
@@ -24,6 +23,12 @@ class FeatureType(str, Enum):
     ENV = "ENV"
     ACTION = "ACTION"
     REWARD = "REWARD"
+    LANGUAGE = "LANGUAGE"
+
+
+class PipelineFeatureType(str, Enum):
+    ACTION = "ACTION"
+    OBSERVATION = "OBSERVATION"
 
 
 class NormalizationMode(str, Enum):
@@ -32,10 +37,6 @@ class NormalizationMode(str, Enum):
     IDENTITY = "IDENTITY"
 
 
-class DictLike(Protocol):
-    def __getitem__(self, key: Any) -> Any: ...
-
-
 @dataclass
 class PolicyFeature:
     type: FeatureType

src/lerobot/datasets/aggregate.py

@@ -93,14 +93,13 @@ def update_data_df(df, src_meta, dst_meta):
         pd.DataFrame: Updated DataFrame with adjusted indices.
     """
 
-    def _update(row):
-        row["episode_index"] = row["episode_index"] + dst_meta.info["total_episodes"]
-        row["index"] = row["index"] + dst_meta.info["total_frames"]
-        task = src_meta.tasks.iloc[row["task_index"]].name
-        row["task_index"] = dst_meta.tasks.loc[task].task_index.item()
-        return row
+    df["episode_index"] = df["episode_index"] + dst_meta.info["total_episodes"]
+    df["index"] = df["index"] + dst_meta.info["total_frames"]
 
-    return df.apply(_update, axis=1)
+    src_task_names = src_meta.tasks.index.take(df["task_index"].to_numpy())
+    df["task_index"] = dst_meta.tasks.loc[src_task_names, "task_index"].to_numpy()
+
+    return df
 
 
 def update_meta_data(
@@ -126,27 +125,21 @@ def update_meta_data(
         pd.DataFrame: Updated DataFrame with adjusted indices and timestamps.
     """
 
-    def _update(row):
-        row["meta/episodes/chunk_index"] = row["meta/episodes/chunk_index"] + meta_idx["chunk"]
-        row["meta/episodes/file_index"] = row["meta/episodes/file_index"] + meta_idx["file"]
-        row["data/chunk_index"] = row["data/chunk_index"] + data_idx["chunk"]
-        row["data/file_index"] = row["data/file_index"] + data_idx["file"]
-        for key, video_idx in videos_idx.items():
-            row[f"videos/{key}/chunk_index"] = row[f"videos/{key}/chunk_index"] + video_idx["chunk"]
-            row[f"videos/{key}/file_index"] = row[f"videos/{key}/file_index"] + video_idx["file"]
-            row[f"videos/{key}/from_timestamp"] = (
-                row[f"videos/{key}/from_timestamp"] + video_idx["latest_duration"]
-            )
-            row[f"videos/{key}/to_timestamp"] = (
-                row[f"videos/{key}/to_timestamp"] + video_idx["latest_duration"]
-            )
+    df["meta/episodes/chunk_index"] = df["meta/episodes/chunk_index"] + meta_idx["chunk"]
+    df["meta/episodes/file_index"] = df["meta/episodes/file_index"] + meta_idx["file"]
+    df["data/chunk_index"] = df["data/chunk_index"] + data_idx["chunk"]
+    df["data/file_index"] = df["data/file_index"] + data_idx["file"]
+    for key, video_idx in videos_idx.items():
+        df[f"videos/{key}/chunk_index"] = df[f"videos/{key}/chunk_index"] + video_idx["chunk"]
+        df[f"videos/{key}/file_index"] = df[f"videos/{key}/file_index"] + video_idx["file"]
+        df[f"videos/{key}/from_timestamp"] = df[f"videos/{key}/from_timestamp"] + video_idx["latest_duration"]
+        df[f"videos/{key}/to_timestamp"] = df[f"videos/{key}/to_timestamp"] + video_idx["latest_duration"]
 
-        row["dataset_from_index"] = row["dataset_from_index"] + dst_meta.info["total_frames"]
-        row["dataset_to_index"] = row["dataset_to_index"] + dst_meta.info["total_frames"]
-        row["episode_index"] = row["episode_index"] + dst_meta.info["total_episodes"]
-        return row
+    df["dataset_from_index"] = df["dataset_from_index"] + dst_meta.info["total_frames"]
+    df["dataset_to_index"] = df["dataset_to_index"] + dst_meta.info["total_frames"]
+    df["episode_index"] = df["episode_index"] + dst_meta.info["total_episodes"]
 
-    return df.apply(_update, axis=1)
+    return df
 
 
 def aggregate_datasets(

src/lerobot/datasets/factory.py

@@ -27,6 +27,7 @@ from lerobot.datasets.lerobot_dataset import (
 )
 from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
 from lerobot.datasets.transforms import ImageTransforms
+from lerobot.utils.constants import ACTION, OBS_PREFIX, REWARD
 
 IMAGENET_STATS = {
     "mean": [[[0.485]], [[0.456]], [[0.406]]],  # (c,1,1)
@@ -54,11 +55,11 @@ def resolve_delta_timestamps(
     """
     delta_timestamps = {}
     for key in ds_meta.features:
-        if key == "next.reward" and cfg.reward_delta_indices is not None:
+        if key == REWARD and cfg.reward_delta_indices is not None:
             delta_timestamps[key] = [i / ds_meta.fps for i in cfg.reward_delta_indices]
-        if key == "action" and cfg.action_delta_indices is not None:
+        if key == ACTION and cfg.action_delta_indices is not None:
             delta_timestamps[key] = [i / ds_meta.fps for i in cfg.action_delta_indices]
-        if key.startswith("observation.") and cfg.observation_delta_indices is not None:
+        if key.startswith(OBS_PREFIX) and cfg.observation_delta_indices is not None:
             delta_timestamps[key] = [i / ds_meta.fps for i in cfg.observation_delta_indices]
 
     if len(delta_timestamps) == 0:

src/lerobot/datasets/lerobot_dataset.py

@@ -31,7 +31,6 @@ import torch.utils
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.errors import RevisionNotFoundError
 
-from lerobot.constants import HF_LEROBOT_HOME
 from lerobot.datasets.compute_stats import aggregate_stats, compute_episode_stats
 from lerobot.datasets.image_writer import AsyncImageWriter, write_image
 from lerobot.datasets.utils import (
@@ -79,6 +78,7 @@ from lerobot.datasets.video_utils import (
     get_video_duration_in_s,
     get_video_info,
 )
+from lerobot.utils.constants import HF_LEROBOT_HOME
 
 CODEBASE_VERSION = "v3.0"
 
@@ -848,11 +848,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         return item
 
-    def _add_padding_keys(self, item: dict, padding: dict[str, list[bool]]) -> dict:
-        for key, val in padding.items():
-            item[key] = torch.BoolTensor(val)
-        return item
-
     def __len__(self):
         return self.num_frames
 
@@ -1032,7 +1027,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             # Reset episode buffer and clean up temporary images (if not already deleted during video encoding)
             self.clear_episode_buffer(delete_images=len(self.meta.image_keys) > 0)
 
-    def _batch_save_episode_video(self, start_episode: int, end_episode: int | None = None):
+    def _batch_save_episode_video(self, start_episode: int, end_episode: int | None = None) -> None:
         """
         Batch save videos for multiple episodes.
 
@@ -1158,7 +1153,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         }
         return metadata
 
-    def _save_episode_video(self, video_key: str, episode_index: int):
+    def _save_episode_video(self, video_key: str, episode_index: int) -> dict:
         # Encode episode frames into a temporary video
         ep_path = self._encode_temporary_episode_video(video_key, episode_index)
         ep_size_in_mb = get_video_size_in_mb(ep_path)
@@ -1263,7 +1258,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         if self.image_writer is not None:
             self.image_writer.wait_until_done()
 
-    def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> dict:
+    def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> Path:
         """
         Use ffmpeg to convert frames stored as png into mp4 videos.
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
@@ -1396,11 +1391,6 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         """
         return {repo_id: i for i, repo_id in enumerate(self.repo_ids)}
 
-    @property
-    def repo_index_to_id(self):
-        """Return the inverse mapping if repo_id_to_index."""
-        return {v: k for k, v in self.repo_id_to_index}
-
     @property
     def fps(self) -> int:
         """Frames per second used during data collection.
@@ -1431,7 +1421,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         """Keys to access image and video stream from cameras."""
         keys = []
         for key, feats in self.features.items():
-            if isinstance(feats, (datasets.Image, VideoFrame)):
+            if isinstance(feats, (datasets.Image | VideoFrame)):
                 keys.append(key)
         return keys
 

src/lerobot/datasets/pipeline_features.py

@@ -0,0 +1,139 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+from collections.abc import Sequence
+from typing import Any
+
+from lerobot.configs.types import PipelineFeatureType
+from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.processor import DataProcessorPipeline
+from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE, OBS_STR
+
+
+def create_initial_features(
+    action: dict[str, Any] | None = None, observation: dict[str, Any] | None = None
+) -> dict[PipelineFeatureType, dict[str, Any]]:
+    """
+    Creates the initial features dict for the dataset from action and observation specs.
+
+    Args:
+        action: A dictionary of action feature names to their types/shapes.
+        observation: A dictionary of observation feature names to their types/shapes.
+
+    Returns:
+        The initial features dictionary structured by PipelineFeatureType.
+    """
+    features = {PipelineFeatureType.ACTION: {}, PipelineFeatureType.OBSERVATION: {}}
+    if action:
+        features[PipelineFeatureType.ACTION] = action
+    if observation:
+        features[PipelineFeatureType.OBSERVATION] = observation
+    return features
+
+
+# Helper to filter state/action keys based on regex patterns.
+def should_keep(key: str, patterns: tuple[str]) -> bool:
+    if patterns is None:
+        return True
+    return any(re.search(pat, key) for pat in patterns)
+
+
+def strip_prefix(key: str, prefixes_to_strip: tuple[str]) -> str:
+    for prefix in prefixes_to_strip:
+        if key.startswith(prefix):
+            return key[len(prefix) :]
+    return key
+
+
+# Define prefixes to strip from feature keys for clean names.
+# Handles both fully qualified (e.g., "action.state") and short (e.g., "state") forms.
+PREFIXES_TO_STRIP = tuple(
+    f"{token}." for const in (ACTION, OBS_STATE, OBS_IMAGES) for token in (const, const.split(".")[-1])
+)
+
+
+def aggregate_pipeline_dataset_features(
+    pipeline: DataProcessorPipeline,
+    initial_features: dict[PipelineFeatureType, dict[str, Any]],
+    *,
+    use_videos: bool = True,
+    patterns: Sequence[str] | None = None,
+) -> dict[str, dict]:
+    """
+    Aggregates and filters pipeline features to create a dataset-ready features dictionary.
+
+    This function transforms initial features using the pipeline, categorizes them as action or observations
+    (image or state), filters them based on `use_videos` and `patterns`, and finally
+    formats them for use with a Hugging Face LeRobot Dataset.
+
+    Args:
+        pipeline: The DataProcessorPipeline to apply.
+        initial_features: A dictionary of raw feature specs for actions and observations.
+        use_videos: If False, image features are excluded.
+        patterns: A sequence of regex patterns to filter action and state features.
+                  Image features are not affected by this filter.
+
+    Returns:
+        A dictionary of features formatted for a Hugging Face LeRobot Dataset.
+    """
+    all_features = pipeline.transform_features(initial_features)
+
+    # Intermediate storage for categorized and filtered features.
+    processed_features: dict[str, dict[str, Any]] = {
+        ACTION: {},
+        OBS_STR: {},
+    }
+    images_token = OBS_IMAGES.split(".")[-1]
+
+    # Iterate through all features transformed by the pipeline.
+    for ptype, feats in all_features.items():
+        if ptype not in [PipelineFeatureType.ACTION, PipelineFeatureType.OBSERVATION]:
+            continue
+
+        for key, value in feats.items():
+            # 1. Categorize the feature.
+            is_action = ptype == PipelineFeatureType.ACTION
+            # Observations are classified as images if their key matches image-related tokens or if the shape of the feature is 3.
+            # All other observations are treated as state.
+            is_image = not is_action and (
+                (isinstance(value, tuple) and len(value) == 3)
+                or (
+                    key.startswith(f"{OBS_IMAGES}.")
+                    or key.startswith(f"{images_token}.")
+                    or f".{images_token}." in key
+                )
+            )
+
+            # 2. Apply filtering rules.
+            if is_image and not use_videos:
+                continue
+            if not is_image and not should_keep(key, patterns):
+                continue
+
+            # 3. Add the feature to the appropriate group with a clean name.
+            name = strip_prefix(key, PREFIXES_TO_STRIP)
+            if is_action:
+                processed_features[ACTION][name] = value
+            else:
+                processed_features[OBS_STR][name] = value
+
+    # Convert the processed features into the final dataset format.
+    dataset_features = {}
+    if processed_features[ACTION]:
+        dataset_features.update(hw_to_dataset_features(processed_features[ACTION], ACTION, use_videos))
+    if processed_features[OBS_STR]:
+        dataset_features.update(hw_to_dataset_features(processed_features[OBS_STR], OBS_STR, use_videos))
+
+    return dataset_features

src/lerobot/datasets/push_dataset_to_hub/utils.py

@@ -13,67 +13,10 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import inspect
-from concurrent.futures import ThreadPoolExecutor
-from pathlib import Path
 
 import datasets
-import numpy
-import PIL
 import torch
 
-from lerobot.datasets.video_utils import encode_video_frames
-
-
-def concatenate_episodes(ep_dicts):
-    data_dict = {}
-
-    keys = ep_dicts[0].keys()
-    for key in keys:
-        if torch.is_tensor(ep_dicts[0][key][0]):
-            data_dict[key] = torch.cat([ep_dict[key] for ep_dict in ep_dicts])
-        else:
-            if key not in data_dict:
-                data_dict[key] = []
-            for ep_dict in ep_dicts:
-                for x in ep_dict[key]:
-                    data_dict[key].append(x)
-
-    total_frames = data_dict["frame_index"].shape[0]
-    data_dict["index"] = torch.arange(0, total_frames, 1)
-    return data_dict
-
-
-def save_images_concurrently(imgs_array: numpy.array, out_dir: Path, max_workers: int = 4):
-    out_dir = Path(out_dir)
-    out_dir.mkdir(parents=True, exist_ok=True)
-
-    def save_image(img_array, i, out_dir):
-        img = PIL.Image.fromarray(img_array)
-        img.save(str(out_dir / f"frame_{i:06d}.png"), quality=100)
-
-    num_images = len(imgs_array)
-    with ThreadPoolExecutor(max_workers=max_workers) as executor:
-        [executor.submit(save_image, imgs_array[i], i, out_dir) for i in range(num_images)]
-
-
-def get_default_encoding() -> dict:
-    """Returns the default ffmpeg encoding parameters used by `encode_video_frames`."""
-    signature = inspect.signature(encode_video_frames)
-    return {
-        k: v.default
-        for k, v in signature.parameters.items()
-        if v.default is not inspect.Parameter.empty and k in ["vcodec", "pix_fmt", "g", "crf"]
-    }
-
-
-def check_repo_id(repo_id: str) -> None:
-    if len(repo_id.split("/")) != 2:
-        raise ValueError(
-            f"""`repo_id` is expected to contain a community or user id `/` the name of the dataset
-            (e.g. 'lerobot/pusht'), but contains '{repo_id}'."""
-        )
-
 
 # TODO(aliberts): remove
 def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> dict[str, torch.Tensor]:

src/lerobot/datasets/streaming_dataset.py

@@ -21,7 +21,6 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
-from lerobot.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE
 from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
 from lerobot.datasets.utils import (
     Backtrackable,
@@ -38,6 +37,7 @@ from lerobot.datasets.video_utils import (
     VideoDecoderCache,
     decode_video_frames_torchcodec,
 )
+from lerobot.utils.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE
 
 
 class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
@@ -298,9 +298,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
 
         return padding_mask
 
-    def make_frame(
-        self, dataset_iterator: Backtrackable, previous_dataset_iterator: Backtrackable | None = None
-    ) -> Generator:
+    def make_frame(self, dataset_iterator: Backtrackable) -> Generator:
         """Makes a frame starting from a dataset iterator"""
         item = next(dataset_iterator)
         item = item_to_torch(item)

src/lerobot/datasets/transforms.py

@@ -120,7 +120,7 @@ class SharpnessJitter(Transform):
         self.sharpness = self._check_input(sharpness)
 
     def _check_input(self, sharpness):
-        if isinstance(sharpness, (int, float)):
+        if isinstance(sharpness, (int | float)):
             if sharpness < 0:
                 raise ValueError("If sharpness is a single number, it must be non negative.")
             sharpness = [1.0 - sharpness, 1.0 + sharpness]

src/lerobot/datasets/utils.py

@@ -21,7 +21,7 @@ from collections import deque
 from collections.abc import Iterable, Iterator
 from pathlib import Path
 from pprint import pformat
-from typing import Any, Deque, Generic, TypeVar
+from typing import Any, Generic, TypeVar
 
 import datasets
 import numpy as np
@@ -43,6 +43,7 @@ from lerobot.datasets.backward_compatibility import (
     BackwardCompatibilityError,
     ForwardCompatibilityError,
 )
+from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_STR
 from lerobot.utils.utils import is_valid_numpy_dtype_string
 
 DEFAULT_CHUNK_SIZE = 1000  # Max number of files per chunk
@@ -66,18 +67,6 @@ DEFAULT_IMAGE_PATH = "images/{image_key}/episode-{episode_index:06d}/frame-{fram
 LEGACY_EPISODES_PATH = "meta/episodes.jsonl"
 LEGACY_EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
 LEGACY_TASKS_PATH = "meta/tasks.jsonl"
-LEGACY_DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
-LEGACY_DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
-
-DATASET_CARD_TEMPLATE = """
----
-# Metadata will go there
----
-This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
-
-## {}
-
-"""
 
 DEFAULT_FEATURES = {
     "timestamp": {"dtype": "float32", "shape": (1,), "names": None},
@@ -150,14 +139,20 @@ def get_video_size_in_mb(mp4_path: Path) -> float:
 
 
 def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
-    """Flatten a nested dictionary structure by collapsing nested keys into one key with a separator.
+    """Flatten a nested dictionary by joining keys with a separator.
 
