feat: add offline training in learner

feat: add RLT algorithm
feat: add RLT policy RL-token encoder-decoder and actor
2026-05-30 18:31:25 +00:00 · 2026-03-22 23:00:07 +01:00 · 2026-03-22 22:59:35 +01:00 · 2026-03-22 22:57:43 +01:00 · 2026-03-22 22:52:56 +01:00 · 2026-03-19 13:42:46 +01:00
106 changed files with 3923 additions and 1423 deletions
--- a/.github/workflows/full_tests.yml
+++ b/.github/workflows/full_tests.yml
@@ -101,9 +101,11 @@ jobs:
    runs-on:
      group: aws-general-8-plus
    if: |
-      (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
-      github.event_name == 'push' ||
-      github.event_name == 'workflow_dispatch'
+      github.repository == 'huggingface/lerobot' && (
+        (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
+        github.event_name == 'push' ||
+        github.event_name == 'workflow_dispatch'
+      )
    outputs:
      image_tag: ${{ steps.set_tag.outputs.image_tag }}
    env:
--- a/.github/workflows/unbound_deps_tests.yml
+++ b/.github/workflows/unbound_deps_tests.yml
@@ -91,6 +91,7 @@ jobs:
    name: Build and Push Docker
    runs-on:
      group: aws-general-8-plus
+    if: github.repository == 'huggingface/lerobot'
    outputs:
      image_tag: ${{ env.DOCKER_IMAGE_NAME }}
    env:
--- a/benchmarks/video/README.md
+++ b/benchmarks/video/README.md
@@ -28,9 +28,9 @@ We don't expect the same optimal settings for a dataset of images from a simulat
 For these reasons, we run this benchmark on four representative datasets:

 - `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
- `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
- `aliberts/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
- `aliberts/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
+- `lerobot/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
+- `lerobot/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
+- `lerobot/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.

 Note: The datasets used for this benchmark need to be image datasets, not video datasets.

@@ -179,7 +179,7 @@ python benchmark/video/run_video_benchmark.py \
    --output-dir outputs/video_benchmark \
    --repo-ids \
        lerobot/pusht_image \
-        aliberts/aloha_mobile_shrimp_image \
+        lerobot/aloha_mobile_shrimp_image \
    --vcodec libx264 libx265 \
    --pix-fmt yuv444p yuv420p \
    --g 2 20 None \
@@ -203,9 +203,9 @@ python benchmark/video/run_video_benchmark.py \
    --output-dir outputs/video_benchmark \
    --repo-ids \
        lerobot/pusht_image \
-        aliberts/aloha_mobile_shrimp_image \
-        aliberts/paris_street \
-        aliberts/kitchen \
+        lerobot/aloha_mobile_shrimp_image \
+        lerobot/paris_street \
+        lerobot/kitchen \
    --vcodec libx264 libx265 \
    --pix-fmt yuv444p yuv420p \
    --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -221,9 +221,9 @@ python benchmark/video/run_video_benchmark.py \
    --output-dir outputs/video_benchmark \
    --repo-ids \
        lerobot/pusht_image \
-        aliberts/aloha_mobile_shrimp_image \
-        aliberts/paris_street \
-        aliberts/kitchen \
+        lerobot/aloha_mobile_shrimp_image \
+        lerobot/paris_street \
+        lerobot/kitchen \
    --vcodec libsvtav1 \
    --pix-fmt yuv420p \
    --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -252,37 +252,37 @@ Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_read

 These tables show the results for `g=2` and `crf=30`, using `timestamps-modes=6_frames` and `backend=pyav`

-| video_images_size_ratio            | vcodec     | pix_fmt |           |           |           |
-| ---------------------------------- | ---------- | ------- | --------- | --------- | --------- |
-|                                    | libx264    |         | libx265   |           | libsvtav1 |
-| repo_id                            | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
-| lerobot/pusht_image                | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
-| aliberts/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
-| aliberts/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
-| aliberts/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
+| video_images_size_ratio           | vcodec     | pix_fmt |           |           |           |
+| --------------------------------- | ---------- | ------- | --------- | --------- | --------- |
+|                                   | libx264    |         | libx265   |           | libsvtav1 |
+| repo_id                           | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
+| lerobot/pusht_image               | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
+| lerobot/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
+| lerobot/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
+| lerobot/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |

-| video_images_load_time_ratio       | vcodec  | pix_fmt |          |         |           |
-| ---------------------------------- | ------- | ------- | -------- | ------- | --------- |
-|                                    | libx264 |         | libx265  |         | libsvtav1 |
-| repo_id                            | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
-| lerobot/pusht_image                | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
-| aliberts/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
-| aliberts/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
-| aliberts/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
+| video_images_load_time_ratio      | vcodec  | pix_fmt |          |         |           |
+| --------------------------------- | ------- | ------- | -------- | ------- | --------- |
+|                                   | libx264 |         | libx265  |         | libsvtav1 |
+| repo_id                           | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
+| lerobot/pusht_image               | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
+| lerobot/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
+| lerobot/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
+| lerobot/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |

-|                                    |          | vcodec   | pix_fmt      |          |           |              |
-| ---------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
-|                                    |          | libx264  |              | libx265  |           | libsvtav1    |
-| repo_id                            | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
-| lerobot/pusht_image                | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
-|                                    | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
-|                                    | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
-| aliberts/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
-|                                    | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
-|                                    | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
-| aliberts/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
-|                                    | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
-|                                    | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
-| aliberts/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
-|                                    | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
-|                                    | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |
+|                                   |          | vcodec   | pix_fmt      |          |           |              |
+| --------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
+|                                   |          | libx264  |              | libx265  |           | libsvtav1    |
+| repo_id                           | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
+| lerobot/pusht_image               | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
+|                                   | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
+|                                   | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
+| lerobot/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
+|                                   | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
+|                                   | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
+| lerobot/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
+|                                   | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
+|                                   | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
+| lerobot/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
+|                                   | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
+|                                   | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |
--- a/docs/source/earthrover_mini_plus.mdx
+++ b/docs/source/earthrover_mini_plus.mdx
@@ -185,7 +185,7 @@ echo $HF_USER
 Use the standard recording command:

 ```bash
-python src/lerobot/scripts/lerobot_record.py \
+lerobot-record \
    --robot.type=earthrover_mini_plus \
    --teleop.type=keyboard_rover \
    --dataset.repo_id=your_username/dataset_name \
--- a/docs/source/hope_jr.mdx
+++ b/docs/source/hope_jr.mdx
@@ -224,7 +224,7 @@ lerobot-record \
    --teleop.port=/dev/tty.usbmodem1201 \
    --teleop.id=right \
    --teleop.side=right \
-    --dataset.repo_id=nepyope/hand_record_test_with_video_data \
+    --dataset.repo_id=<USER>/hand_record_test_with_video_data \
    --dataset.single_task="Hand recording test with video data" \
    --dataset.num_episodes=1 \
    --dataset.episode_time_s=5 \
@@ -241,7 +241,7 @@ lerobot-replay \
    --robot.port=/dev/tty.usbmodem58760432281 \
    --robot.id=right \
    --robot.side=right \
-    --dataset.repo_id=nepyope/hand_record_test_with_camera \
+    --dataset.repo_id=<USER>/hand_record_test_with_camera \
    --dataset.episode=0
 ```

@@ -249,13 +249,13 @@ lerobot-replay \

 ```bash
 lerobot-train \
-  --dataset.repo_id=nepyope/hand_record_test_with_video_data \
+  --dataset.repo_id=<USER>/hand_record_test_with_video_data \
  --policy.type=act \
  --output_dir=outputs/train/hopejr_hand \
  --job_name=hopejr \
  --policy.device=mps \
  --wandb.enable=true \
-  --policy.repo_id=nepyope/hand_test_policy
+  --policy.repo_id=<USER>/hand_test_policy
 ```