-    For example:
-    ```
-    >>> dct = {"a": {"b": 1, "c": {"d": 2}}, "e": 3}`
-    >>> print(flatten_dict(dct))
-    {"a/b": 1, "a/c/d": 2, "e": 3}
-    ```
+    Example:
+        >>> dct = {"a": {"b": 1, "c": {"d": 2}}, "e": 3}
+        >>> print(flatten_dict(dct))
+        {'a/b': 1, 'a/c/d': 2, 'e': 3}
+
+    Args:
+        d (dict): The dictionary to flatten.
+        parent_key (str): The base key to prepend to the keys in this level.
+        sep (str): The separator to use between keys.
+
+    Returns:
+        dict: A flattened dictionary.
     """
     items = []
     for k, v in d.items():
@@ -170,6 +165,20 @@ def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
 
 
 def unflatten_dict(d: dict, sep: str = "/") -> dict:
+    """Unflatten a dictionary with delimited keys into a nested dictionary.
+
+    Example:
+        >>> flat_dct = {"a/b": 1, "a/c/d": 2, "e": 3}
+        >>> print(unflatten_dict(flat_dct))
+        {'a': {'b': 1, 'c': {'d': 2}}, 'e': 3}
+
+    Args:
+        d (dict): A dictionary with flattened keys.
+        sep (str): The separator used in the keys.
+
+    Returns:
+        dict: A nested dictionary.
+    """
     outdict = {}
     for key, value in d.items():
         parts = key.split(sep)
@@ -183,15 +192,28 @@ def unflatten_dict(d: dict, sep: str = "/") -> dict:
 
 
 def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
+    """Serialize a dictionary containing tensors or numpy arrays to be JSON-compatible.
+
+    Converts torch.Tensor, np.ndarray, and np.generic types to lists or native Python types.
+
+    Args:
+        stats (dict): A dictionary that may contain non-serializable numeric types.
+
+    Returns:
+        dict: A dictionary with all values converted to JSON-serializable types.
+
+    Raises:
+        NotImplementedError: If a value has an unsupported type.
+    """
     serialized_dict = {}
     for key, value in flatten_dict(stats).items():
-        if isinstance(value, (torch.Tensor, np.ndarray)):
+        if isinstance(value, (torch.Tensor | np.ndarray)):
             serialized_dict[key] = value.tolist()
-        elif isinstance(value, list) and isinstance(value[0], (int, float, list)):
+        elif isinstance(value, list) and isinstance(value[0], (int | float | list)):
             serialized_dict[key] = value
         elif isinstance(value, np.generic):
             serialized_dict[key] = value.item()
-        elif isinstance(value, (int, float)):
+        elif isinstance(value, (int | float)):
             serialized_dict[key] = value
         else:
             raise NotImplementedError(f"The value '{value}' of type '{type(value)}' is not supported.")
@@ -199,6 +221,17 @@ def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
 
 
 def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
+    """Embed image bytes into the dataset table before saving to Parquet.
+
+    This function prepares a Hugging Face dataset for serialization by converting
+    image objects into an embedded format that can be stored in Arrow/Parquet.
+
+    Args:
+        dataset (datasets.Dataset): The input dataset, possibly containing image features.
+
+    Returns:
+        datasets.Dataset: The dataset with images embedded in the table storage.
+    """
     # Embed image bytes into the table before saving to parquet
     format = dataset.format
     dataset = dataset.with_format("arrow")
@@ -208,11 +241,27 @@ def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
 
 
 def load_json(fpath: Path) -> Any:
+    """Load data from a JSON file.
+
+    Args:
+        fpath (Path): Path to the JSON file.
+
+    Returns:
+        Any: The data loaded from the JSON file.
+    """
     with open(fpath) as f:
         return json.load(f)
 
 
 def write_json(data: dict, fpath: Path) -> None:
+    """Write data to a JSON file.
+
+    Creates parent directories if they don't exist.
+
+    Args:
+        data (dict): The dictionary to write.
+        fpath (Path): The path to the output JSON file.
+    """
     fpath.parent.mkdir(exist_ok=True, parents=True)
     with open(fpath, "w") as f:
         json.dump(data, f, indent=4, ensure_ascii=False)
@@ -223,6 +272,16 @@ def write_info(info: dict, local_dir: Path) -> None:
 
 
 def load_info(local_dir: Path) -> dict:
+    """Load dataset info metadata from its standard file path.
+
+    Also converts shape lists to tuples for consistency.
+
+    Args:
+        local_dir (Path): The root directory of the dataset.
+
+    Returns:
+        dict: The dataset information dictionary.
+    """
     info = load_json(local_dir / INFO_PATH)
     for ft in info["features"].values():
         ft["shape"] = tuple(ft["shape"])
@@ -230,16 +289,40 @@ def load_info(local_dir: Path) -> dict:
 
 
 def write_stats(stats: dict, local_dir: Path) -> None:
+    """Serialize and write dataset statistics to their standard file path.
+
+    Args:
+        stats (dict): The statistics dictionary (can contain tensors/numpy arrays).
+        local_dir (Path): The root directory of the dataset.
+    """
     serialized_stats = serialize_dict(stats)
     write_json(serialized_stats, local_dir / STATS_PATH)
 
 
 def cast_stats_to_numpy(stats: dict) -> dict[str, dict[str, np.ndarray]]:
+    """Recursively cast numerical values in a stats dictionary to numpy arrays.
+
+    Args:
+        stats (dict): The statistics dictionary.
+
+    Returns:
+        dict: The statistics dictionary with values cast to numpy arrays.
+    """
     stats = {key: np.array(value) for key, value in flatten_dict(stats).items()}
     return unflatten_dict(stats)
 
 
 def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]] | None:
+    """Load dataset statistics and cast numerical values to numpy arrays.
+
+    Returns None if the stats file doesn't exist.
+
+    Args:
+        local_dir (Path): The root directory of the dataset.
+
+    Returns:
+        A dictionary of statistics or None if the file is not found.
+    """
     if not (local_dir / STATS_PATH).exists():
         return None
     stats = load_json(local_dir / STATS_PATH)
@@ -288,15 +371,21 @@ def load_episodes(local_dir: Path) -> datasets.Dataset:
     return episodes
 
 
-def backward_compatible_episodes_stats(
-    stats: dict[str, dict[str, np.ndarray]], episodes: list[int]
-) -> dict[int, dict[str, dict[str, np.ndarray]]]:
-    return dict.fromkeys(episodes, stats)
-
-
 def load_image_as_numpy(
     fpath: str | Path, dtype: np.dtype = np.float32, channel_first: bool = True
 ) -> np.ndarray:
+    """Load an image from a file into a numpy array.
+
+    Args:
+        fpath (str | Path): Path to the image file.
+        dtype (np.dtype): The desired data type of the output array. If floating,
+            pixels are scaled to [0, 1].
+        channel_first (bool): If True, converts the image to (C, H, W) format.
+            Otherwise, it remains in (H, W, C) format.
+
+    Returns:
+        np.ndarray: The image as a numpy array.
+    """
     img = PILImage.open(fpath).convert("RGB")
     img_array = np.array(img, dtype=dtype)
     if channel_first:  # (H, W, C) -> (C, H, W)
@@ -307,10 +396,19 @@ def load_image_as_numpy(
 
 
 def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[torch.Tensor | str]]:
-    """Get a transform function that convert items from Hugging Face dataset (pyarrow)
-    to torch tensors. Importantly, images are converted from PIL, which corresponds to
-    a channel last representation (h w c) of uint8 type, to a torch image representation
-    with channel first (c h w) of float32 type in range [0,1].
+    """Convert a batch from a Hugging Face dataset to torch tensors.
+
+    This transform function converts items from Hugging Face dataset format (pyarrow)
+    to torch tensors. Importantly, images are converted from PIL objects (H, W, C, uint8)
+    to a torch image representation (C, H, W, float32) in the range [0, 1]. Other
+    types are converted to torch.tensor.
+
+    Args:
+        items_dict (dict): A dictionary representing a batch of data from a
+            Hugging Face dataset.
+
+    Returns:
+        dict: The batch with items converted to torch tensors.
     """
     for key in items_dict:
         first_item = items_dict[key][0]
@@ -325,6 +423,14 @@ def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[to
 
 
 def is_valid_version(version: str) -> bool:
+    """Check if a string is a valid PEP 440 version.
+
+    Args:
+        version (str): The version string to check.
+
+    Returns:
+        bool: True if the version string is valid, False otherwise.
+    """
     try:
         packaging.version.parse(version)
         return True
@@ -338,6 +444,18 @@ def check_version_compatibility(
     current_version: str | packaging.version.Version,
     enforce_breaking_major: bool = True,
 ) -> None:
+    """Check for version compatibility between a dataset and the current codebase.
+
+    Args:
+        repo_id (str): The repository ID for logging purposes.
+        version_to_check (str | packaging.version.Version): The version of the dataset.
+        current_version (str | packaging.version.Version): The current version of the codebase.
+        enforce_breaking_major (bool): If True, raise an error on major version mismatch.
+
+    Raises:
+        BackwardCompatibilityError: If the dataset version is from a newer, incompatible
+            major version of the codebase.
+    """
     v_check = (
         packaging.version.parse(version_to_check)
         if not isinstance(version_to_check, packaging.version.Version)
@@ -355,7 +473,14 @@ def check_version_compatibility(
 
 
 def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
-    """Returns available valid versions (branches and tags) on given repo."""
+    """Return available valid versions (branches and tags) on a given Hub repo.
+
+    Args:
+        repo_id (str): The repository ID on the Hugging Face Hub.
+
+    Returns:
+        list[packaging.version.Version]: A list of valid versions found.
+    """
     api = HfApi()
     repo_refs = api.list_repo_refs(repo_id, repo_type="dataset")
     repo_refs = [b.name for b in repo_refs.branches + repo_refs.tags]
@@ -368,9 +493,22 @@ def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
 
 
 def get_safe_version(repo_id: str, version: str | packaging.version.Version) -> str:
-    """
-    Returns the version if available on repo or the latest compatible one.
-    Otherwise, will throw a `CompatibilityError`.
+    """Return the specified version if available on repo, or the latest compatible one.
+
+    If the exact version is not found, it looks for the latest version with the
+    same major version number that is less than or equal to the target minor version.
+
+    Args:
+        repo_id (str): The repository ID on the Hugging Face Hub.
+        version (str | packaging.version.Version): The target version.
+
+    Returns:
+        str: The safe version string (e.g., "v1.2.3") to use as a revision.
+
+    Raises:
+        RevisionNotFoundError: If the repo has no version tags.
+        BackwardCompatibilityError: If only older major versions are available.
+        ForwardCompatibilityError: If only newer major versions are available.
     """
     target_version = (
         packaging.version.parse(version) if not isinstance(version, packaging.version.Version) else version
@@ -412,6 +550,17 @@ def get_safe_version(repo_id: str, version: str | packaging.version.Version) ->
 
 
 def get_hf_features_from_features(features: dict) -> datasets.Features:
+    """Convert a LeRobot features dictionary to a `datasets.Features` object.
+
+    Args:
+        features (dict): A LeRobot-style feature dictionary.
+
+    Returns:
+        datasets.Features: The corresponding Hugging Face `datasets.Features` object.
+
+    Raises:
+        ValueError: If a feature has an unsupported shape.
+    """
     hf_features = {}
     for key, ft in features.items():
         if ft["dtype"] == "video":
@@ -439,6 +588,14 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
 
 
 def _validate_feature_names(features: dict[str, dict]) -> None:
+    """Validate that feature names do not contain invalid characters.
+
+    Args:
+        features (dict): The LeRobot features dictionary.
+
+    Raises:
+        ValueError: If any feature name contains '/'.
+    """
     invalid_features = {name: ft for name, ft in features.items() if "/" in name}
     if invalid_features:
         raise ValueError(f"Feature names should not contain '/'. Found '/' in '{invalid_features}'.")
@@ -447,18 +604,38 @@ def _validate_feature_names(features: dict[str, dict]) -> None:
 def hw_to_dataset_features(
     hw_features: dict[str, type | tuple], prefix: str, use_video: bool = True
 ) -> dict[str, dict]:
+    """Convert hardware-specific features to a LeRobot dataset feature dictionary.
+
+    This function takes a dictionary describing hardware outputs (like joint states
+    or camera image shapes) and formats it into the standard LeRobot feature
+    specification.
+
+    Args:
+        hw_features (dict): Dictionary mapping feature names to their type (float for
+            joints) or shape (tuple for images).
+        prefix (str): The prefix to add to the feature keys (e.g., "observation"
+            or "action").
+        use_video (bool): If True, image features are marked as "video", otherwise "image".
+
+    Returns:
+        dict: A LeRobot features dictionary.
+    """
     features = {}
-    joint_fts = {key: ftype for key, ftype in hw_features.items() if ftype is float}
+    joint_fts = {
+        key: ftype
+        for key, ftype in hw_features.items()
+        if ftype is float or (isinstance(ftype, PolicyFeature) and ftype.type != FeatureType.VISUAL)
+    }
     cam_fts = {key: shape for key, shape in hw_features.items() if isinstance(shape, tuple)}
 
-    if joint_fts and prefix == "action":
+    if joint_fts and prefix == ACTION:
         features[prefix] = {
             "dtype": "float32",
             "shape": (len(joint_fts),),
             "names": list(joint_fts),
         }
 
-    if joint_fts and prefix == "observation":
+    if joint_fts and prefix == OBS_STR:
         features[f"{prefix}.state"] = {
             "dtype": "float32",
             "shape": (len(joint_fts),),
@@ -479,6 +656,20 @@ def hw_to_dataset_features(
 def build_dataset_frame(
     ds_features: dict[str, dict], values: dict[str, Any], prefix: str
 ) -> dict[str, np.ndarray]:
+    """Construct a single data frame from raw values based on dataset features.
+
+    A "frame" is a dictionary containing all the data for a single timestep,
+    formatted as numpy arrays according to the feature specification.
+
+    Args:
+        ds_features (dict): The LeRobot dataset features dictionary.
+        values (dict): A dictionary of raw values from the hardware/environment.
+        prefix (str): The prefix to filter features by (e.g., "observation"
+            or "action").
+
+    Returns:
+        dict: A dictionary representing a single frame of data.
+    """
     frame = {}
     for key, ft in ds_features.items():
         if key in DEFAULT_FEATURES or not key.startswith(prefix):
@@ -492,6 +683,21 @@ def build_dataset_frame(
 
 
 def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFeature]:
+    """Convert dataset features to policy features.
+
+    This function transforms the dataset's feature specification into a format
+    that a policy can use, classifying features by type (e.g., visual, state,
+    action) and ensuring correct shapes (e.g., channel-first for images).
+
+    Args:
+        features (dict): The LeRobot dataset features dictionary.
+
+    Returns:
+        dict: A dictionary mapping feature keys to `PolicyFeature` objects.
+
+    Raises:
+        ValueError: If an image feature does not have a 3D shape.
+    """
     # TODO(aliberts): Implement "type" in dataset features and simplify this
     policy_features = {}
     for key, ft in features.items():
@@ -505,11 +711,11 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
             # Backward compatibility for "channel" which is an error introduced in LeRobotDataset v2.0 for ported datasets.
             if names[2] in ["channel", "channels"]:  # (h, w, c) -> (c, h, w)
                 shape = (shape[2], shape[0], shape[1])
-        elif key == "observation.environment_state":
+        elif key == OBS_ENV_STATE:
             type = FeatureType.ENV
-        elif key.startswith("observation"):
+        elif key.startswith(OBS_STR):
             type = FeatureType.STATE
-        elif key.startswith("action"):
+        elif key.startswith(ACTION):
             type = FeatureType.ACTION
         else:
             continue
@@ -522,6 +728,58 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
     return policy_features
 
 
+def combine_feature_dicts(*dicts: dict) -> dict:
+    """Merge LeRobot grouped feature dicts.
+
+    - For 1D numeric specs (dtype not image/video/string) with "names": we merge the names and recompute the shape.
+    - For others (e.g. `observation.images.*`), the last one wins (if they are identical).
+
+    Args:
+        *dicts: A variable number of LeRobot feature dictionaries to merge.
+
+    Returns:
+        dict: A single merged feature dictionary.
+
+    Raises:
+        ValueError: If there's a dtype mismatch for a feature being merged.
+    """
+    out: dict = {}
+    for d in dicts:
+        for key, value in d.items():
+            if not isinstance(value, dict):
+                out[key] = value
+                continue
+
+            dtype = value.get("dtype")
+            shape = value.get("shape")
+            is_vector = (
+                dtype not in ("image", "video", "string")
+                and isinstance(shape, tuple)
+                and len(shape) == 1
+                and "names" in value
+            )
+
+            if is_vector:
+                # Initialize or retrieve the accumulating dict for this feature key
+                target = out.setdefault(key, {"dtype": dtype, "names": [], "shape": (0,)})
+                # Ensure consistent data types across merged entries
+                if "dtype" in target and dtype != target["dtype"]:
+                    raise ValueError(f"dtype mismatch for '{key}': {target['dtype']} vs {dtype}")
+
+                # Merge feature names: append only new ones to preserve order without duplicates
+                seen = set(target["names"])
+                for n in value["names"]:
+                    if n not in seen:
+                        target["names"].append(n)
+                        seen.add(n)
+                # Recompute the shape to reflect the updated number of features
+                target["shape"] = (len(target["names"]),)
+            else:
+                # For images/videos and non-1D entries: override with the latest definition
+                out[key] = value
+    return out
+
+
 def create_empty_dataset_info(
     codebase_version: str,
     fps: int,
@@ -532,6 +790,18 @@ def create_empty_dataset_info(
     data_files_size_in_mb: int | None = None,
     video_files_size_in_mb: int | None = None,
 ) -> dict:
+    """Create a template dictionary for a new dataset's `info.json`.
+
+    Args:
+        codebase_version (str): The version of the LeRobot codebase.
+        fps (int): The frames per second of the data.
+        features (dict): The LeRobot features dictionary for the dataset.
+        use_videos (bool): Whether the dataset will store videos.
+        robot_type (str | None): The type of robot used, if any.
+
+    Returns:
+        dict: A dictionary with the initial dataset metadata.
+    """
     return {
         "codebase_version": codebase_version,
         "robot_type": robot_type,
@@ -552,9 +822,23 @@ def create_empty_dataset_info(
 def check_delta_timestamps(
     delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
 ) -> bool:
-    """This will check if all the values in delta_timestamps are multiples of 1/fps +/- tolerance.
-    This is to ensure that these delta_timestamps added to any timestamp from a dataset will themselves be
-    actual timestamps from the dataset.
+    """Check if delta timestamps are multiples of 1/fps +/- tolerance.
+
+    This ensures that adding these delta timestamps to any existing timestamp in
+    the dataset will result in a value that aligns with the dataset's frame rate.
+
+    Args:
+        delta_timestamps (dict): A dictionary where values are lists of time
+            deltas in seconds.
+        fps (int): The frames per second of the dataset.
+        tolerance_s (float): The allowed tolerance in seconds.
+        raise_value_error (bool): If True, raises an error on failure.
+
+    Returns:
+        bool: True if all deltas are valid, False otherwise.
+
+    Raises:
+        ValueError: If any delta is outside the tolerance and `raise_value_error` is True.
     """
     outside_tolerance = {}
     for key, delta_ts in delta_timestamps.items():
@@ -580,6 +864,15 @@ def check_delta_timestamps(
 
 
 def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dict[str, list[int]]:
+    """Convert delta timestamps in seconds to delta indices in frames.
+
+    Args:
+        delta_timestamps (dict): A dictionary of time deltas in seconds.
+        fps (int): The frames per second of the dataset.
+
+    Returns:
+        dict: A dictionary of frame delta indices.
+    """
     delta_indices = {}
     for key, delta_ts in delta_timestamps.items():
         delta_indices[key] = [round(d * fps) for d in delta_ts]
@@ -588,9 +881,17 @@ def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dic
 
 
 def cycle(iterable: Any) -> Iterator[Any]:
-    """The equivalent of itertools.cycle, but safe for Pytorch dataloaders.
+    """Create a dataloader-safe cyclical iterator.
 