 ### Evaluate
@@ -270,7 +270,7 @@ lerobot-record \
  --robot.side=right \
  --robot.cameras='{"main": {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}}' \
  --display_data=false \
-  --dataset.repo_id=nepyope/eval_hopejr \
+  --dataset.repo_id=<USER>/eval_hopejr \
  --dataset.single_task="Evaluate hopejr hand policy" \
  --dataset.num_episodes=10 \
  --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
--- a/docs/source/installation.mdx
+++ b/docs/source/installation.mdx
@@ -1,13 +1,15 @@
 # Installation

-## Install [`miniforge`](https://conda-forge.org/download/)
+This guide uses conda (via miniforge) to manage environments. If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.10 and ffmpeg installed with the `libsvtav1` encoder, then skip ahead to [Install LeRobot](#step-3-install-lerobot-).
+
+## Step 1: Install [`miniforge`](https://conda-forge.org/download/)

 ```bash
 wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
 bash Miniforge3-$(uname)-$(uname -m).sh
 ```

-## Environment Setup
+## Step 2: Environment Setup

 Create a virtual environment with Python 3.10, using conda:

@@ -38,7 +40,7 @@ conda install ffmpeg -c conda-forge
 >
 > - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.

-## Install LeRobot 🤗
+## Step 3: Install LeRobot 🤗

 ### From Source

--- a/docs/source/pi0.mdx
+++ b/docs/source/pi0.mdx
@@ -60,7 +60,7 @@ policy.type=pi0
 For training π₀, you can use the standard LeRobot training script with the appropriate configuration:

 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
    --dataset.repo_id=your_dataset \
    --policy.type=pi0 \
    --output_dir=./outputs/pi0_training \
--- a/docs/source/pi05.mdx
+++ b/docs/source/pi05.mdx
@@ -56,7 +56,7 @@ policy.type=pi05
 Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:

 ```bash
-python src/lerobot/scripts/lerobot_train.py\
+lerobot-train \
    --dataset.repo_id=your_dataset \
    --policy.type=pi05 \
    --output_dir=./outputs/pi05_training \
--- a/docs/source/sarm.mdx
+++ b/docs/source/sarm.mdx
@@ -269,7 +269,7 @@ This generates visualizations showing video frames with subtask boundaries overl
 Train with **no annotations** - uses linear progress from 0 to 1:

 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=single_stage \
@@ -288,7 +288,7 @@ python src/lerobot/scripts/lerobot_train.py \
 Train with **dense annotations only** (sparse auto-generated):

 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=dense_only \
@@ -307,7 +307,7 @@ python src/lerobot/scripts/lerobot_train.py \
 Train with **both sparse and dense annotations**:

 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=dual \
@@ -468,7 +468,7 @@ This script:
 Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`). Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:

 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=pi0 \
  --use_rabc=true \
--- a/docs/source/unitree_g1.mdx
+++ b/docs/source/unitree_g1.mdx
@@ -216,7 +216,7 @@ lerobot-teleoperate \
 ### Record Dataset in Simulation