-    See https://github.com/pytorch/pytorch/issues/23900 for information on why itertools.cycle is not safe.
+    This is an equivalent of `itertools.cycle` but is safe for use with
+    PyTorch DataLoaders with multiple workers.
+    See https://github.com/pytorch/pytorch/issues/23900 for details.
+
+    Args:
+        iterable: The iterable to cycle over.
+
+    Yields:
+        Items from the iterable, restarting from the beginning when exhausted.
     """
     iterator = iter(iterable)
     while True:
@@ -601,8 +902,14 @@ def cycle(iterable: Any) -> Iterator[Any]:
 
 
 def create_branch(repo_id: str, *, branch: str, repo_type: str | None = None) -> None:
-    """Create a branch on a existing Hugging Face repo. Delete the branch if it already
-    exists before creating it.
+    """Create a branch on an existing Hugging Face repo.
+
+    Deletes the branch if it already exists before creating it.
+
+    Args:
+        repo_id (str): The ID of the repository.
+        branch (str): The name of the branch to create.
+        repo_type (str | None): The type of the repository (e.g., "dataset").
     """
     api = HfApi()
 
@@ -620,9 +927,20 @@ def create_lerobot_dataset_card(
     dataset_info: dict | None = None,
     **kwargs,
 ) -> DatasetCard:
-    """
-    Keyword arguments will be used to replace values in src/lerobot/datasets/card_template.md.
-    Note: If specified, license must be one of https://huggingface.co/docs/hub/repositories-licenses.
+    """Create a `DatasetCard` for a LeRobot dataset.
+
+    Keyword arguments are used to replace values in the card template.
+    Note: If specified, `license` must be a valid license identifier from
+    https://huggingface.co/docs/hub/repositories-licenses.
+
+    Args:
+        tags (list | None): A list of tags to add to the dataset card.
+        dataset_info (dict | None): The dataset's info dictionary, which will
+            be displayed on the card.
+        **kwargs: Additional keyword arguments to populate the card template.
+
+    Returns:
+        DatasetCard: The generated dataset card object.
     """
     card_tags = ["LeRobot"]
 
@@ -675,6 +993,15 @@ def validate_frame(frame: dict, features: dict) -> None:
 
 
 def validate_features_presence(actual_features: set[str], expected_features: set[str]) -> str:
+    """Check for missing or extra features in a frame.
+
+    Args:
+        actual_features (set[str]): The set of feature names present in the frame.
+        expected_features (set[str]): The set of feature names expected in the frame.
+
+    Returns:
+        str: An error message string if there's a mismatch, otherwise an empty string.
+    """
     error_message = ""
     missing_features = expected_features - actual_features
     extra_features = actual_features - expected_features
@@ -692,6 +1019,19 @@ def validate_features_presence(actual_features: set[str], expected_features: set
 def validate_feature_dtype_and_shape(
     name: str, feature: dict, value: np.ndarray | PILImage.Image | str
 ) -> str:
+    """Validate the dtype and shape of a single feature's value.
+
+    Args:
+        name (str): The name of the feature.
+        feature (dict): The feature specification from the LeRobot features dictionary.
+        value: The value of the feature to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+
+    Raises:
+        NotImplementedError: If the feature dtype is not supported for validation.
+    """
     expected_dtype = feature["dtype"]
     expected_shape = feature["shape"]
     if is_valid_numpy_dtype_string(expected_dtype):
@@ -707,6 +1047,17 @@ def validate_feature_dtype_and_shape(
 def validate_feature_numpy_array(
     name: str, expected_dtype: str, expected_shape: list[int], value: np.ndarray
 ) -> str:
+    """Validate a feature that is expected to be a numpy array.
+
+    Args:
+        name (str): The name of the feature.
+        expected_dtype (str): The expected numpy dtype as a string.
+        expected_shape (list[int]): The expected shape.
+        value (np.ndarray): The numpy array to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
     error_message = ""
     if isinstance(value, np.ndarray):
         actual_dtype = value.dtype
@@ -726,6 +1077,18 @@ def validate_feature_numpy_array(
 def validate_feature_image_or_video(
     name: str, expected_shape: list[str], value: np.ndarray | PILImage.Image
 ) -> str:
+    """Validate a feature that is expected to be an image or video frame.
+
+    Accepts `np.ndarray` (channel-first or channel-last) or `PIL.Image.Image`.
+
+    Args:
+        name (str): The name of the feature.
+        expected_shape (list[str]): The expected shape (C, H, W).
+        value: The image data to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
     # Note: The check of pixels range ([0,1] for float and [0,255] for uint8) is done by the image writer threads.
     error_message = ""
     if isinstance(value, np.ndarray):
@@ -742,12 +1105,35 @@ def validate_feature_image_or_video(
 
 
 def validate_feature_string(name: str, value: str) -> str:
+    """Validate a feature that is expected to be a string.
+
+    Args:
+        name (str): The name of the feature.
+        value (str): The value to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
     if not isinstance(value, str):
         return f"The feature '{name}' is expected to be of type 'str', but type '{type(value)}' provided instead.\n"
     return ""
 
 
 def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict) -> None:
+    """Validate the episode buffer before it's written to disk.
+
+    Ensures the buffer has the required keys, contains at least one frame, and
+    has features consistent with the dataset's specification.
+
+    Args:
+        episode_buffer (dict): The buffer containing data for a single episode.
+        total_episodes (int): The current total number of episodes in the dataset.
+        features (dict): The LeRobot features dictionary for the dataset.
+
+    Raises:
+        ValueError: If the buffer is invalid.
+        NotImplementedError: If the episode index is manually set and doesn't match.
+    """
     if "size" not in episode_buffer:
         raise ValueError("size key not found in episode_buffer")
 
@@ -793,7 +1179,7 @@ def item_to_torch(item: dict) -> dict:
         dict: Dictionary with all tensor-like items converted to torch.Tensor.
     """
     for key, val in item.items():
-        if isinstance(val, (np.ndarray, list)) and key not in ["task"]:
+        if isinstance(val, (np.ndarray | list)) and key not in ["task"]:
             # Convert numpy arrays and lists to torch tensors
             item[key] = torch.tensor(val)
     return item
@@ -867,8 +1253,8 @@ class Backtrackable(Generic[T]):
             raise ValueError("lookahead must be > 0")
 
         self._source: Iterator[T] = iter(iterable)
-        self._back_buf: Deque[T] = deque(maxlen=history)
-        self._ahead_buf: Deque[T] = deque(maxlen=lookahead) if lookahead > 0 else deque()
+        self._back_buf: deque[T] = deque(maxlen=history)
+        self._ahead_buf: deque[T] = deque(maxlen=lookahead) if lookahead > 0 else deque()
         self._cursor: int = 0
         self._history = history
         self._lookahead = lookahead
@@ -942,12 +1328,6 @@ class Backtrackable(Generic[T]):
         # When cursor<0, slice so the order remains chronological
         return list(self._back_buf)[: self._cursor or None]
 
-    def lookahead_buffer(self) -> list[T]:
-        """
-        Return a copy of the current lookahead buffer.
-        """
-        return list(self._ahead_buf)
-
     def can_peek_back(self, steps: int = 1) -> bool:
         """
         Check if we can go back `steps` items without raising an IndexError.
@@ -973,31 +1353,6 @@ class Backtrackable(Generic[T]):
         except StopIteration:
             return False
 
-    def reset_cursor(self) -> None:
-        """
-        Reset cursor to the most recent position (equivalent to calling next()
-        until you're back to the latest item).
-        """
-        self._cursor = 0
-
-    def clear_ahead_buffer(self) -> None:
-        """
-        Clear the ahead buffer, discarding any pre-fetched items.
-        """
-        self._ahead_buf.clear()
-
-    def switch_source_iterable(self, new_source: Iterable[T]) -> None:
-        """
-        Switch the source of the backtrackable to a new iterable, keeping the history.
-
-        This is useful when iterating over a sequence of datasets. The history from the
-        previous source is kept, but the lookahead buffer is cleared. The cursor is reset
-        to the present.
-        """
-        self._source = iter(new_source)
-        self.clear_ahead_buffer()
-        self.reset_cursor()
-
 
 def safe_shard(dataset: datasets.IterableDataset, index: int, num_shards: int) -> datasets.Dataset:
     """

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -34,6 +34,7 @@ python src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py \
 """
 
 import argparse
+import logging
 import shutil
 from pathlib import Path
 from typing import Any
@@ -46,7 +47,6 @@ from datasets import Dataset, Features, Image
 from huggingface_hub import HfApi, snapshot_download
 from requests import HTTPError
 
-from lerobot.constants import HF_LEROBOT_HOME
 from lerobot.datasets.compute_stats import aggregate_stats
 from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
 from lerobot.datasets.utils import (
@@ -71,6 +71,8 @@ from lerobot.datasets.utils import (
     write_tasks,
 )
 from lerobot.datasets.video_utils import concatenate_video_files, get_video_duration_in_s
+from lerobot.utils.constants import HF_LEROBOT_HOME
+from lerobot.utils.utils import init_logging
 
 V21 = "v2.1"
 
@@ -144,6 +146,7 @@ def legacy_load_tasks(local_dir: Path) -> tuple[dict, dict]:
 
 
 def convert_tasks(root, new_root):
+    logging.info(f"Converting tasks from {root} to {new_root}")
     tasks, _ = legacy_load_tasks(root)
     task_indices = tasks.keys()
     task_strings = tasks.values()
@@ -185,7 +188,10 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
     num_frames = 0
     paths_to_cat = []
     episodes_metadata = []
-    for ep_path in ep_paths:
+
+    logging.info(f"Converting data files from {len(ep_paths)} episodes")
+
+    for ep_path in tqdm.tqdm(ep_paths, desc="convert data files"):
         ep_size_in_mb = get_parquet_file_size_in_mb(ep_path)
         ep_num_frames = get_parquet_num_frames(ep_path)
         ep_metadata = {
@@ -209,7 +215,6 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
 
         # Reset for the next file
         size_in_mb = ep_size_in_mb
-        num_frames = ep_num_frames
         paths_to_cat = [ep_path]
 
         chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
@@ -236,6 +241,8 @@ def get_image_keys(root):
 
 
 def convert_videos(root: Path, new_root: Path, video_file_size_in_mb: int):
+    logging.info(f"Converting videos from {root} to {new_root}")
+
     video_keys = get_video_keys(root)
     if len(video_keys) == 0:
         return None
@@ -254,7 +261,7 @@ def convert_videos(root: Path, new_root: Path, video_file_size_in_mb: int):
     episods_metadata = []
     num_cameras = len(video_keys)
     num_episodes = num_eps_per_cam[0]
-    for ep_idx in range(num_episodes):
+    for ep_idx in tqdm.tqdm(range(num_episodes), desc="convert videos"):
         # Sanity check
         ep_ids = [eps_metadata_per_cam[cam_idx][ep_idx]["episode_index"] for cam_idx in range(num_cameras)]
         ep_ids += [ep_idx]
@@ -281,6 +288,7 @@ def convert_videos_of_camera(root: Path, new_root: Path, video_key: str, video_f
     duration_in_s = 0.0
     paths_to_cat = []
     episodes_metadata = []
+
     for ep_path in tqdm.tqdm(ep_paths, desc=f"convert videos of {video_key}"):
         ep_size_in_mb = get_video_size_in_mb(ep_path)
         ep_duration_in_s = get_video_duration_in_s(ep_path)
@@ -374,6 +382,8 @@ def generate_episode_metadata_dict(
 
 
 def convert_episodes_metadata(root, new_root, episodes_metadata, episodes_video_metadata=None):
+    logging.info(f"Converting episodes metadata from {root} to {new_root}")
+
     episodes_legacy_metadata = legacy_load_episodes(root)
     episodes_stats = legacy_load_episodes_stats(root)
 
@@ -404,7 +414,8 @@ def convert_info(root, new_root, data_file_size_in_mb, video_file_size_in_mb):
     info["video_files_size_in_mb"] = video_file_size_in_mb
     info["data_path"] = DEFAULT_DATA_PATH
     info["video_path"] = DEFAULT_VIDEO_PATH
-    info["fps"] = float(info["fps"])
+    info["fps"] = int(info["fps"])
+    logging.info(f"Converting info from {root} to {new_root}")
     for key in info["features"]:
         if info["features"][key]["dtype"] == "video":
             # already has fps in video_info
@@ -469,6 +480,7 @@ def convert_dataset(
 
 
 if __name__ == "__main__":
+    init_logging()
     parser = argparse.ArgumentParser()
     parser.add_argument(
         "--repo-id",

src/lerobot/datasets/video_utils.py

@@ -428,7 +428,7 @@ def concatenate_video_files(
     with tempfile.NamedTemporaryFile(mode="w", suffix=".ffconcat", delete=False) as tmp_concatenate_file:
         tmp_concatenate_file.write("ffconcat version 1.0\n")
         for input_path in input_video_paths:
-            tmp_concatenate_file.write(f"file '{str(input_path)}'\n")
+            tmp_concatenate_file.write(f"file '{str(input_path.resolve())}'\n")
         tmp_concatenate_file.flush()
         tmp_concatenate_path = tmp_concatenate_file.name
 
@@ -437,7 +437,9 @@ def concatenate_video_files(
         tmp_concatenate_path, mode="r", format="concat", options={"safe": "0"}
     )  # safe = 0 allows absolute paths as well as relative paths
 
-    tmp_output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False).name
+    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_named_file:
+        tmp_output_video_path = tmp_named_file.name
+
     output_container = av.open(
         tmp_output_video_path, mode="w", options={"movflags": "faststart"}
     )  # faststart is to move the metadata to the beginning of the file to speed up loading
@@ -585,19 +587,6 @@ def get_video_pixel_channels(pix_fmt: str) -> int:
         raise ValueError("Unknown format")
 
 
-def get_image_pixel_channels(image: Image):
-    if image.mode == "L":
-        return 1  # Grayscale
-    elif image.mode == "LA":
-        return 2  # Grayscale + Alpha
-    elif image.mode == "RGB":
-        return 3  # RGB
-    elif image.mode == "RGBA":
-        return 4  # RGBA
-    else:
-        raise ValueError("Unknown format")
-
-
 def get_video_duration_in_s(video_path: Path | str) -> float:
     """
     Get the duration of a video file in seconds using PyAV.

src/lerobot/envs/configs.py

@@ -19,9 +19,9 @@ from typing import Any
 import draccus
 
 from lerobot.configs.types import FeatureType, PolicyFeature
-from lerobot.constants import ACTION, OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
 from lerobot.robots import RobotConfig
 from lerobot.teleoperators.config import TeleoperatorConfig
+from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
 
 
 @dataclass
@@ -30,6 +30,8 @@ class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
     fps: int = 30
     features: dict[str, PolicyFeature] = field(default_factory=dict)
     features_map: dict[str, str] = field(default_factory=dict)
+    max_parallel_tasks: int = 1
+    disable_env_checker: bool = True
 
     @property
     def type(self) -> str:
@@ -51,12 +53,12 @@ class AlohaEnv(EnvConfig):
     render_mode: str = "rgb_array"
     features: dict[str, PolicyFeature] = field(
         default_factory=lambda: {
-            "action": PolicyFeature(type=FeatureType.ACTION, shape=(14,)),
+            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(14,)),
         }
     )
     features_map: dict[str, str] = field(
         default_factory=lambda: {
-            "action": ACTION,
+            ACTION: ACTION,
             "agent_pos": OBS_STATE,
             "top": f"{OBS_IMAGE}.top",
             "pixels/top": f"{OBS_IMAGES}.top",
@@ -91,13 +93,13 @@ class PushtEnv(EnvConfig):
     visualization_height: int = 384
     features: dict[str, PolicyFeature] = field(
         default_factory=lambda: {
-            "action": PolicyFeature(type=FeatureType.ACTION, shape=(2,)),
+            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(2,)),
             "agent_pos": PolicyFeature(type=FeatureType.STATE, shape=(2,)),
         }
     )
     features_map: dict[str, str] = field(
         default_factory=lambda: {
-            "action": ACTION,
+            ACTION: ACTION,
             "agent_pos": OBS_STATE,
             "environment_state": OBS_ENV_STATE,
             "pixels": OBS_IMAGE,
@@ -133,13 +135,13 @@ class XarmEnv(EnvConfig):
     visualization_height: int = 384
     features: dict[str, PolicyFeature] = field(
         default_factory=lambda: {
-            "action": PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
+            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
             "pixels": PolicyFeature(type=FeatureType.VISUAL, shape=(84, 84, 3)),
         }
     )
     features_map: dict[str, str] = field(
         default_factory=lambda: {
-            "action": ACTION,
+            ACTION: ACTION,
             "agent_pos": OBS_STATE,
             "pixels": OBS_IMAGE,
         }
@@ -161,33 +163,69 @@ class XarmEnv(EnvConfig):
 
 
 @dataclass
-class VideoRecordConfig:
-    """Configuration for video recording in ManiSkill environments."""
-
-    enabled: bool = False
-    record_dir: str = "videos"
-    trajectory_name: str = "trajectory"
+class ImagePreprocessingConfig:
+    crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None
+    resize_size: tuple[int, int] | None = None
 
 
 @dataclass
-class EnvTransformConfig:
-    """Configuration for environment wrappers."""
+class RewardClassifierConfig:
+    """Configuration for reward classification."""
+
+    pretrained_path: str | None = None
+    success_threshold: float = 0.5
+    success_reward: float = 1.0
+
+
+@dataclass
+class InverseKinematicsConfig:
+    """Configuration for inverse kinematics processing."""
+
+    urdf_path: str | None = None
+    target_frame_name: str | None = None
+    end_effector_bounds: dict[str, list[float]] | None = None
+    end_effector_step_sizes: dict[str, float] | None = None
+
+
+@dataclass
+class ObservationConfig:
+    """Configuration for observation processing."""
 