 ```bash
-python -m lerobot.scripts.lerobot_record \
+lerobot-record \
    --robot.type=unitree_g1 \
    --robot.is_simulation=true \
    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
@@ -266,7 +266,7 @@ lerobot-teleoperate \
 ### Record Dataset on Real Robot

 ```bash
-python -m lerobot.scripts.lerobot_record \
+lerobot-record \
    --robot.type=unitree_g1 \
    --robot.is_simulation=false \
    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "172.18.129.215", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
--- a/docs/source/using_dataset_tools.mdx
+++ b/docs/source/using_dataset_tools.mdx
@@ -12,6 +12,7 @@ LeRobot provides several utilities for manipulating datasets:
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
 6. **Convert to Video** - Convert image-based datasets to video format for efficient storage
+7. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.

 The core implementation is in `lerobot.datasets.dataset_tools`.
 An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
@@ -156,6 +157,30 @@ lerobot-edit-dataset \

 **Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). All episodes, stats, and tasks are preserved.

+### Show the information of datasets
+
+Show the information of datasets such as number of episode, number of frame, File size and so on.
+No change will be made to the dataset
+
+```bash
+
+# Show dataset information without feature details
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type info \
+
+# Show dataset information with feature details
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type info \
+    --operation.show_features true
+
+```
+
+**Parameters:**
+
+- `parameters`: The flag to control show or no show dataset information with feature details.(default=false)
+
 ### Push to Hub

 Add the `--push_to_hub true` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:
--- a/docs/source/walloss.mdx
+++ b/docs/source/walloss.mdx
@@ -45,7 +45,7 @@ policy.type=wall_x
 For training WallX, you can use the standard LeRobot training script with the appropriate configuration:

 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
    --dataset.repo_id=your_dataset \
    --policy.type=wall_x \
    --output_dir=./outputs/wallx_training \
--- a/docs/source/xvla.mdx
+++ b/docs/source/xvla.mdx
@@ -154,7 +154,7 @@ lerobot-train \

 ```bash
 lerobot-train \
-  --dataset.repo_id=pepijn223/bimanual-so100-handover-cube \
+  --dataset.repo_id=<USER>/bimanual-so100-handover-cube \
  --output_dir=./outputs/xvla_bimanual \
  --job_name=xvla_so101_training \
  --policy.path="lerobot/xvla-base" \
--- a/examples/backward_compatibility/replay.py
+++ b/examples/backward_compatibility/replay.py
@@ -22,7 +22,7 @@ lerobot-replay \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=black \
-    --dataset.repo_id=aliberts/record-test \
+    --dataset.repo_id=<USER>/record-test \
    --dataset.episode=2
 ```
 """
--- a/examples/rtc/eval_dataset.py
+++ b/examples/rtc/eval_dataset.py
@@ -27,8 +27,8 @@ measuring consistency and ground truth alignment.
 Usage:
    # Basic usage with smolvla policy
    uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
        --rtc.execution_horizon=8 \
        --device=mps \
        --rtc.max_guidance_weight=10.0 \
@@ -58,16 +58,16 @@ Usage:
        --device=cuda

    uv run python examples/rtc/eval_dataset.py \
-        --policy.path=lipsop/reuben_pi0 \
-        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
+        --policy.path=<USER>/reuben_pi0 \
+        --dataset.repo_id=<USER>/so101_cube_in_cup \
        --rtc.execution_horizon=8 \
        --device=cuda

    # With torch.compile for faster inference (PyTorch 2.0+)
    # Note: CUDA graphs disabled by default due to in-place ops in denoising loop
    uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
        --rtc.execution_horizon=8 \
        --device=mps \
        --use_torch_compile=true \
@@ -75,8 +75,8 @@ Usage:

    # With torch.compile on CUDA (CUDA graphs disabled by default)
    uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
        --rtc.execution_horizon=8 \
        --device=cuda \
        --use_torch_compile=true \
@@ -84,8 +84,8 @@ Usage:

    # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
    uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
        --use_torch_compile=true \
        --torch_compile_backend=inductor \
        --torch_compile_mode=max-autotune \
--- a/examples/rtc/eval_with_real_robot.py
+++ b/examples/rtc/eval_with_real_robot.py
@@ -28,7 +28,7 @@ For simulation environments, see eval_with_simulation.py
 Usage:
    # Run RTC with Real robot with RTC
    uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
        --policy.device=mps \
        --rtc.enabled=true \
        --rtc.execution_horizon=20 \
@@ -41,7 +41,7 @@ Usage:

    # Run RTC with Real robot without RTC
    uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
        --policy.device=mps \
        --rtc.enabled=false \
        --robot.type=so100_follower \
@@ -53,7 +53,7 @@ Usage:

    # Run RTC with Real robot with pi0.5 policy
    uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/pi05_check_rtc \
+        --policy.path=<USER>/pi05_check_rtc \
        --policy.device=mps \
        --rtc.enabled=true \
        --rtc.execution_horizon=20 \
--- a/examples/tutorial/rl/hilserl_example.py
+++ b/examples/tutorial/rl/hilserl_example.py
@@ -4,7 +4,6 @@ from pathlib import Path
 from queue import Empty, Full

 import torch
-import torch.optim as optim

 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
@@ -12,6 +11,7 @@ from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
 from lerobot.policies.sac.configuration_sac import SACConfig
 from lerobot.policies.sac.modeling_sac import SACPolicy
 from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
+from lerobot.rl.algorithms.sac import SACAlgorithm, SACAlgorithmConfig
 from lerobot.rl.buffer import ReplayBuffer
 from lerobot.rl.gym_manipulator import make_robot_env
 from lerobot.robots.so_follower import SO100FollowerConfig
@@ -40,8 +40,9 @@ def run_learner(
    policy_learner.train()
    policy_learner.to(device)

-    # Create Adam optimizer from scratch - simple and clean
-    optimizer = optim.Adam(policy_learner.parameters(), lr=lr)
+    algo_config = SACAlgorithmConfig.from_policy_config(policy_learner.config)
+    algorithm = SACAlgorithm(policy=policy_learner, config=algo_config)
+    algorithm.make_optimizers()

    print(f"[LEARNER] Online buffer capacity: {online_buffer.capacity}")
    print(f"[LEARNER] Offline buffer capacity: {offline_buffer.capacity}")
@@ -83,24 +84,26 @@ def run_learner(
                else:
                    batch[key] = online_batch[key]

-            loss, _ = policy_learner.forward(batch)
+            def batch_iter(b=batch):
+                while True:
+                    yield b

-            optimizer.zero_grad()
-            loss.backward()
-            optimizer.step()
+            stats = algorithm.update(batch_iter())
            training_step += 1

            if training_step % LOG_EVERY == 0:
+                log_dict = stats.to_log_dict()
                print(
-                    f"[LEARNER] Training step {training_step}, Loss: {loss.item():.4f}, "
+                    f"[LEARNER] Training step {training_step}, "
+                    f"critic_loss: {log_dict.get('critic', 'N/A'):.4f}, "
                    f"Buffers: Online={len(online_buffer)}, Offline={len(offline_buffer)}"
                )

            # Send updated parameters to actor every 10 training steps
            if training_step % SEND_EVERY == 0:
                try:
-                    state_dict = {k: v.cpu() for k, v in policy_learner.state_dict().items()}
-                    parameters_queue.put_nowait(state_dict)
+                    weights = algorithm.get_weights()
+                    parameters_queue.put_nowait(weights)
                    print("[LEARNER] Sent updated parameters to actor")
                except Full:
                    # Missing write due to queue not being consumed (should happen rarely)
@@ -144,15 +147,15 @@ def run_actor(

            while step < MAX_STEPS_PER_EPISODE and not shutdown_event.is_set():
                try:
-                    new_params = parameters_queue.get_nowait()
-                    policy_actor.load_state_dict(new_params)
+                    new_weights = parameters_queue.get_nowait()
+                    policy_actor.load_state_dict(new_weights)
                    print("[ACTOR] Updated policy parameters from learner")
                except Empty:  # No new updated parameters available from learner, waiting
                    pass

-                # Get action from policy
+                # Get action from policy (returns full action: continuous + discrete)
                policy_obs = make_policy_obs(obs, device=device)
-                action_tensor = policy_actor.select_action(policy_obs)  # predicts a single action
+                action_tensor = policy_actor.select_action(policy_obs)
                action = action_tensor.squeeze(0).cpu().numpy()

                # Step environment
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -76,9 +76,9 @@ dependencies = [
    "pyserial>=3