-    # ee_action_space_params: EEActionSpaceConfig = field(default_factory=EEActionSpaceConfig)
-    control_mode: str = "gamepad"
-    display_cameras: bool = False
     add_joint_velocity_to_observation: bool = False
     add_current_to_observation: bool = False
-    add_ee_pose_to_observation: bool = False
-    crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None
-    resize_size: tuple[int, int] | None = None
-    control_time_s: float = 20.0
+    display_cameras: bool = False
+
+
+@dataclass
+class GripperConfig:
+    """Configuration for gripper control and penalties."""
+
+    use_gripper: bool = True
+    gripper_penalty: float = 0.0
+
+
+@dataclass
+class ResetConfig:
+    """Configuration for environment reset behavior."""
+
     fixed_reset_joint_positions: Any | None = None
     reset_time_s: float = 5.0
-    use_gripper: bool = True
-    gripper_quantization_threshold: float | None = 0.8
-    gripper_penalty: float = 0.0
-    gripper_penalty_in_reward: bool = False
+    control_time_s: float = 20.0
+    terminate_on_success: bool = True
+
+
+@dataclass
+class HILSerlProcessorConfig:
+    """Configuration for environment processing pipeline."""
+
+    control_mode: str = "gamepad"
+    observation: ObservationConfig | None = None
+    image_preprocessing: ImagePreprocessingConfig | None = None
+    gripper: GripperConfig | None = None
+    reset: ResetConfig | None = None
+    inverse_kinematics: InverseKinematicsConfig | None = None
+    reward_classifier: RewardClassifierConfig | None = None
+    max_gripper_pos: float | None = 100.0
 
 
 @EnvConfig.register_subclass(name="gym_manipulator")
@@ -197,77 +235,62 @@ class HILSerlRobotEnvConfig(EnvConfig):
 
     robot: RobotConfig | None = None
     teleop: TeleoperatorConfig | None = None
-    wrapper: EnvTransformConfig | None = None
-    fps: int = 10
+    processor: HILSerlProcessorConfig = field(default_factory=HILSerlProcessorConfig)
+
     name: str = "real_robot"
-    mode: str | None = None  # Either "record", "replay", None
-    repo_id: str | None = None
-    dataset_root: str | None = None
-    task: str | None = ""
-    num_episodes: int = 10  # only for record mode
-    episode: int = 0
-    device: str = "cuda"
-    push_to_hub: bool = True
-    pretrained_policy_name_or_path: str | None = None
-    reward_classifier_pretrained_path: str | None = None
-    # For the reward classifier, to record more positive examples after a success
-    number_of_steps_after_success: int = 0
 
     @property
     def gym_kwargs(self) -> dict:
         return {}
 
 
-@EnvConfig.register_subclass("hil")
+@EnvConfig.register_subclass("libero")
 @dataclass
-class HILEnvConfig(EnvConfig):
-    """Configuration for the HIL environment."""
-
-    name: str = "PandaPickCube"
-    task: str | None = "PandaPickCubeKeyboard-v0"
-    use_viewer: bool = True
-    gripper_penalty: float = 0.0
-    use_gamepad: bool = True
-    state_dim: int = 18
-    action_dim: int = 4
-    fps: int = 100
-    episode_length: int = 100
-    video_record: VideoRecordConfig = field(default_factory=VideoRecordConfig)
+class LiberoEnv(EnvConfig):
+    task: str = "libero_10"  # can also choose libero_spatial, libero_object, etc.
+    fps: int = 30
+    episode_length: int = 520
+    obs_type: str = "pixels_agent_pos"
+    render_mode: str = "rgb_array"
+    camera_name: str = "agentview_image,robot0_eye_in_hand_image"
+    init_states: bool = True
+    camera_name_mapping: dict[str, str] | None = (None,)
     features: dict[str, PolicyFeature] = field(
         default_factory=lambda: {
-            "action": PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
-            "observation.image": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 128, 128)),
-            "observation.state": PolicyFeature(type=FeatureType.STATE, shape=(18,)),
+            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
         }
     )
     features_map: dict[str, str] = field(
         default_factory=lambda: {
-            "action": ACTION,
-            "observation.image": OBS_IMAGE,
-            "observation.state": OBS_STATE,
+            ACTION: ACTION,
+            "agent_pos": OBS_STATE,
+            "pixels/agentview_image": f"{OBS_IMAGES}.image",
+            "pixels/robot0_eye_in_hand_image": f"{OBS_IMAGES}.image2",
         }
     )
-    ################# args from hilserlrobotenv
-    reward_classifier_pretrained_path: str | None = None
-    robot_config: RobotConfig | None = None
-    teleop_config: TeleoperatorConfig | None = None
-    wrapper: EnvTransformConfig | None = None
-    mode: str | None = None  # Either "record", "replay", None
-    repo_id: str | None = None
-    dataset_root: str | None = None
-    num_episodes: int = 10  # only for record mode
-    episode: int = 0
-    device: str = "cuda"
-    push_to_hub: bool = True
-    pretrained_policy_name_or_path: str | None = None
-    # For the reward classifier, to record more positive examples after a success
-    number_of_steps_after_success: int = 0
-    ############################
+
+    def __post_init__(self):
+        if self.obs_type == "pixels":
+            self.features["pixels/agentview_image"] = PolicyFeature(
+                type=FeatureType.VISUAL, shape=(360, 360, 3)
+            )
+            self.features["pixels/robot0_eye_in_hand_image"] = PolicyFeature(
+                type=FeatureType.VISUAL, shape=(360, 360, 3)
+            )
+        elif self.obs_type == "pixels_agent_pos":
+            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(8,))
+            self.features["pixels/agentview_image"] = PolicyFeature(
+                type=FeatureType.VISUAL, shape=(360, 360, 3)
+            )
+            self.features["pixels/robot0_eye_in_hand_image"] = PolicyFeature(
+                type=FeatureType.VISUAL, shape=(360, 360, 3)
+            )
+        else:
+            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
 
     @property
     def gym_kwargs(self) -> dict:
         return {
-            "use_viewer": self.use_viewer,
-            "use_gamepad": self.use_gamepad,
-            "gripper_penalty": self.gripper_penalty,
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
         }

src/lerobot/envs/factory.py

@@ -17,7 +17,7 @@ import importlib
 
 import gymnasium as gym
 
-from lerobot.envs.configs import AlohaEnv, EnvConfig, HILEnvConfig, PushtEnv, XarmEnv
+from lerobot.envs.configs import AlohaEnv, EnvConfig, LiberoEnv, PushtEnv, XarmEnv
 
 
 def make_env_config(env_type: str, **kwargs) -> EnvConfig:
@@ -27,13 +27,15 @@ def make_env_config(env_type: str, **kwargs) -> EnvConfig:
         return PushtEnv(**kwargs)
     elif env_type == "xarm":
         return XarmEnv(**kwargs)
-    elif env_type == "hil":
-        return HILEnvConfig(**kwargs)
+    elif env_type == "libero":
+        return LiberoEnv(**kwargs)
     else:
         raise ValueError(f"Policy type '{env_type}' is not available.")
 
 
-def make_env(cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False) -> gym.vector.VectorEnv | None:
+def make_env(
+    cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False
+) -> dict[str, dict[int, gym.vector.VectorEnv]]:
     """Makes a gym vector environment according to the config.
 
     Args:
@@ -47,13 +49,30 @@ def make_env(cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False) -> g
         ModuleNotFoundError: If the requested env package is not installed
 
     Returns:
-        gym.vector.VectorEnv: The parallelized gym.env instance.
+        dict[str, dict[int, gym.vector.VectorEnv]]:
+            A mapping from suite name to indexed vectorized environments.
+            - For multi-task benchmarks (e.g., LIBERO): one entry per suite, and one vec env per task_id.
+            - For single-task environments: a single suite entry (cfg.type) with task_id=0.
+
     """
     if n_envs < 1:
-        raise ValueError("`n_envs must be at least 1")
+        raise ValueError("`n_envs` must be at least 1")
+
+    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
+
+    if "libero" in cfg.type:
+        from lerobot.envs.libero import create_libero_envs
+
+        return create_libero_envs(
+            task=cfg.task,
+            n_envs=n_envs,
+            camera_name=cfg.camera_name,
+            init_states=cfg.init_states,
+            gym_kwargs=cfg.gym_kwargs,
+            env_cls=env_cls,
+        )
 
     package_name = f"gym_{cfg.type}"
-
     try:
         importlib.import_module(package_name)
     except ModuleNotFoundError as e:
@@ -62,10 +81,11 @@ def make_env(cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False) -> g
 
     gym_handle = f"{package_name}/{cfg.task}"
 
-    # batched version of the env that returns an observation of shape (b, c)
-    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
-    env = env_cls(
-        [lambda: gym.make(gym_handle, disable_env_checker=True, **cfg.gym_kwargs) for _ in range(n_envs)]
-    )
+    def _make_one():
+        return gym.make(gym_handle, disable_env_checker=cfg.disable_env_checker, **(cfg.gym_kwargs or {}))
 
-    return env
+    vec = env_cls([_make_one for _ in range(n_envs)])
+
+    # normalize to {suite: {task_id: vec_env}} for consistency
+    suite_name = cfg.type  # e.g., "pusht", "aloha"
+    return {suite_name: {0: vec}}

src/lerobot/envs/libero.py

@@ -0,0 +1,377 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import annotations
+
+import os
+from collections import defaultdict
+from collections.abc import Callable, Iterable, Mapping, Sequence
+from functools import partial
+from pathlib import Path
+from typing import Any
+
+import gymnasium as gym
+import numpy as np
+import torch
+from gymnasium import spaces
+from libero.libero import benchmark, get_libero_path
+from libero.libero.envs import OffScreenRenderEnv
+from robosuite.utils.transform_utils import quat2axisangle
+
+
+def _parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:
+    """Normalize camera_name into a non-empty list of strings."""
+    if isinstance(camera_name, str):
+        cams = [c.strip() for c in camera_name.split(",") if c.strip()]
+    elif isinstance(camera_name, (list | tuple)):
+        cams = [str(c).strip() for c in camera_name if str(c).strip()]
+    else:
+        raise TypeError(f"camera_name must be str or sequence[str], got {type(camera_name).__name__}")
+    if not cams:
+        raise ValueError("camera_name resolved to an empty list.")
+    return cams
+
+
+def _get_suite(name: str) -> benchmark.Benchmark:
+    """Instantiate a LIBERO suite by name with clear validation."""
+    bench = benchmark.get_benchmark_dict()
+    if name not in bench:
+        raise ValueError(f"Unknown LIBERO suite '{name}'. Available: {', '.join(sorted(bench.keys()))}")
+    suite = bench[name]()
+    if not getattr(suite, "tasks", None):
+        raise ValueError(f"Suite '{name}' has no tasks.")
+    return suite
+
+
+def _select_task_ids(total_tasks: int, task_ids: Iterable[int] | None) -> list[int]:
+    """Validate/normalize task ids. If None → all tasks."""
+    if task_ids is None:
+        return list(range(total_tasks))
+    ids = sorted({int(t) for t in task_ids})
+    for t in ids:
+        if t < 0 or t >= total_tasks:
+            raise ValueError(f"task_id {t} out of range [0, {total_tasks - 1}].")
+    return ids
+
+
+def get_task_init_states(task_suite: Any, i: int) -> np.ndarray:
+    init_states_path = (
+        Path(get_libero_path("init_states"))
+        / task_suite.tasks[i].problem_folder
+        / task_suite.tasks[i].init_states_file
+    )
+    init_states = torch.load(init_states_path, weights_only=False)  # nosec B614
+    return init_states
+
+
+def get_libero_dummy_action():
+    """Get dummy/no-op action, used to roll out the simulation while the robot does nothing."""
+    return [0, 0, 0, 0, 0, 0, -1]
+
+
+OBS_STATE_DIM = 8
+ACTION_DIM = 7
+AGENT_POS_LOW = -1000.0
+AGENT_POS_HIGH = 1000.0
+ACTION_LOW = -1.0
+ACTION_HIGH = 1.0
+TASK_SUITE_MAX_STEPS: dict[str, int] = {
+    "libero_spatial": 280,  # longest training demo has 193 steps
+    "libero_object": 280,  # longest training demo has 254 steps
+    "libero_goal": 300,  # longest training demo has 270 steps
+    "libero_10": 520,  # longest training demo has 505 steps
+    "libero_90": 400,  # longest training demo has 373 steps
+}
+
+
+class LiberoEnv(gym.Env):
+    metadata = {"render_modes": ["rgb_array"], "render_fps": 80}
+
+    def __init__(
+        self,
+        task_suite: Any,
+        task_id: int,
+        task_suite_name: str,
+        camera_name: str | Sequence[str] = "agentview_image,robot0_eye_in_hand_image",
+        obs_type: str = "pixels",
+        render_mode: str = "rgb_array",
+        observation_width: int = 256,
+        observation_height: int = 256,
+        visualization_width: int = 640,
+        visualization_height: int = 480,
+        init_states: bool = True,
+        episode_index: int = 0,
+        camera_name_mapping: dict[str, str] | None = None,
+        num_steps_wait: int = 10,
+    ):
+        super().__init__()
+        self.task_id = task_id
+        self.obs_type = obs_type
+        self.render_mode = render_mode
+        self.observation_width = observation_width
+        self.observation_height = observation_height
+        self.visualization_width = visualization_width
+        self.visualization_height = visualization_height
+        self.init_states = init_states
+        self.camera_name = _parse_camera_names(
+            camera_name
+        )  # agentview_image (main) or robot0_eye_in_hand_image (wrist)
+
+        # Map raw camera names to "image1" and "image2".
+        # The preprocessing step `preprocess_observation` will then prefix these with `.images.*`,
+        # following the LeRobot convention (e.g., `observation.images.image`, `observation.images.image2`).
+        # This ensures the policy consistently receives observations in the
+        # expected format regardless of the original camera naming.
+        if camera_name_mapping is None:
+            camera_name_mapping = {
+                "agentview_image": "image",
+                "robot0_eye_in_hand_image": "image2",
+            }
+        self.camera_name_mapping = camera_name_mapping
+        self.num_steps_wait = num_steps_wait
+        self.episode_index = episode_index
+        # Load once and keep
+        self._init_states = get_task_init_states(task_suite, self.task_id) if self.init_states else None
+        self._init_state_id = self.episode_index  # tie each sub-env to a fixed init state
+
+        self._env = self._make_envs_task(task_suite, self.task_id)
+        default_steps = 500
+        self._max_episode_steps = TASK_SUITE_MAX_STEPS.get(task_suite_name, default_steps)
+
+        images = {}
+        for cam in self.camera_name:
+            images[self.camera_name_mapping[cam]] = spaces.Box(
+                low=0,
+                high=255,
+                shape=(self.observation_height, self.observation_width, 3),
+                dtype=np.uint8,
+            )
+
+        if self.obs_type == "state":
+            raise NotImplementedError(
+                "The 'state' observation type is not supported in LiberoEnv. "
+                "Please switch to an image-based obs_type (e.g. 'pixels', 'pixels_agent_pos')."
+            )
+
+        elif self.obs_type == "pixels":
+            self.observation_space = spaces.Dict(
+                {
+                    "pixels": spaces.Dict(images),
+                }
+            )
+        elif self.obs_type == "pixels_agent_pos":
+            self.observation_space = spaces.Dict(
+                {
+                    "pixels": spaces.Dict(images),
+                    "agent_pos": spaces.Box(
+                        low=AGENT_POS_LOW,
+                        high=AGENT_POS_HIGH,
+                        shape=(OBS_STATE_DIM,),
+                        dtype=np.float64,
+                    ),
+                }
+            )
+
+        self.action_space = spaces.Box(
+            low=ACTION_LOW, high=ACTION_HIGH, shape=(ACTION_DIM,), dtype=np.float32
+        )
+
+    def render(self):
+        raw_obs = self._env.env._get_observations()
+        image = self._format_raw_obs(raw_obs)["pixels"]["image"]
+        return image
+
+    def _make_envs_task(self, task_suite: Any, task_id: int = 0):
+        task = task_suite.get_task(task_id)
+        self.task = task.name
+        self.task_description = task.language
+        task_bddl_file = os.path.join(get_libero_path("bddl_files"), task.problem_folder, task.bddl_file)
+
+        env_args = {
+            "bddl_file_name": task_bddl_file,
+            "camera_heights": self.observation_height,
+            "camera_widths": self.observation_width,
+        }
+        env = OffScreenRenderEnv(**env_args)
+        env.reset()
+        return env
+
+    def _format_raw_obs(self, raw_obs: dict[str, Any]) -> dict[str, Any]:
+        images = {}
+        for camera_name in self.camera_name:
+            image = raw_obs[camera_name]
+            image = image[::-1, ::-1]  # rotate 180 degrees
+            images[self.camera_name_mapping[camera_name]] = image
+        state = np.concatenate(
+            (
+                raw_obs["robot0_eef_pos"],
+                quat2axisangle(raw_obs["robot0_eef_quat"]),
+                raw_obs["robot0_gripper_qpos"],
+            )
+        )
+        agent_pos = state
+        if self.obs_type == "pixels":
+            return {"pixels": images.copy()}
+        if self.obs_type == "pixels_agent_pos":
+            return {
+                "pixels": images.copy(),
+                "agent_pos": agent_pos,
+            }
+        raise NotImplementedError(
+            f"The observation type '{self.obs_type}' is not supported in LiberoEnv. "
+            "Please switch to an image-based obs_type (e.g. 'pixels', 'pixels_agent_pos')."
+        )
+
+    def reset(self, seed=None, **kwargs):
+        super().reset(seed=seed)
+        self._env.seed(seed)
+        if self.init_states and self._init_states is not None:
+            self._env.set_init_state(self._init_states[self._init_state_id])
+        raw_obs = self._env.reset()
+
+        # After reset, objects may be unstable (slightly floating, intersecting, etc.).
+        # Step the simulator with a no-op action for a few frames so everything settles.
+        # Increasing this value can improve determinism and reproducibility across resets.
+        for _ in range(self.num_steps_wait):
+            raw_obs, _, _, _ = self._env.step(get_libero_dummy_action())
+        observation = self._format_raw_obs(raw_obs)
+        info = {"is_success": False}
+        return observation, info
+
+    def step(self, action: np.ndarray) -> tuple[dict[str, Any], float, bool, bool, dict[str, Any]]:
+        if action.ndim != 1:
+            raise ValueError(
+                f"Expected action to be 1-D (shape (action_dim,)), "
+                f"but got shape {action.shape} with ndim={action.ndim}"
+            )
+        raw_obs, reward, done, info = self._env.step(action)
+
+        is_success = self._env.check_success()
+        terminated = done or is_success
+        info["is_success"] = is_success
+
+        observation = self._format_raw_obs(raw_obs)
+        if done:
+            self.reset()
+            info.update(
+                {
+                    "task": self.task,
+                    "task_id": self.task_id,
+                    "done": done,
+                    "is_success": is_success,
+                }
+            )
+        truncated = False
+        return observation, reward, terminated, truncated, info
+
+    def close(self):
+        self._env.close()
+
+
+def _make_env_fns(
+    *,
+    suite,
+    suite_name: str,
+    task_id: int,
+    n_envs: int,
+    camera_names: list[str],
+    init_states: bool,
+    gym_kwargs: Mapping[str, Any],
+) -> list[Callable[[], LiberoEnv]]:
+    """Build n_envs factory callables for a single (suite, task_id)."""
+
+    def _make_env(episode_index: int, **kwargs) -> LiberoEnv:
+        local_kwargs = dict(kwargs)
+        return LiberoEnv(
+            task_suite=suite,
+            task_id=task_id,
+            task_suite_name=suite_name,
+            camera_name=camera_names,
+            init_states=init_states,
+            episode_index=episode_index,
+            **local_kwargs,
+        )
+
+    fns: list[Callable[[], LiberoEnv]] = []
+    for episode_index in range(n_envs):
+        fns.append(partial(_make_env, episode_index, **gym_kwargs))
+    return fns
+
+
+# ---- Main API ----------------------------------------------------------------
+
+
+def create_libero_envs(
+    task: str,
+    n_envs: int,
+    gym_kwargs: dict[str, Any] | None = None,
+    camera_name: str | Sequence[str] = "agentview_image,robot0_eye_in_hand_image",
+    init_states: bool = True,
+    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
+) -> dict[str, dict[int, Any]]:
+    """
+    Create vectorized LIBERO environments with a consistent return shape.
+
+    Returns:
+        dict[suite_name][task_id] -> vec_env (env_cls([...]) with exactly n_envs factories)
+    Notes:
+        - n_envs is the number of rollouts *per task* (episode_index = 0..n_envs-1).
+        - `task` can be a single suite or a comma-separated list of suites.
+        - You may pass `task_ids` (list[int]) inside `gym_kwargs` to restrict tasks per suite.
+    """
+    if env_cls is None or not callable(env_cls):
+        raise ValueError("env_cls must be a callable that wraps a list of environment factory callables.")
+    if not isinstance(n_envs, int) or n_envs <= 0:
+        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
+
+    gym_kwargs = dict(gym_kwargs or {})
+    task_ids_filter = gym_kwargs.pop("task_ids", None)  # optional: limit to specific tasks
+
+    camera_names = _parse_camera_names(camera_name)
+    suite_names = [s.strip() for s in str(task).split(",") if s.strip()]
+    if not suite_names:
+        raise ValueError("`task` must contain at least one LIBERO suite name.")
+
+    print(
+        f"Creating LIBERO envs | suites={suite_names} | n_envs(per task)={n_envs} | init_states={init_states}"
+    )
+    if task_ids_filter is not None:
+        print(f"Restricting to task_ids={task_ids_filter}")
+
+    out: dict[str, dict[int, Any]] = defaultdict(dict)
+
+    for suite_name in suite_names:
+        suite = _get_suite(suite_name)
+        total = len(suite.tasks)
+        selected = _select_task_ids(total, task_ids_filter)
+
+        if not selected:
+            raise ValueError(f"No tasks selected for suite '{suite_name}' (available: {total}).")
+
+        for tid in selected:
+            fns = _make_env_fns(
+                suite=suite,
+                suite_name=suite_name,
+                task_id=tid,
+                n_envs=n_envs,
+                camera_names=camera_names,
+                init_states=init_states,
+                gym_kwargs=gym_kwargs,
+            )
+            out[suite_name][tid] = env_cls(fns)
+            print(f"Built vec env | suite={suite_name} | task_id={tid} | n_envs={n_envs}")
+
+    # return plain dicts for predictability
+    return {suite: dict(task_map) for suite, task_map in out.items()}

src/lerobot/envs/utils.py

@@ -14,6 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import warnings
+from collections.abc import Mapping, Sequence
+from functools import singledispatch
 from typing import Any
 
 import einops
@@ -24,6 +26,7 @@ from torch import Tensor
 
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.envs.configs import EnvConfig
+from lerobot.utils.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
 from lerobot.utils.utils import get_channel_first_image_shape
 
 
@@ -39,9 +42,9 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
     return_observations = {}
     if "pixels" in observations:
         if isinstance(observations["pixels"], dict):
-            imgs = {f"observation.images.{key}": img for key, img in observations["pixels"].items()}
+            imgs = {f"{OBS_IMAGES}.{key}": img for key, img in observations["pixels"].items()}
         else:
-            imgs = {"observation.image": observations["pixels"]}
+            imgs = {OBS_IMAGE: observations["pixels"]}
 
         for imgkey, img in imgs.items():
             # TODO(aliberts, rcadene): use transforms.ToTensor()?
@@ -70,13 +73,13 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
         if env_state.dim() == 1:
             env_state = env_state.unsqueeze(0)
 
-        return_observations["observation.environment_state"] = env_state
+        return_observations[OBS_ENV_STATE] = env_state
 
     # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing
     agent_pos = torch.from_numpy(observations["agent_pos"]).float()
     if agent_pos.dim() == 1:
         agent_pos = agent_pos.unsqueeze(0)
-    return_observations["observation.state"] = agent_pos
+    return_observations[OBS_STATE] = agent_pos
 
     return return_observations
 
@@ -127,10 +130,68 @@ def check_env_attributes_and_types(env: gym.vector.VectorEnv) -> None:
 def add_envs_task(env: gym.vector.VectorEnv, observation: dict[str, Any]) -> dict[str, Any]:
     """Adds task feature to the observation dict with respect to the first environment attribute."""
     if hasattr(env.envs[0], "task_description"):
-        observation["task"] = env.call("task_description")
+        task_result = env.call("task_description")
+
+        if isinstance(task_result, tuple):
+            task_result = list(task_result)
+
+        if not isinstance(task_result, list):
+            raise TypeError(f"Expected task_description to return a list, got {type(task_result)}")
+        if not all(isinstance(item, str) for item in task_result):
+            raise TypeError("All items in task_description result must be strings")
+
+        observation["task"] = task_result
     elif hasattr(env.envs[0], "task"):
-        observation["task"] = env.call("task")
+        task_result = env.call("task")
+
+        if isinstance(task_result, tuple):
+            task_result = list(task_result)
+
+        if not isinstance(task_result, list):
+            raise TypeError(f"Expected task to return a list, got {type(task_result)}")
+        if not all(isinstance(item, str) for item in task_result):
+            raise TypeError("All items in task result must be strings")
+
+        observation["task"] = task_result
     else:  #  For envs without language instructions, e.g. aloha transfer cube and etc.
         num_envs = observation[list(observation.keys())[0]].shape[0]
         observation["task"] = ["" for _ in range(num_envs)]
     return observation
+
+
+def _close_single_env(env: Any) -> None:
+    try:
+        env.close()
+    except Exception as exc:
+        print(f"Exception while closing env {env}: {exc}")
+
+
+@singledispatch
+def close_envs(obj: Any) -> None:
+    """Default: raise if the type is not recognized."""
+    raise NotImplementedError(f"close_envs not implemented for type {type(obj).__name__}")
+
+
+@close_envs.register
+def _(env: Mapping) -> None:
+    for v in env.values():
+        if isinstance(v, Mapping):
+            close_envs(v)
+        elif hasattr(v, "close"):
+            _close_single_env(v)
+
+
+@close_envs.register
+def _(envs: Sequence) -> None:
+    if isinstance(envs, (str | bytes)):
+        return
+    for v in envs:
+        if isinstance(v, Mapping) or isinstance(v, Sequence) and not isinstance(v, (str | bytes)):
+            close_envs(v)
+        elif hasattr(v, "close"):
+            _close_single_env(v)
+
+
+@close_envs.register
+def _(env: gym.Env) -> None:
+    _close_single_env(env)

src/lerobot/motors/dynamixel/dynamixel.py

@@ -22,7 +22,7 @@ import logging
 from copy import deepcopy
 from enum import Enum
 
-from lerobot.utils.encoding_utils import decode_twos_complement, encode_twos_complement
+from lerobot.motors.encoding_utils import decode_twos_complement, encode_twos_complement
 
 from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
 from .tables import (

src/lerobot/motors/feetech/feetech.py

@@ -17,7 +17,7 @@ from copy import deepcopy
 from enum import Enum
 from pprint import pformat
 
-from lerobot.utils.encoding_utils import decode_sign_magnitude, encode_sign_magnitude
+from lerobot.motors.encoding_utils import decode_sign_magnitude, encode_sign_magnitude
 
 from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
 from .tables import (

src/lerobot/motors/motors_bus.py

@@ -32,7 +32,7 @@ import serial
 from deepdiff import DeepDiff
 from tqdm import tqdm
 
-from lerobot.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 from lerobot.utils.utils import enter_pressed, move_cursor_up
 
 NameOrID: TypeAlias = str | int
@@ -99,12 +99,6 @@ class Motor:
     norm_mode: MotorNormMode
 
 
-class JointOutOfRangeError(Exception):
-    def __init__(self, message="Joint is out of range"):
-        self.message = message
-        super().__init__(self.message)
-
-
 class PortHandler(Protocol):
     def __init__(self, port_name):
         self.is_open: bool
@@ -348,7 +342,7 @@ class MotorsBus(abc.ABC):
             raise TypeError(motors)
 
     def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> list[str]:
-        if isinstance(values, (int, float)):
+        if isinstance(values, (int | float)):
             return dict.fromkeys(self.ids, values)
         elif isinstance(values, dict):
             return {self.motors[motor].id: val for motor, val in values.items()}
@@ -675,7 +669,7 @@ class MotorsBus(abc.ABC):
         """
         if motors is None:
             motors = list(self.motors)
-        elif isinstance(motors, (str, int)):
+        elif isinstance(motors, (str | int)):
             motors = [motors]
         elif not isinstance(motors, list):
             raise TypeError(motors)
@@ -703,7 +697,7 @@ class MotorsBus(abc.ABC):
         """
         if motors is None:
             motors = list(self.motors)
-        elif isinstance(motors, (str, int)):
+        elif isinstance(motors, (str | int)):
             motors = [motors]
         elif not isinstance(motors, list):
             raise TypeError(motors)
@@ -739,7 +733,7 @@ class MotorsBus(abc.ABC):
         """
         if motors is None:
             motors = list(self.motors)
-        elif isinstance(motors, (str, int)):
+        elif isinstance(motors, (str | int)):
             motors = [motors]
         elif not isinstance(motors, list):
             raise TypeError(motors)

src/lerobot/optim/optimizers.py

@@ -22,11 +22,11 @@ import draccus
 import torch
 from safetensors.torch import load_file, save_file
 
-from lerobot.constants import (
+from lerobot.datasets.utils import flatten_dict, unflatten_dict, write_json
+from lerobot.utils.constants import (
     OPTIMIZER_PARAM_GROUPS,
     OPTIMIZER_STATE,
 )
-from lerobot.datasets.utils import flatten_dict, unflatten_dict, write_json
 from lerobot.utils.io_utils import deserialize_json_into_object
 
 

src/lerobot/optim/schedulers.py

@@ -22,8 +22,8 @@ import draccus
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LambdaLR, LRScheduler
 
-from lerobot.constants import SCHEDULER_STATE
 from lerobot.datasets.utils import write_json
+from lerobot.utils.constants import SCHEDULER_STATE
 from lerobot.utils.io_utils import deserialize_json_into_object
 
 

src/lerobot/policies/__init__.py

@@ -15,6 +15,17 @@
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
+from .pi0.processor_pi0 import Pi0NewLineProcessor
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
+from .smolvla.processor_smolvla import SmolVLANewLineProcessor
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
+
+__all__ = [
+    "ACTConfig",
+    "DiffusionConfig",
+    "PI0Config",
+    "SmolVLAConfig",
+    "TDMPCConfig",
+    "VQBeTConfig",
+]

src/lerobot/policies/act/modeling_act.py

@@ -33,10 +33,9 @@ from torch import Tensor, nn
 from torchvision.models._utils import IntermediateLayerGetter
 from torchvision.ops.misc import FrozenBatchNorm2d
 
-from lerobot.constants import ACTION, OBS_IMAGES
 from lerobot.policies.act.configuration_act import ACTConfig
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGES, OBS_STATE
 
 
 class ACTPolicy(PreTrainedPolicy):
@@ -51,27 +50,16 @@ class ACTPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: ACTConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
         Args:
             config: Policy configuration class instance or None, in which case the default instantiation of
                     the configuration class is used.
-            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
-                that they will be passed with a call to `load_state_dict` before the policy is used.
         """
         super().__init__(config)
         config.validate_features()
         self.config = config
 
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         self.model = ACT(config)
 
         if config.temporal_ensemble_coeff is not None:
@@ -137,23 +125,19 @@ class ACTPolicy(PreTrainedPolicy):
         """Predict a chunk of actions given environment observations."""
         self.eval()
 
-        batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGES] = [batch[key] for key in self.config.image_features]
 
         actions = self.model(batch)[0]
-        actions = self.unnormalize_outputs({ACTION: actions})[ACTION]
         return actions
 
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
         """Run the batch through the model and compute the loss for training or validation."""
-        batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGES] = [batch[key] for key in self.config.image_features]
 
-        batch = self.normalize_targets(batch)
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
 
         l1_loss = (
@@ -410,25 +394,22 @@ class ACT(nn.Module):
             latent dimension.
         """
         if self.config.use_vae and self.training:
-            assert "action" in batch, (
+            assert ACTION in batch, (
                 "actions must be provided when using the variational objective in training mode."
             )
 
-        if "observation.images" in batch:
-            batch_size = batch["observation.images"][0].shape[0]
-        else:
-            batch_size = batch["observation.environment_state"].shape[0]
+        batch_size = batch[OBS_IMAGES][0].shape[0] if OBS_IMAGES in batch else batch[OBS_ENV_STATE].shape[0]
 
         # Prepare the latent for input to the transformer encoder.
-        if self.config.use_vae and "action" in batch and self.training:
+        if self.config.use_vae and ACTION in batch and self.training:
             # Prepare the input to the VAE encoder: [cls, *joint_space_configuration, *action_sequence].
             cls_embed = einops.repeat(
                 self.vae_encoder_cls_embed.weight, "1 d -> b 1 d", b=batch_size
             )  # (B, 1, D)
             if self.config.robot_state_feature:
-                robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"])
+                robot_state_embed = self.vae_encoder_robot_state_input_proj(batch[OBS_STATE])
                 robot_state_embed = robot_state_embed.unsqueeze(1)  # (B, 1, D)
-            action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)
+            action_embed = self.vae_encoder_action_input_proj(batch[ACTION])  # (B, S, D)
 
             if self.config.robot_state_feature:
                 vae_encoder_input = [cls_embed, robot_state_embed, action_embed]  # (B, S+2, D)
@@ -446,7 +427,7 @@ class ACT(nn.Module):
             cls_joint_is_pad = torch.full(
                 (batch_size, 2 if self.config.robot_state_feature else 1),
                 False,
-                device=batch["observation.state"].device,
+                device=batch[OBS_STATE].device,
             )
             key_padding_mask = torch.cat(
                 [cls_joint_is_pad, batch["action_is_pad"]], axis=1
@@ -470,7 +451,7 @@ class ACT(nn.Module):
             mu = log_sigma_x2 = None
             # TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use buffer
             latent_sample = torch.zeros([batch_size, self.config.latent_dim], dtype=torch.float32).to(
-                batch["observation.state"].device
+                batch[OBS_STATE].device
             )
 
         # Prepare transformer encoder inputs.
@@ -478,18 +459,16 @@ class ACT(nn.Module):
         encoder_in_pos_embed = list(self.encoder_1d_feature_pos_embed.weight.unsqueeze(1))
         # Robot state token.
         if self.config.robot_state_feature:
-            encoder_in_tokens.append(self.encoder_robot_state_input_proj(batch["observation.state"]))
+            encoder_in_tokens.append(self.encoder_robot_state_input_proj(batch[OBS_STATE]))
         # Environment state token.
         if self.config.env_state_feature:
-            encoder_in_tokens.append(
-                self.encoder_env_state_input_proj(batch["observation.environment_state"])
-            )
+            encoder_in_tokens.append(self.encoder_env_state_input_proj(batch[OBS_ENV_STATE]))
 
         if self.config.image_features:
             # For a list of images, the H and W may vary but H*W is constant.
             # NOTE: If modifying this section, verify on MPS devices that
             # gradients remain stable (no explosions or NaNs).
-            for img in batch["observation.images"]:
+            for img in batch[OBS_IMAGES]:
                 cam_features = self.backbone(img)["feature_map"]
                 cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features).to(dtype=cam_features.dtype)
                 cam_features = self.encoder_img_feat_input_proj(cam_features)

src/lerobot/policies/act/processor_act.py

@@ -0,0 +1,85 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Tony Z. Zhao and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any
+
+import torch
+
+from lerobot.policies.act.configuration_act import ACTConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    RenameObservationsProcessorStep,
+    UnnormalizerProcessorStep,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+
+def make_act_pre_post_processors(
+    config: ACTConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """Creates the pre- and post-processing pipelines for the ACT policy.
+
+    The pre-processing pipeline handles normalization, batching, and device placement for the model inputs.
+    The post-processing pipeline handles unnormalization and moves the model outputs back to the CPU.
+
+    Args:
+        config (ACTConfig): The ACT policy configuration object.
+        dataset_stats (dict[str, dict[str, torch.Tensor]] | None): A dictionary containing dataset
+            statistics (e.g., mean and std) used for normalization. Defaults to None.
+
+    Returns:
+        tuple[PolicyProcessorPipeline[dict[str, Any], dict[str, Any]], PolicyProcessorPipeline[PolicyAction, PolicyAction]]: A tuple containing the
+        pre-processor pipeline and the post-processor pipeline.
+    """
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+            device=config.device,
+        ),
+    ]
+    output_steps = [
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -33,9 +33,7 @@ from diffusers.schedulers.scheduling_ddim import DDIMScheduler
 from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
 from torch import Tensor, nn
 
-from lerobot.constants import ACTION, OBS_ENV_STATE, OBS_IMAGES, OBS_STATE
 from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.utils import (
     get_device_from_parameters,
@@ -43,6 +41,7 @@ from lerobot.policies.utils import (
     get_output_shape,
     populate_queues,
 )
+from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGES, OBS_STATE
 
 
 class DiffusionPolicy(PreTrainedPolicy):
@@ -57,7 +56,6 @@ class DiffusionPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: DiffusionConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
         Args:
@@ -70,14 +68,6 @@ class DiffusionPolicy(PreTrainedPolicy):
         config.validate_features()
         self.config = config
 
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         # queues are populated during rollout of the policy, they contain the n latest observations and actions
         self._queues = None
 
@@ -91,13 +81,13 @@ class DiffusionPolicy(PreTrainedPolicy):
     def reset(self):
         """Clear observation and action queues. Should be called on `env.reset()`"""
         self._queues = {
-            "observation.state": deque(maxlen=self.config.n_obs_steps),
-            "action": deque(maxlen=self.config.n_action_steps),
+            OBS_STATE: deque(maxlen=self.config.n_obs_steps),
+            ACTION: deque(maxlen=self.config.n_action_steps),
         }
         if self.config.image_features:
-            self._queues["observation.images"] = deque(maxlen=self.config.n_obs_steps)
+            self._queues[OBS_IMAGES] = deque(maxlen=self.config.n_obs_steps)
         if self.config.env_state_feature:
-            self._queues["observation.environment_state"] = deque(maxlen=self.config.n_obs_steps)
+            self._queues[OBS_ENV_STATE] = deque(maxlen=self.config.n_obs_steps)
 
     @torch.no_grad()
     def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
@@ -106,9 +96,6 @@ class DiffusionPolicy(PreTrainedPolicy):
         batch = {k: torch.stack(list(self._queues[k]), dim=1) for k in batch if k in self._queues}
         actions = self.diffusion.generate_actions(batch)
 
-        # TODO(rcadene): make above methods return output dictionary?
-        actions = self.unnormalize_outputs({ACTION: actions})[ACTION]
-
         return actions
 
     @torch.no_grad()
@@ -137,7 +124,6 @@ class DiffusionPolicy(PreTrainedPolicy):
         if ACTION in batch:
             batch.pop(ACTION)
 
-        batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
@@ -153,11 +139,9 @@ class DiffusionPolicy(PreTrainedPolicy):
 
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, None]:
         """Run the batch through the model and compute the loss for training or validation."""
-        batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
-        batch = self.normalize_targets(batch)
         loss = self.diffusion.compute_loss(batch)
         # no output_dict so returning None
         return loss, None
@@ -250,7 +234,7 @@ class DiffusionModel(nn.Module):
         if self.config.image_features:
             if self.config.use_separate_rgb_encoder_per_camera:
                 # Combine batch and sequence dims while rearranging to make the camera index dimension first.
-                images_per_camera = einops.rearrange(batch["observation.images"], "b s n ... -> n (b s) ...")
+                images_per_camera = einops.rearrange(batch[OBS_IMAGES], "b s n ... -> n (b s) ...")
                 img_features_list = torch.cat(
                     [
                         encoder(images)
@@ -265,7 +249,7 @@ class DiffusionModel(nn.Module):
             else:
                 # Combine batch, sequence, and "which camera" dims before passing to shared encoder.
                 img_features = self.rgb_encoder(
-                    einops.rearrange(batch["observation.images"], "b s n ... -> (b s n) ...")
+                    einops.rearrange(batch[OBS_IMAGES], "b s n ... -> (b s n) ...")
                 )
                 # Separate batch dim and sequence dim back out. The camera index dim gets absorbed into the
                 # feature dim (effectively concatenating the camera features).
@@ -291,7 +275,7 @@ class DiffusionModel(nn.Module):
             "observation.environment_state": (B, n_obs_steps, environment_dim)
         }
         """
-        batch_size, n_obs_steps = batch["observation.state"].shape[:2]
+        batch_size, n_obs_steps = batch[OBS_STATE].shape[:2]
         assert n_obs_steps == self.config.n_obs_steps
 
         # Encode image features and concatenate them all together along with the state vector.
@@ -322,10 +306,10 @@ class DiffusionModel(nn.Module):
         }
         """
         # Input validation.
-        assert set(batch).issuperset({"observation.state", "action", "action_is_pad"})
-        assert "observation.images" in batch or "observation.environment_state" in batch
-        n_obs_steps = batch["observation.state"].shape[1]
-        horizon = batch["action"].shape[1]
+        assert set(batch).issuperset({OBS_STATE, ACTION, "action_is_pad"})
+        assert OBS_IMAGES in batch or OBS_ENV_STATE in batch
+        n_obs_steps = batch[OBS_STATE].shape[1]
+        horizon = batch[ACTION].shape[1]
         assert horizon == self.config.horizon
         assert n_obs_steps == self.config.n_obs_steps
 
@@ -333,7 +317,7 @@ class DiffusionModel(nn.Module):
         global_cond = self._prepare_global_conditioning(batch)  # (B, global_cond_dim)
 
         # Forward diffusion.
-        trajectory = batch["action"]
+        trajectory = batch[ACTION]
         # Sample noise to add to the trajectory.
         eps = torch.randn(trajectory.shape, device=trajectory.device)
         # Sample a random noising timestep for each item in the batch.
@@ -354,7 +338,7 @@ class DiffusionModel(nn.Module):
         if self.config.prediction_type == "epsilon":
             target = eps
         elif self.config.prediction_type == "sample":
-            target = batch["action"]
+            target = batch[ACTION]
         else:
             raise ValueError(f"Unsupported prediction type {self.config.prediction_type}")
 

src/lerobot/policies/diffusion/processor_diffusion.py

@@ -0,0 +1,92 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Columbia Artificial Intelligence, Robotics Lab,
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any
+
+import torch
+
+from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    RenameObservationsProcessorStep,
+    UnnormalizerProcessorStep,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+
+def make_diffusion_pre_post_processors(
+    config: DiffusionConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Constructs pre-processor and post-processor pipelines for a diffusion policy.
+
+    The pre-processing pipeline prepares the input data for the model by:
+    1. Renaming features.
+    2. Normalizing the input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Moving the data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving the data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the diffusion policy,
+            containing feature definitions, normalization mappings, and device information.
+        dataset_stats: A dictionary of statistics used for normalization.
+            Defaults to None.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+    output_steps = [
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )

src/lerobot/policies/factory.py

@@ -14,9 +14,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
+from __future__ import annotations
 
-from torch import nn
+import logging
+from typing import Any, TypedDict
+
+import torch
+from typing_extensions import Unpack
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType
@@ -34,10 +38,33 @@ from lerobot.policies.sac.reward_model.configuration_classifier import RewardCla
 from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig
 from lerobot.policies.tdmpc.configuration_tdmpc import TDMPCConfig
 from lerobot.policies.vqbet.configuration_vqbet import VQBeTConfig
+from lerobot.processor import PolicyAction, PolicyProcessorPipeline
+from lerobot.processor.converters import (
+    batch_to_transition,
+    policy_action_to_transition,
+    transition_to_batch,
+    transition_to_policy_action,
+)
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 
-def get_policy_class(name: str) -> PreTrainedPolicy:
-    """Get the policy's class and config class given a name (matching the policy class' `name` attribute)."""
+def get_policy_class(name: str) -> type[PreTrainedPolicy]:
+    """
+    Retrieves a policy class by its registered name.
+
+    This function uses dynamic imports to avoid loading all policy classes into memory
+    at once, improving startup time and reducing dependencies.
+
+    Args:
+        name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
+              "vqbet", "pi0", "pi0fast", "sac", "reward_classifier", "smolvla".
+
+    Returns:
+        The policy class corresponding to the given name.
+
+    Raises:
+        NotImplementedError: If the policy name is not recognized.
+    """
     if name == "tdmpc":
         from lerobot.policies.tdmpc.modeling_tdmpc import TDMPCPolicy
 
@@ -79,6 +106,24 @@ def get_policy_class(name: str) -> PreTrainedPolicy:
 
 
 def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
+    """
+    Instantiates a policy configuration object based on the policy type.
+
+    This factory function simplifies the creation of policy configuration objects by
+    mapping a string identifier to the corresponding config class.
+
+    Args:
+        policy_type: The type of the policy. Supported types include "tdmpc",
+                     "diffusion", "act", "vqbet", "pi0", "pi0fast", "sac", "smolvla",
+                     "reward_classifier".
+        **kwargs: Keyword arguments to be passed to the configuration class constructor.
+
+    Returns:
+        An instance of a `PreTrainedConfig` subclass.
+
+    Raises:
+        ValueError: If the `policy_type` is not recognized.
+    """
     if policy_type == "tdmpc":
         return TDMPCConfig(**kwargs)
     elif policy_type == "diffusion":
@@ -101,30 +146,187 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         raise ValueError(f"Policy type '{policy_type}' is not available.")
 
 
+class ProcessorConfigKwargs(TypedDict, total=False):
+    """
+    A TypedDict defining the keyword arguments for processor configuration.
+
+    This provides type hints for the optional arguments passed to `make_pre_post_processors`,
+    improving code clarity and enabling static analysis.
+
+    Attributes:
+        preprocessor_config_filename: The filename for the preprocessor configuration.
+        postprocessor_config_filename: The filename for the postprocessor configuration.
+        preprocessor_overrides: A dictionary of overrides for the preprocessor configuration.
+        postprocessor_overrides: A dictionary of overrides for the postprocessor configuration.
+        dataset_stats: Dataset statistics for normalization.
+    """
+
+    preprocessor_config_filename: str | None
+    postprocessor_config_filename: str | None
+    preprocessor_overrides: dict[str, Any] | None
+    postprocessor_overrides: dict[str, Any] | None
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None
+
+
+def make_pre_post_processors(
+    policy_cfg: PreTrainedConfig,
+    pretrained_path: str | None = None,
+    **kwargs: Unpack[ProcessorConfigKwargs],
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Create or load pre- and post-processor pipelines for a given policy.
+
+    This function acts as a factory. It can either load existing processor pipelines
+    from a pretrained path or create new ones from scratch based on the policy
+    configuration. Each policy type has a dedicated factory function for its
+    processors (e.g., `make_tdmpc_pre_post_processors`).
+
+    Args:
+        policy_cfg: The configuration of the policy for which to create processors.
+        pretrained_path: An optional path to load pretrained processor pipelines from.
+            If provided, pipelines are loaded from this path.
+        **kwargs: Keyword arguments for processor configuration, as defined in
+            `ProcessorConfigKwargs`.
+
+    Returns:
+        A tuple containing the input (pre-processor) and output (post-processor) pipelines.
+
+    Raises:
+        NotImplementedError: If a processor factory is not implemented for the given
+            policy configuration type.
+    """
+    if pretrained_path:
+        return (
+            PolicyProcessorPipeline.from_pretrained(
+                pretrained_model_name_or_path=pretrained_path,
+                config_filename=kwargs.get(
+                    "preprocessor_config_filename", f"{POLICY_PREPROCESSOR_DEFAULT_NAME}.json"
+                ),
+                overrides=kwargs.get("preprocessor_overrides", {}),
+                to_transition=batch_to_transition,
+                to_output=transition_to_batch,
+            ),
+            PolicyProcessorPipeline.from_pretrained(
+                pretrained_model_name_or_path=pretrained_path,
+                config_filename=kwargs.get(
+                    "postprocessor_config_filename", f"{POLICY_POSTPROCESSOR_DEFAULT_NAME}.json"
+                ),
+                overrides=kwargs.get("postprocessor_overrides", {}),
+                to_transition=policy_action_to_transition,
+                to_output=transition_to_policy_action,
+            ),
+        )
+
+    # Create a new processor based on policy type
+    if isinstance(policy_cfg, TDMPCConfig):
+        from lerobot.policies.tdmpc.processor_tdmpc import make_tdmpc_pre_post_processors
+
+        processors = make_tdmpc_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
+    elif isinstance(policy_cfg, DiffusionConfig):
+        from lerobot.policies.diffusion.processor_diffusion import make_diffusion_pre_post_processors
+
+        processors = make_diffusion_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
+    elif isinstance(policy_cfg, ACTConfig):
+        from lerobot.policies.act.processor_act import make_act_pre_post_processors
+
+        processors = make_act_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
+    elif isinstance(policy_cfg, VQBeTConfig):
+        from lerobot.policies.vqbet.processor_vqbet import make_vqbet_pre_post_processors
+
+        processors = make_vqbet_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
+    elif isinstance(policy_cfg, PI0Config):
+        from lerobot.policies.pi0.processor_pi0 import make_pi0_pre_post_processors
+
+        processors = make_pi0_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
+    elif isinstance(policy_cfg, PI0FASTConfig):
+        from lerobot.policies.pi0fast.processor_pi0fast import make_pi0fast_pre_post_processors
+
+        processors = make_pi0fast_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
+    elif isinstance(policy_cfg, SACConfig):
+        from lerobot.policies.sac.processor_sac import make_sac_pre_post_processors
+
+        processors = make_sac_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
+    elif isinstance(policy_cfg, RewardClassifierConfig):
+        from lerobot.policies.sac.reward_model.processor_classifier import make_classifier_processor
+
+        processors = make_classifier_processor(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
+    elif isinstance(policy_cfg, SmolVLAConfig):
+        from lerobot.policies.smolvla.processor_smolvla import make_smolvla_pre_post_processors
+
+        processors = make_smolvla_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
+    else:
+        raise NotImplementedError(f"Processor for policy type '{policy_cfg.type}' is not implemented.")
+
+    return processors
+
+
 def make_policy(
     cfg: PreTrainedConfig,
     ds_meta: LeRobotDatasetMetadata | None = None,
     env_cfg: EnvConfig | None = None,
 ) -> PreTrainedPolicy:
-    """Make an instance of a policy class.
+    """
+    Instantiate a policy model.
 
-    This function exists because (for now) we need to parse features from either a dataset or an environment
-    in order to properly dimension and instantiate a policy for that dataset or environment.
+    This factory function handles the logic of creating a policy, which requires
+    determining the input and output feature shapes. These shapes can be derived
+    either from a `LeRobotDatasetMetadata` object or an `EnvConfig` object. The function
+    can either initialize a new policy from scratch or load a pretrained one.
 
     Args:
-        cfg (PreTrainedConfig): The config of the policy to make. If `pretrained_path` is set, the policy will
-            be loaded with the weights from that path.
-        ds_meta (LeRobotDatasetMetadata | None, optional): Dataset metadata to take input/output shapes and
-            statistics to use for (un)normalization of inputs/outputs in the policy. Defaults to None.
-        env_cfg (EnvConfig | None, optional): The config of a gym environment to parse features from. Must be
-            provided if ds_meta is not. Defaults to None.
-
-    Raises:
-        ValueError: Either ds_meta or env and env_cfg must be provided.
-        NotImplementedError: if the policy.type is 'vqbet' and the policy device 'mps' (due to an incompatibility)
+        cfg: The configuration for the policy to be created. If `cfg.pretrained_path` is
+             set, the policy will be loaded with weights from that path.
+        ds_meta: Dataset metadata used to infer feature shapes and types. Also provides
+                 statistics for normalization layers.
+        env_cfg: Environment configuration used to infer feature shapes and types.
+                 One of `ds_meta` or `env_cfg` must be provided.
 
     Returns:
-        PreTrainedPolicy: _description_
+        An instantiated and device-placed policy model.
+
+    Raises:
+        ValueError: If both or neither of `ds_meta` and `env_cfg` are provided.
+        NotImplementedError: If attempting to use an unsupported policy-backend
+                             combination (e.g., VQBeT with 'mps').
     """
     if bool(ds_meta) == bool(env_cfg):
         raise ValueError("Either one of a dataset metadata or a sim env must be provided.")
@@ -147,7 +349,6 @@ def make_policy(
     kwargs = {}
     if ds_meta is not None:
         features = dataset_to_policy_features(ds_meta.features)
-        kwargs["dataset_stats"] = ds_meta.stats
     else:
         if not cfg.pretrained_path:
             logging.warning(
@@ -155,6 +356,8 @@ def make_policy(
                 "rather than a dataset. Normalization modules inside the policy will have infinite values "
                 "by default without stats from a dataset."
             )
+        if env_cfg is None:
+            raise ValueError("env_cfg cannot be None when ds_meta is not provided")
         features = env_to_policy_features(env_cfg)
 
     cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
@@ -171,7 +374,7 @@ def make_policy(
         policy = policy_cls(**kwargs)
 
     policy.to(cfg.device)
-    assert isinstance(policy, nn.Module)
+    assert isinstance(policy, torch.nn.Module)
 
     # policy = torch.compile(policy, mode="reduce-overhead")
 

src/lerobot/policies/normalize.py

@@ -1,420 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-import numpy as np
-import torch
-from torch import Tensor, nn
-
-from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
-
-
-def create_stats_buffers(
-    features: dict[str, PolicyFeature],
-    norm_map: dict[str, NormalizationMode],
-    stats: dict[str, dict[str, Tensor]] | None = None,
-) -> dict[str, dict[str, nn.ParameterDict]]:
-    """
-    Create buffers per modality (e.g. "observation.image", "action") containing their mean, std, min, max
-    statistics.
-
-    Args: (see Normalize and Unnormalize)
-
-    Returns:
-        dict: A dictionary where keys are modalities and values are `nn.ParameterDict` containing
-            `nn.Parameters` set to `requires_grad=False`, suitable to not be updated during backpropagation.
-    """
-    stats_buffers = {}
-
-    for key, ft in features.items():
-        norm_mode = norm_map.get(ft.type, NormalizationMode.IDENTITY)
-        if norm_mode is NormalizationMode.IDENTITY:
-            continue
-
-        assert isinstance(norm_mode, NormalizationMode)
-
-        shape = tuple(ft.shape)
-
-        if ft.type is FeatureType.VISUAL:
-            # sanity checks
-            assert len(shape) == 3, f"number of dimensions of {key} != 3 ({shape=}"
-            c, h, w = shape
-            assert c < h and c < w, f"{key} is not channel first ({shape=})"
-            # override image shape to be invariant to height and width
-            shape = (c, 1, 1)
-
-        # Note: we initialize mean, std, min, max to infinity. They should be overwritten
-        # downstream by `stats` or `policy.load_state_dict`, as expected. During forward,
-        # we assert they are not infinity anymore.
-
-        buffer = {}
-        if norm_mode is NormalizationMode.MEAN_STD:
-            mean = torch.ones(shape, dtype=torch.float32) * torch.inf
-            std = torch.ones(shape, dtype=torch.float32) * torch.inf
-            buffer = nn.ParameterDict(
-                {
-                    "mean": nn.Parameter(mean, requires_grad=False),
-                    "std": nn.Parameter(std, requires_grad=False),
-                }
-            )
-        elif norm_mode is NormalizationMode.MIN_MAX:
-            min = torch.ones(shape, dtype=torch.float32) * torch.inf
-            max = torch.ones(shape, dtype=torch.float32) * torch.inf
-            buffer = nn.ParameterDict(
-                {
-                    "min": nn.Parameter(min, requires_grad=False),
-                    "max": nn.Parameter(max, requires_grad=False),
-                }
-            )
-
-        # TODO(aliberts, rcadene): harmonize this to only use one framework (np or torch)
-        if stats:
-            if isinstance(stats[key]["mean"], np.ndarray):
-                if norm_mode is NormalizationMode.MEAN_STD:
-                    buffer["mean"].data = torch.from_numpy(stats[key]["mean"]).to(dtype=torch.float32)
-                    buffer["std"].data = torch.from_numpy(stats[key]["std"]).to(dtype=torch.float32)
-                elif norm_mode is NormalizationMode.MIN_MAX:
-                    buffer["min"].data = torch.from_numpy(stats[key]["min"]).to(dtype=torch.float32)
-                    buffer["max"].data = torch.from_numpy(stats[key]["max"]).to(dtype=torch.float32)
-            elif isinstance(stats[key]["mean"], torch.Tensor):
-                # Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
-                # tensors anywhere (for example, when we use the same stats for normalization and
-                # unnormalization). See the logic here
-                # https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
-                if norm_mode is NormalizationMode.MEAN_STD:
-                    buffer["mean"].data = stats[key]["mean"].clone().to(dtype=torch.float32)
-                    buffer["std"].data = stats[key]["std"].clone().to(dtype=torch.float32)
-                elif norm_mode is NormalizationMode.MIN_MAX:
-                    buffer["min"].data = stats[key]["min"].clone().to(dtype=torch.float32)
-                    buffer["max"].data = stats[key]["max"].clone().to(dtype=torch.float32)
-            else:
-                type_ = type(stats[key]["mean"])
-                raise ValueError(f"np.ndarray or torch.Tensor expected, but type is '{type_}' instead.")
-
-        stats_buffers[key] = buffer
-    return stats_buffers
-
-
-def _no_stats_error_str(name: str) -> str:
-    return (
-        f"`{name}` is infinity. You should either initialize with `stats` as an argument, or use a "
-        "pretrained model."
-    )
-
-
-class Normalize(nn.Module):
-    """Normalizes data (e.g. "observation.image") for more stable and faster convergence during training."""
-
-    def __init__(
-        self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
-        stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        """
-        Args:
-            shapes (dict): A dictionary where keys are input modalities (e.g. "observation.image") and values
-            are their shapes (e.g. `[3,96,96]`]). These shapes are used to create the tensor buffer containing
-            mean, std, min, max statistics. If the provided `shapes` contain keys related to images, the shape
-            is adjusted to be invariant to height and width, assuming a channel-first (c, h, w) format.
-            modes (dict): A dictionary where keys are output modalities (e.g. "observation.image") and values
-                are their normalization modes among:
-                    - "mean_std": subtract the mean and divide by standard deviation.
-                    - "min_max": map to [-1, 1] range.
-            stats (dict, optional): A dictionary where keys are output modalities (e.g. "observation.image")
-                and values are dictionaries of statistic types and their values (e.g.
-                `{"mean": torch.randn(3,1,1)}, "std": torch.randn(3,1,1)}`). If provided, as expected for
-                training the model for the first time, these statistics will overwrite the default buffers. If
-                not provided, as expected for finetuning or evaluation, the default buffers should to be
-                overwritten by a call to `policy.load_state_dict(state_dict)`. That way, initializing the
-                dataset is not needed to get the stats, since they are already in the policy state_dict.
-        """
-        super().__init__()
-        self.features = features
-        self.norm_map = norm_map
-        self.stats = stats
-        stats_buffers = create_stats_buffers(features, norm_map, stats)
-        for key, buffer in stats_buffers.items():
-            setattr(self, "buffer_" + key.replace(".", "_"), buffer)
-
-    # TODO(rcadene): should we remove torch.no_grad?
-    @torch.no_grad()
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        # TODO: Remove this shallow copy
-        batch = dict(batch)  # shallow copy avoids mutating the input batch
-        for key, ft in self.features.items():
-            if key not in batch:
-                # FIXME(aliberts, rcadene): This might lead to silent fail!
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
-            buffer = getattr(self, "buffer_" + key.replace(".", "_"))
-
-            if norm_mode is NormalizationMode.MEAN_STD:
-                mean = buffer["mean"]
-                std = buffer["std"]
-                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
-                assert not torch.isinf(std).any(), _no_stats_error_str("std")
-                batch[key] = (batch[key] - mean) / (std + 1e-8)
-            elif norm_mode is NormalizationMode.MIN_MAX:
-                min = buffer["min"]
-                max = buffer["max"]
-                assert not torch.isinf(min).any(), _no_stats_error_str("min")
-                assert not torch.isinf(max).any(), _no_stats_error_str("max")
-                # normalize to [0,1]
-                batch[key] = (batch[key] - min) / (max - min + 1e-8)
-                # normalize to [-1, 1]
-                batch[key] = batch[key] * 2 - 1
-            else:
-                raise ValueError(norm_mode)
-        return batch
-
-
-class Unnormalize(nn.Module):
-    """
-    Similar to `Normalize` but unnormalizes output data (e.g. `{"action": torch.randn(b,c)}`) in their
-    original range used by the environment.
-    """
-
-    def __init__(
-        self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
-        stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        """
-        Args:
-            shapes (dict): A dictionary where keys are input modalities (e.g. "observation.image") and values
-            are their shapes (e.g. `[3,96,96]`]). These shapes are used to create the tensor buffer containing
-            mean, std, min, max statistics. If the provided `shapes` contain keys related to images, the shape
-            is adjusted to be invariant to height and width, assuming a channel-first (c, h, w) format.
-            modes (dict): A dictionary where keys are output modalities (e.g. "observation.image") and values
-                are their normalization modes among:
-                    - "mean_std": subtract the mean and divide by standard deviation.
-                    - "min_max": map to [-1, 1] range.
-            stats (dict, optional): A dictionary where keys are output modalities (e.g. "observation.image")
-                and values are dictionaries of statistic types and their values (e.g.
-                `{"mean": torch.randn(3,1,1)}, "std": torch.randn(3,1,1)}`). If provided, as expected for
-                training the model for the first time, these statistics will overwrite the default buffers. If
-                not provided, as expected for finetuning or evaluation, the default buffers should to be
-                overwritten by a call to `policy.load_state_dict(state_dict)`. That way, initializing the
-                dataset is not needed to get the stats, since they are already in the policy state_dict.
-        """
-        super().__init__()
-        self.features = features
-        self.norm_map = norm_map
-        self.stats = stats
-        # `self.buffer_observation_state["mean"]` contains `torch.tensor(state_dim)`
-        stats_buffers = create_stats_buffers(features, norm_map, stats)
-        for key, buffer in stats_buffers.items():
-            setattr(self, "buffer_" + key.replace(".", "_"), buffer)
-
-    # TODO(rcadene): should we remove torch.no_grad?
-    @torch.no_grad()
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        batch = dict(batch)  # shallow copy avoids mutating the input batch
-        for key, ft in self.features.items():
-            if key not in batch:
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
-            buffer = getattr(self, "buffer_" + key.replace(".", "_"))
-
-            if norm_mode is NormalizationMode.MEAN_STD:
-                mean = buffer["mean"]
-                std = buffer["std"]
-                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
-                assert not torch.isinf(std).any(), _no_stats_error_str("std")
-                batch[key] = batch[key] * std + mean
-            elif norm_mode is NormalizationMode.MIN_MAX:
-                min = buffer["min"]
-                max = buffer["max"]
-                assert not torch.isinf(min).any(), _no_stats_error_str("min")
-                assert not torch.isinf(max).any(), _no_stats_error_str("max")
-                batch[key] = (batch[key] + 1) / 2
-                batch[key] = batch[key] * (max - min) + min
-            else:
-                raise ValueError(norm_mode)
-        return batch
-
-
-# TODO (azouitine): We should replace all normalization on the policies with register_buffer normalization
-#       and remove the `Normalize` and `Unnormalize` classes.
-def _initialize_stats_buffers(
-    module: nn.Module,
-    features: dict[str, PolicyFeature],
-    norm_map: dict[str, NormalizationMode],
-    stats: dict[str, dict[str, Tensor]] | None = None,
-) -> None:
-    """Register statistics buffers (mean/std or min/max) on the given *module*.
-
-    The logic matches the previous constructors of `NormalizeBuffer` and `UnnormalizeBuffer`,
-    but is factored out so it can be reused by both classes and stay in sync.
-    """
-    for key, ft in features.items():
-        norm_mode = norm_map.get(ft.type, NormalizationMode.IDENTITY)
-        if norm_mode is NormalizationMode.IDENTITY:
-            continue
-
-        shape: tuple[int, ...] = tuple(ft.shape)
-        if ft.type is FeatureType.VISUAL:
-            # reduce spatial dimensions, keep channel dimension only
-            c, *_ = shape
-            shape = (c, 1, 1)
-
-        prefix = key.replace(".", "_")
-
-        if norm_mode is NormalizationMode.MEAN_STD:
-            mean = torch.full(shape, torch.inf, dtype=torch.float32)
-            std = torch.full(shape, torch.inf, dtype=torch.float32)
-
-            if stats and key in stats and "mean" in stats[key] and "std" in stats[key]:
-                mean_data = stats[key]["mean"]
-                std_data = stats[key]["std"]
-                if isinstance(mean_data, torch.Tensor):
-                    # Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
-                    # tensors anywhere (for example, when we use the same stats for normalization and
-                    # unnormalization). See the logic here
-                    # https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
-                    mean = mean_data.clone().to(dtype=torch.float32)
-                    std = std_data.clone().to(dtype=torch.float32)
-                else:
-                    raise ValueError(f"Unsupported stats type for key '{key}' (expected ndarray or Tensor).")
-
-            module.register_buffer(f"{prefix}_mean", mean)
-            module.register_buffer(f"{prefix}_std", std)
-            continue
-
-        if norm_mode is NormalizationMode.MIN_MAX:
-            min_val = torch.full(shape, torch.inf, dtype=torch.float32)
-            max_val = torch.full(shape, torch.inf, dtype=torch.float32)
-
-            if stats and key in stats and "min" in stats[key] and "max" in stats[key]:
-                min_data = stats[key]["min"]
-                max_data = stats[key]["max"]
-                if isinstance(min_data, torch.Tensor):
-                    min_val = min_data.clone().to(dtype=torch.float32)
-                    max_val = max_data.clone().to(dtype=torch.float32)
-                else:
-                    raise ValueError(f"Unsupported stats type for key '{key}' (expected ndarray or Tensor).")
-
-            module.register_buffer(f"{prefix}_min", min_val)
-            module.register_buffer(f"{prefix}_max", max_val)
-            continue
-
-        raise ValueError(norm_mode)
-
-
-class NormalizeBuffer(nn.Module):
-    """Same as `Normalize` but statistics are stored as registered buffers rather than parameters."""
-
-    def __init__(
-        self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
-        stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        super().__init__()
-        self.features = features
-        self.norm_map = norm_map
-
-        _initialize_stats_buffers(self, features, norm_map, stats)
-
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        batch = dict(batch)
-        for key, ft in self.features.items():
-            if key not in batch:
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
-            prefix = key.replace(".", "_")
-
-            if norm_mode is NormalizationMode.MEAN_STD:
-                mean = getattr(self, f"{prefix}_mean")
-                std = getattr(self, f"{prefix}_std")
-                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
-                assert not torch.isinf(std).any(), _no_stats_error_str("std")
-                batch[key] = (batch[key] - mean) / (std + 1e-8)
-                continue
-
-            if norm_mode is NormalizationMode.MIN_MAX:
-                min_val = getattr(self, f"{prefix}_min")
-                max_val = getattr(self, f"{prefix}_max")
-                assert not torch.isinf(min_val).any(), _no_stats_error_str("min")
-                assert not torch.isinf(max_val).any(), _no_stats_error_str("max")
-                batch[key] = (batch[key] - min_val) / (max_val - min_val + 1e-8)
-                batch[key] = batch[key] * 2 - 1
-                continue
-
-            raise ValueError(norm_mode)
-
-        return batch
-
-
-class UnnormalizeBuffer(nn.Module):
-    """Inverse operation of `NormalizeBuffer`. Uses registered buffers for statistics."""
-
-    def __init__(
-        self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
-        stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        super().__init__()
-        self.features = features
-        self.norm_map = norm_map
-
-        _initialize_stats_buffers(self, features, norm_map, stats)
-
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        # batch = dict(batch)
-        for key, ft in self.features.items():
-            if key not in batch:
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
-            prefix = key.replace(".", "_")
-
-            if norm_mode is NormalizationMode.MEAN_STD:
-                mean = getattr(self, f"{prefix}_mean")
-                std = getattr(self, f"{prefix}_std")
-                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
-                assert not torch.isinf(std).any(), _no_stats_error_str("std")
-                batch[key] = batch[key] * std + mean
-                continue
-
-            if norm_mode is NormalizationMode.MIN_MAX:
-                min_val = getattr(self, f"{prefix}_min")
-                max_val = getattr(self, f"{prefix}_max")
-                assert not torch.isinf(min_val).any(), _no_stats_error_str("min")
-                assert not torch.isinf(max_val).any(), _no_stats_error_str("max")
-                batch[key] = (batch[key] + 1) / 2
-                batch[key] = batch[key] * (max_val - min_val) + min_val
-                continue
-
-            raise ValueError(norm_mode)
-
-        return batch

src/lerobot/policies/pi0/configuration_pi0.py

@@ -20,6 +20,7 @@ from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import (
     CosineDecayWithWarmupSchedulerConfig,
 )
+from lerobot.utils.constants import OBS_IMAGES
 
 
 @PreTrainedConfig.register_subclass("pi0")
@@ -113,7 +114,7 @@ class PI0Config(PreTrainedConfig):
         #     raise ValueError("You must provide at least one image or the environment state among the inputs.")
 
         for i in range(self.empty_cameras):
-            key = f"observation.images.empty_camera_{i}"
+            key = f"{OBS_IMAGES}.empty_camera_{i}"
             empty_camera = PolicyFeature(
                 type=FeatureType.VISUAL,
                 shape=(3, 480, 640),

src/lerobot/policies/pi0/conversion_scripts/compare_with_jax.py

@@ -21,6 +21,7 @@ import torch
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
 from lerobot.policies.factory import make_policy
+from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
 
 
 def display(tensor: torch.Tensor):
@@ -60,26 +61,26 @@ def main():
 
     # Override stats
     dataset_meta = LeRobotDatasetMetadata(dataset_repo_id)
-    dataset_meta.stats["observation.state"]["mean"] = torch.tensor(
+    dataset_meta.stats[OBS_STATE]["mean"] = torch.tensor(
         norm_stats["norm_stats"]["state"]["mean"][:num_motors], dtype=torch.float32
     )
-    dataset_meta.stats["observation.state"]["std"] = torch.tensor(
+    dataset_meta.stats[OBS_STATE]["std"] = torch.tensor(
         norm_stats["norm_stats"]["state"]["std"][:num_motors], dtype=torch.float32
     )
 
     # Create LeRobot batch from Jax
     batch = {}
     for cam_key, uint_chw_array in example["images"].items():
-        batch[f"observation.images.{cam_key}"] = torch.from_numpy(uint_chw_array) / 255.0
-    batch["observation.state"] = torch.from_numpy(example["state"])
-    batch["action"] = torch.from_numpy(outputs["actions"])
+        batch[f"{OBS_IMAGES}.{cam_key}"] = torch.from_numpy(uint_chw_array) / 255.0
+    batch[OBS_STATE] = torch.from_numpy(example["state"])
+    batch[ACTION] = torch.from_numpy(outputs["actions"])
     batch["task"] = example["prompt"]
 
     if model_name == "pi0_aloha_towel":
-        del batch["observation.images.cam_low"]
+        del batch[f"{OBS_IMAGES}.cam_low"]
     elif model_name == "pi0_aloha_sim":
-        batch["observation.images.top"] = batch["observation.images.cam_high"]
-        del batch["observation.images.cam_high"]
+        batch[f"{OBS_IMAGES}.top"] = batch[f"{OBS_IMAGES}.cam_high"]
+        del batch[f"{OBS_IMAGES}.cam_high"]
 
     # Batchify
     for key in batch:
@@ -116,7 +117,7 @@ def main():
         actions.append(action)
 
     actions = torch.stack(actions, dim=1)
-    pi_actions = batch["action"]
+    pi_actions = batch[ACTION]
     print("actions")
     display(actions)
     print()

src/lerobot/policies/pi0/modeling_pi0.py

@@ -56,18 +56,15 @@ from collections import deque
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
-from transformers import AutoTokenizer
 
-from lerobot.constants import ACTION, OBS_STATE
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.policies.pi0.paligemma_with_expert import (
     PaliGemmaWithExpertConfig,
     PaliGemmaWithExpertModel,
 )
 from lerobot.policies.pretrained import PreTrainedPolicy
-from lerobot.policies.utils import log_model_loading_keys
-from lerobot.utils.utils import get_safe_dtype, init_logging
+from lerobot.utils.constants import ACTION, OBS_LANGUAGE_ATTENTION_MASK, OBS_LANGUAGE_TOKENS, OBS_STATE
+from lerobot.utils.utils import get_safe_dtype
 
 
 def create_sinusoidal_pos_embedding(
@@ -223,28 +220,17 @@ class PI0Policy(PreTrainedPolicy):
     def __init__(
         self,
         config: PI0Config,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
         Args:
             config: Policy configuration class instance or None, in which case the default instantiation of
                     the configuration class is used.
-            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
-                that they will be passed with a call to `load_state_dict` before the policy is used.
         """
 
         super().__init__(config)
         config.validate_features()
         self.config = config
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
 
-        self.language_tokenizer = AutoTokenizer.from_pretrained("google/paligemma-3b-pt-224")
         self.model = PI0FlowMatching(config)
 
         self.reset()
@@ -253,99 +239,6 @@ class PI0Policy(PreTrainedPolicy):
         """This should be called whenever the environment is reset."""
         self._action_queue = deque([], maxlen=self.config.n_action_steps)
 
-    @classmethod
-    def _transform_state_dict_keys(cls, state_dict: dict) -> dict:
-        """
-        Transform state dict keys to match expected model structure.
-
-        Transformations:
-        - model.paligemma_with_expert.paligemma.language_model.lm_head ->
-          model.paligemma_with_expert.paligemma.lm_head
-        - model.paligemma_with_expert.paligemma.language_model.model ->
-          model.paligemma_with_expert.paligemma.model.language_model
-        - model.paligemma_with_expert.paligemma.vision_tower ->
-          model.paligemma_with_expert.paligemma.model.vision_tower
-        - model.paligemma_with_expert.paligemma.multi_modal_projector ->
-          model.paligemma_with_expert.paligemma.model.multi_modal_projector
-
-        Also handles tied weights between lm_head.weight and
-        embed_tokens.weight.
-        """
-        import re
-
-        transformed_dict = {}
-
-        transformations = [
-            (
-                re.compile(r"\.paligemma_with_expert\.paligemma\.language_model\.lm_head"),
-                ".paligemma_with_expert.paligemma.lm_head",
-            ),
-            (
-                re.compile(r"\.paligemma_with_expert\.paligemma\.language_model\.model"),
-                ".paligemma_with_expert.paligemma.model.language_model",
-            ),
-            (
-                re.compile(r"\.paligemma_with_expert\.paligemma\.vision_tower"),
-                ".paligemma_with_expert.paligemma.model.vision_tower",
-            ),
-            (
-                re.compile(r"\.paligemma_with_expert\.paligemma\.multi_modal_projector"),
-                ".paligemma_with_expert.paligemma.model.multi_modal_projector",
-            ),
-        ]
-
-        for key, value in state_dict.items():
-            new_key = key
-            for pattern, replacement in transformations:
-                new_key = pattern.sub(replacement, new_key)
-            transformed_dict[new_key] = value
-
-        # Handle tied weights: lm_head.weight and embed_tokens.weight share memory
-        lm_head_key = None
-        embed_tokens_key = None
-
-        for key in transformed_dict:
-            if key.endswith(".paligemma_with_expert.paligemma.lm_head.weight"):
-                lm_head_key = key
-            elif key.endswith(".paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"):
-                embed_tokens_key = key
-            if lm_head_key and embed_tokens_key:
-                break
-
-        if lm_head_key and not embed_tokens_key:
-            embed_tokens_key = lm_head_key.replace(
-                ".lm_head.weight", ".model.language_model.embed_tokens.weight"
-            )
-            transformed_dict[embed_tokens_key] = transformed_dict[lm_head_key]
-        elif embed_tokens_key and not lm_head_key:
-            lm_head_key = embed_tokens_key.replace(
-                ".model.language_model.embed_tokens.weight", ".lm_head.weight"
-            )
-            transformed_dict[lm_head_key] = transformed_dict[embed_tokens_key]
-
-        return transformed_dict
-
-    @classmethod
-    def _load_as_safetensor(
-        cls, model: "PI0Policy", model_file: str, map_location: str, strict: bool
-    ) -> "PI0Policy":
-        """Override to apply key transformations before loading."""
-        from safetensors.torch import load_file
-
-        init_logging()
-        # Load the state dict from file safely
-        state_dict = load_file(model_file, device=map_location)
-
-        # Apply key transformations
-        transformed_state_dict = cls._transform_state_dict_keys(state_dict)
-
-        # Load the transformed state dict
-        msg = model.load_state_dict(transformed_state_dict, strict=strict)
-
-        # Log message
-        log_model_loading_keys(msg.missing_keys, msg.unexpected_keys)
-        return model
-
     def get_optim_params(self) -> dict:
         return self.parameters()
 
@@ -377,14 +270,13 @@ class PI0Policy(PreTrainedPolicy):
         if self.config.adapt_to_pi_aloha:
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
 
-        batch = self.normalize_inputs(batch)
-
         # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
         # querying the policy.
         if len(self._action_queue) == 0:
             images, img_masks = self.prepare_images(batch)
             state = self.prepare_state(batch)
-            lang_tokens, lang_masks = self.prepare_language(batch)
+            lang_tokens = batch[f"{OBS_LANGUAGE_TOKENS}"]
+            lang_masks = batch[f"{OBS_LANGUAGE_ATTENTION_MASK}"]
 
             actions = self.model.sample_actions(
                 images, img_masks, lang_tokens, lang_masks, state, noise=noise
@@ -394,8 +286,6 @@ class PI0Policy(PreTrainedPolicy):
             original_action_dim = self.config.action_feature.shape[0]
             actions = actions[:, :, :original_action_dim]
 
-            actions = self.unnormalize_outputs({"action": actions})["action"]
-
             if self.config.adapt_to_pi_aloha:
                 actions = self._pi_aloha_encode_actions(actions)
 
@@ -410,12 +300,10 @@ class PI0Policy(PreTrainedPolicy):
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
             batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
 
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-
         images, img_masks = self.prepare_images(batch)
         state = self.prepare_state(batch)
-        lang_tokens, lang_masks = self.prepare_language(batch)
+        lang_tokens = batch[f"{OBS_LANGUAGE_TOKENS}"]
+        lang_masks = batch[f"{OBS_LANGUAGE_ATTENTION_MASK}"]
         actions = self.prepare_action(batch)
         actions_is_pad = batch.get("action_is_pad")
 
@@ -482,26 +370,6 @@ class PI0Policy(PreTrainedPolicy):
 
         return images, img_masks
 
-    def prepare_language(self, batch) -> tuple[Tensor, Tensor]:
-        """Tokenize the text input"""
-        device = batch[OBS_STATE].device
-        tasks = batch["task"]
-
-        # PaliGemma prompt has to end with a new line
-        tasks = [task if task.endswith("\n") else f"{task}\n" for task in tasks]
-
-        tokenized_prompt = self.language_tokenizer.__call__(
-            tasks,
-            padding="max_length",
-            padding_side="right",
-            max_length=self.config.tokenizer_max_length,
-            return_tensors="pt",
-        )
-        lang_tokens = tokenized_prompt["input_ids"].to(device=device)
-        lang_masks = tokenized_prompt["attention_mask"].to(device=device, dtype=torch.bool)
-
-        return lang_tokens, lang_masks
-
     def _pi_aloha_decode_state(self, state):
         # Flip the joints.
         for motor_idx in [1, 2, 8, 9]:
@@ -567,7 +435,7 @@ class PI0FlowMatching(nn.Module):
     └──────────────────────────────┘
     """
 
-    def __init__(self, config):
+    def __init__(self, config: PI0Config):
         super().__init__()
         self.config = config
 

src/lerobot/policies/pi0/processor_pi0.py

@@ -0,0 +1,166 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any
+
+import torch
+
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    ComplementaryDataProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    RenameObservationsProcessorStep,
+    TokenizerProcessorStep,
+    UnnormalizerProcessorStep,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+
+@ProcessorStepRegistry.register(name="pi0_new_line_processor")
+class Pi0NewLineProcessor(ComplementaryDataProcessorStep):
+    """
+    Ensures that the task description string ends with a newline character.
+
+    This processing step is required for compatibility with the PaliGemma tokenizer,
+    which expects a newline at the end of the text prompt. It handles both single
+    strings and lists of strings for the 'task' key in complementary data.
+    """
+
+    def complementary_data(self, complementary_data):
+        """
+        Adds a newline to the 'task' field if it doesn't already have one.
+
+        Args:
+            complementary_data: A dictionary that may contain a 'task' key with a
+                                string or list of strings.
+
+        Returns:
+            A new dictionary with the modified 'task' field.
+        """
+        if "task" not in complementary_data:
+            return complementary_data
+
+        task = complementary_data["task"]
+        if task is None:
+            return complementary_data
+
+        new_complementary_data = dict(complementary_data)
+
+        # Handle both string and list of strings
+        if isinstance(task, str):
+            # Single string: add newline if not present
+            if not task.endswith("\n"):
+                new_complementary_data["task"] = f"{task}\n"
+        elif isinstance(task, list) and all(isinstance(t, str) for t in task):
+            # List of strings: add newline to each if not present
+            new_complementary_data["task"] = [t if t.endswith("\n") else f"{t}\n" for t in task]
+        # If task is neither string nor list of strings, leave unchanged
+
+        return new_complementary_data
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        This step does not alter the feature definitions.
+
+        Args:
+            features: The input feature dictionary.
+
+        Returns:
+            The unchanged feature dictionary.
+        """
+        return features
+
+
+def make_pi0_pre_post_processors(
+    config: PI0Config,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Constructs pre-processor and post-processor pipelines for the PI0 policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1. Renaming features to match pretrained configurations.
+    2. Normalizing input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Appending a newline character to the task description for tokenizer compatibility.
+    5. Tokenizing the text prompt using the PaliGemma tokenizer.
+    6. Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the PI0 policy.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+
+    # Add remaining processors
+    input_steps: list[ProcessorStep] = [
+        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
+        AddBatchDimensionProcessorStep(),
+        Pi0NewLineProcessor(),  # Add newlines before tokenization for PaliGemma
+        TokenizerProcessorStep(
+            tokenizer_name="google/paligemma-3b-pt-224",
+            max_length=config.tokenizer_max_length,
+            padding_side="right",
+            padding="max_length",
+        ),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+
+    output_steps: list[ProcessorStep] = [
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )

src/lerobot/policies/pi0fast/configuration_pi0fast.py

@@ -6,6 +6,7 @@ from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import (
     CosineDecayWithWarmupSchedulerConfig,
 )
+from lerobot.utils.constants import OBS_IMAGES
 
 
 @PreTrainedConfig.register_subclass("pi0fast")
@@ -99,7 +100,7 @@ class PI0FASTConfig(PreTrainedConfig):
 
     def validate_features(self) -> None:
         for i in range(self.empty_cameras):
-            key = f"observation.images.empty_camera_{i}"
+            key = f"{OBS_IMAGES}.empty_camera_{i}"
             empty_camera = PolicyFeature(
                 type=FeatureType.VISUAL,
                 shape=(3, 480, 640),

src/lerobot/policies/pi0fast/modeling_pi0fast.py

@@ -57,10 +57,9 @@ from transformers import AutoProcessor, AutoTokenizer, PaliGemmaForConditionalGe
 from transformers.cache_utils import HybridCache, StaticCache
 from transformers.models.auto import CONFIG_MAPPING
 
-from lerobot.constants import ACTION, OBS_STATE
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
 from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.utils.constants import ACTION, OBS_STATE
 
 PRECISION = {
     "float16": torch.float16,
@@ -146,14 +145,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
         config.validate_features()
         self.config = config
 
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         self.language_tokenizer = AutoProcessor.from_pretrained("google/paligemma-3b-pt-224")
         self.model = PI0FAST(config)
 
@@ -221,8 +212,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
         if self.config.adapt_to_pi_aloha:
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
 
-        batch = self.normalize_inputs(batch)
-
         # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
         # querying the policy.
         if len(self._action_queue) == 0:
@@ -235,8 +224,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
             ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
             actions = actions[:, :, :original_action_dim]
 
-            actions = self.unnormalize_outputs({"action": actions})["action"]
-
             if self.config.adapt_to_pi_aloha:
                 actions = self._pi_aloha_encode_actions(actions)
 
@@ -249,8 +236,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
         if self.config.adapt_to_pi_aloha:
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
             batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
         loss_dict = self.model.forward(batch)
         return loss_dict["loss"], loss_dict
 

src/lerobot/policies/pi0fast/processor_pi0fast.py

@@ -0,0 +1,92 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any
+
+import torch
+
+from lerobot.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    RenameObservationsProcessorStep,
+    UnnormalizerProcessorStep,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+
+def make_pi0fast_pre_post_processors(
+    config: PI0FASTConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Constructs pre-processor and post-processor pipelines for the PI0Fast policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1. Renaming features to match pretrained configurations.
+    2. Normalizing input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the PI0Fast policy.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+    output_steps = [
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )

src/lerobot/policies/pretrained.py

@@ -246,7 +246,9 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
             base_model=base_model,
         )
 
-        template_card = files("lerobot.templates").joinpath("lerobot_modelcard_template.md").read_text()
+        template_card = (
+            files("lerobot.templates").joinpath("lerobot_modelcard_template.md").read_text(encoding="utf-8")
+        )
         card = ModelCard.from_template(card_data, template_str=template_card)
         card.validate()
         return card

src/lerobot/policies/sac/configuration_sac.py

@@ -19,8 +19,8 @@ from dataclasses import dataclass, field
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import NormalizationMode
-from lerobot.constants import ACTION, OBS_IMAGE, OBS_STATE
 from lerobot.optim.optimizers import MultiAdamConfig
+from lerobot.utils.constants import ACTION, OBS_IMAGE, OBS_STATE
 
 
 def is_image_feature(key: str) -> bool:
@@ -139,8 +139,6 @@ class SACConfig(PreTrainedConfig):
     # Training parameter
     # Number of steps for online training
     online_steps: int = 1000000
-    # Seed for the online environment
-    online_env_seed: int = 10000
     # Capacity of the online replay buffer
     online_buffer_capacity: int = 100000
     # Capacity of the offline replay buffer
@@ -225,7 +223,7 @@ class SACConfig(PreTrainedConfig):
                 "You must provide either 'observation.state' or an image observation (key starting with 'observation.image') in the input features"
             )
 
-        if "action" not in self.output_features:
+        if ACTION not in self.output_features:
             raise ValueError("You must provide 'action' in the output features")
 
     @property

src/lerobot/policies/sac/modeling_sac.py

@@ -28,10 +28,10 @@ import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 from torch.distributions import MultivariateNormal, TanhTransform, Transform, TransformedDistribution
 
-from lerobot.policies.normalize import NormalizeBuffer
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.sac.configuration_sac import SACConfig, is_image_feature
 from lerobot.policies.utils import get_device_from_parameters
+from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_STATE
 
 DISCRETE_DIMENSION_INDEX = -1  # Gripper is always the last dimension
 
@@ -45,15 +45,13 @@ class SACPolicy(
     def __init__(
         self,
         config: SACConfig | None = None,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         super().__init__(config)
         config.validate_features()
         self.config = config
 
         # Determine action dimension and initialize all components
-        continuous_action_dim = config.output_features["action"].shape[0]
-        self._init_normalization(dataset_stats)
+        continuous_action_dim = config.output_features[ACTION].shape[0]
         self._init_encoders()
         self._init_critics(continuous_action_dim)
         self._init_actor(continuous_action_dim)
@@ -88,8 +86,7 @@ class SACPolicy(
 
         observations_features = None
         if self.shared_encoder and self.actor.encoder.has_images:
-            # Cache and normalize image features
-            observations_features = self.actor.encoder.get_cached_image_features(batch, normalize=True)
+            observations_features = self.actor.encoder.get_cached_image_features(batch)
 
         actions, _, _ = self.actor(batch, observations_features)
 
@@ -161,7 +158,7 @@ class SACPolicy(
             The computed loss tensor
         """
         # Extract common components from batch
-        actions: Tensor = batch["action"]
+        actions: Tensor = batch[ACTION]
         observations: dict[str, Tensor] = batch["state"]
         observation_features: Tensor = batch.get("observation_feature")
 
@@ -391,28 +388,12 @@ class SACPolicy(
         actor_loss = ((self.temperature * log_probs) - min_q_preds).mean()
         return actor_loss
 
-    def _init_normalization(self, dataset_stats):
-        """Initialize input/output normalization modules."""
-        self.normalize_inputs = nn.Identity()
-        self.normalize_targets = nn.Identity()
-        if self.config.dataset_stats is not None:
-            params = _convert_normalization_params_to_tensor(self.config.dataset_stats)
-            self.normalize_inputs = NormalizeBuffer(
-                self.config.input_features, self.config.normalization_mapping, params
-            )
-            stats = dataset_stats or params
-            self.normalize_targets = NormalizeBuffer(
-                self.config.output_features, self.config.normalization_mapping, stats
-            )
-
     def _init_encoders(self):
         """Initialize shared or separate encoders for actor and critic."""
         self.shared_encoder = self.config.shared_encoder
-        self.encoder_critic = SACObservationEncoder(self.config, self.normalize_inputs)
+        self.encoder_critic = SACObservationEncoder(self.config)
         self.encoder_actor = (
-            self.encoder_critic
-            if self.shared_encoder
-            else SACObservationEncoder(self.config, self.normalize_inputs)
+            self.encoder_critic if self.shared_encoder else SACObservationEncoder(self.config)
         )
 
     def _init_critics(self, continuous_action_dim):
@@ -424,9 +405,7 @@ class SACPolicy(
             )
             for _ in range(self.config.num_critics)
         ]
-        self.critic_ensemble = CriticEnsemble(
-            encoder=self.encoder_critic, ensemble=heads, output_normalization=self.normalize_targets
-        )
+        self.critic_ensemble = CriticEnsemble(encoder=self.encoder_critic, ensemble=heads)
         target_heads = [
             CriticHead(
                 input_dim=self.encoder_critic.output_dim + continuous_action_dim,
@@ -434,9 +413,7 @@ class SACPolicy(
             )
             for _ in range(self.config.num_critics)
         ]
-        self.critic_target = CriticEnsemble(
-            encoder=self.encoder_critic, ensemble=target_heads, output_normalization=self.normalize_targets
-        )
+        self.critic_target = CriticEnsemble(encoder=self.encoder_critic, ensemble=target_heads)
         self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
 
         if self.config.use_torch_compile:
@@ -490,10 +467,9 @@ class SACPolicy(
 class SACObservationEncoder(nn.Module):
     """Encode image and/or state vector observations."""
 
-    def __init__(self, config: SACConfig, input_normalizer: nn.Module) -> None:
+    def __init__(self, config: SACConfig) -> None:
         super().__init__()
         self.config = config
-        self.input_normalization = input_normalizer
         self._init_image_layers()
         self._init_state_layers()
         self._compute_output_dim()
@@ -538,17 +514,17 @@ class SACObservationEncoder(nn.Module):
             )
 
     def _init_state_layers(self) -> None:
-        self.has_env = "observation.environment_state" in self.config.input_features
-        self.has_state = "observation.state" in self.config.input_features
+        self.has_env = OBS_ENV_STATE in self.config.input_features
+        self.has_state = OBS_STATE in self.config.input_features
         if self.has_env:
-            dim = self.config.input_features["observation.environment_state"].shape[0]
+            dim = self.config.input_features[OBS_ENV_STATE].shape[0]
             self.env_encoder = nn.Sequential(
                 nn.Linear(dim, self.config.latent_dim),
                 nn.LayerNorm(self.config.latent_dim),
                 nn.Tanh(),
             )
         if self.has_state:
-            dim = self.config.input_features["observation.state"].shape[0]
+            dim = self.config.input_features[OBS_STATE].shape[0]
             self.state_encoder = nn.Sequential(
                 nn.Linear(dim, self.config.latent_dim),
                 nn.LayerNorm(self.config.latent_dim),
@@ -568,16 +544,15 @@ class SACObservationEncoder(nn.Module):
     def forward(
         self, obs: dict[str, Tensor], cache: dict[str, Tensor] | None = None, detach: bool = False
     ) -> Tensor:
-        obs = self.input_normalization(obs)
         parts = []
         if self.has_images:
             if cache is None:
-                cache = self.get_cached_image_features(obs, normalize=False)
+                cache = self.get_cached_image_features(obs)
             parts.append(self._encode_images(cache, detach))
         if self.has_env:
-            parts.append(self.env_encoder(obs["observation.environment_state"]))
+            parts.append(self.env_encoder(obs[OBS_ENV_STATE]))
         if self.has_state:
-            parts.append(self.state_encoder(obs["observation.state"]))
+            parts.append(self.state_encoder(obs[OBS_STATE]))
         if parts:
             return torch.cat(parts, dim=-1)
 
@@ -585,7 +560,7 @@ class SACObservationEncoder(nn.Module):
             "No parts to concatenate, you should have at least one image or environment state or state"
         )
 
-    def get_cached_image_features(self, obs: dict[str, Tensor], normalize: bool = False) -> dict[str, Tensor]:
+    def get_cached_image_features(self, obs: dict[str, Tensor]) -> dict[str, Tensor]:
         """Extract and optionally cache image features from observations.
 
         This function processes image observations through the vision encoder once and returns
@@ -597,26 +572,17 @@ class SACObservationEncoder(nn.Module):
         - The vision encoder forward pass is typically the main computational bottleneck during training and inference
         - Caching these features can provide 2-4x speedup in training and inference
 
-        Normalization behavior:
-        - When called from inside forward(): set normalize=False since inputs are already normalized
-        - When called from outside forward(): set normalize=True to ensure proper input normalization
-
         Usage patterns:
-        - Called in select_action() with normalize=True
+        - Called in select_action()
         - Called in learner.py's get_observation_features() to pre-compute features for all policy components
-        - Called internally by forward() with normalize=False
+        - Called internally by forward()
 
         Args:
             obs: Dictionary of observation tensors containing image keys
-            normalize: Whether to normalize observations before encoding
-                      Set to True when calling directly from outside the encoder's forward method
-                      Set to False when calling from within forward() where inputs are already normalized
 
         Returns:
             Dictionary mapping image keys to their corresponding encoded features
         """
-        if normalize:
-            obs = self.input_normalization(obs)
         batched = torch.cat([obs[k] for k in self.image_keys], dim=0)
         out = self.image_encoder(batched)
         chunks = torch.chunk(out, len(self.image_keys), dim=0)
@@ -747,7 +713,6 @@ class CriticEnsemble(nn.Module):
     Args:
         encoder (SACObservationEncoder): encoder for observations.
         ensemble (List[CriticHead]): list of critic heads.
-        output_normalization (nn.Module): normalization layer for actions.
         init_final (float | None): optional initializer scale for final layers.
 
     Forward returns a tensor of shape (num_critics, batch_size) containing Q-values.
@@ -757,13 +722,11 @@ class CriticEnsemble(nn.Module):
         self,
         encoder: SACObservationEncoder,
         ensemble: list[CriticHead],
-        output_normalization: nn.Module,
         init_final: float | None = None,
     ):
         super().__init__()
         self.encoder = encoder
         self.init_final = init_final
-        self.output_normalization = output_normalization
         self.critics = nn.ModuleList(ensemble)
 
     def forward(
@@ -775,11 +738,6 @@ class CriticEnsemble(nn.Module):
         device = get_device_from_parameters(self)
         # Move each tensor in observations to device
         observations = {k: v.to(device) for k, v in observations.items()}
-        # NOTE: We normalize actions it helps for sample efficiency
-        actions: dict[str, torch.tensor] = {"action": actions}
-        # NOTE: Normalization layer took dict in input and outputs a dict that why
-        actions = self.output_normalization(actions)["action"]
-        actions = actions.to(device)
 
         obs_enc = self.encoder(observations, cache=observation_features)
 
@@ -1103,15 +1061,3 @@ class TanhMultivariateNormalDiag(TransformedDistribution):
             x = transform(x)
 
         return x
-
-
-def _convert_normalization_params_to_tensor(normalization_params: dict) -> dict:
-    converted_params = {}
-    for outer_key, inner_dict in normalization_params.items():
-        converted_params[outer_key] = {}
-        for key, value in inner_dict.items():
-            converted_params[outer_key][key] = torch.tensor(value)
-            if "image" in outer_key:
-                converted_params[outer_key][key] = converted_params[outer_key][key].view(3, 1, 1)
-
-    return converted_params

src/lerobot/policies/sac/processor_sac.py

@@ -0,0 +1,92 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any
+
+import torch
+
+from lerobot.policies.sac.configuration_sac import SACConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    RenameObservationsProcessorStep,
+    UnnormalizerProcessorStep,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+
+def make_sac_pre_post_processors(
+    config: SACConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Constructs pre-processor and post-processor pipelines for the SAC policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1. Renaming features to match pretrained configurations.
+    2. Normalizing input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the SAC policy.
+        dataset_stats: A dictionary of statistics for normalization.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+
+    # Add remaining processors
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+    output_steps = [
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )

src/lerobot/policies/sac/reward_model/configuration_classifier.py

@@ -19,6 +19,7 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import NormalizationMode
 from lerobot.optim.optimizers import AdamWConfig, OptimizerConfig
 from lerobot.optim.schedulers import LRSchedulerConfig
+from lerobot.utils.constants import OBS_IMAGE
 
 
 @PreTrainedConfig.register_subclass(name="reward_classifier")
@@ -69,7 +70,7 @@ class RewardClassifierConfig(PreTrainedConfig):
 
     def validate_features(self) -> None:
         """Validate feature configurations."""
-        has_image = any(key.startswith("observation.image") for key in self.input_features)
+        has_image = any(key.startswith(OBS_IMAGE) for key in self.input_features)
         if not has_image:
             raise ValueError(
                 "You must provide an image observation (key starting with 'observation.image') in the input features"

src/lerobot/policies/sac/reward_model/modeling_classifier.py

@@ -19,10 +19,9 @@ import logging
 import torch
 from torch import Tensor, nn
 
-from lerobot.constants import OBS_IMAGE, REWARD
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
+from lerobot.utils.constants import OBS_IMAGE, REWARD
 
 
 class ClassifierOutput:
@@ -108,22 +107,12 @@ class Classifier(PreTrainedPolicy):
     def __init__(
         self,
         config: RewardClassifierConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         from transformers import AutoModel
 
         super().__init__(config)
         self.config = config
 
-        # Initialize normalization (standardized with the policy framework)
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         # Set up encoder
         encoder = AutoModel.from_pretrained(self.config.model_name, trust_remote_code=True)
         # Extract vision model if we're given a multimodal model
@@ -247,10 +236,6 @@ class Classifier(PreTrainedPolicy):
 
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict[str, Tensor]]:
         """Standard forward pass for training compatible with train.py."""
-        # Normalize inputs if needed
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-
         # Extract images and labels
         images, labels = self.extract_images_and_labels(batch)
 

src/lerobot/policies/sac/reward_model/processor_classifier.py

@@ -0,0 +1,82 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any
+
+import torch
+
+from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
+from lerobot.processor import (
+    DeviceProcessorStep,
+    IdentityProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+
+
+def make_classifier_processor(
+    config: RewardClassifierConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Constructs pre-processor and post-processor pipelines for the reward classifier.
+
+    The pre-processing pipeline prepares input data for the classifier by:
+    1. Normalizing both input and output features based on dataset statistics.
+    2. Moving the data to the specified device.
+
+    The post-processing pipeline handles the classifier's output by:
+    1. Moving the data to the CPU.
+    2. Applying an identity step, as no unnormalization is needed for the output logits.
+
+    Args:
+        config: The configuration object for the RewardClassifier.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+
+    input_steps = [
+        NormalizerProcessorStep(
+            features=config.input_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        NormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        DeviceProcessorStep(device=config.device),
+    ]
+    output_steps = [DeviceProcessorStep(device="cpu"), IdentityProcessorStep()]
+
+    return (
+        PolicyProcessorPipeline(
+            steps=input_steps,
+            name="classifier_preprocessor",
+        ),
+        PolicyProcessorPipeline(
+            steps=output_steps,
+            name="classifier_postprocessor",
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )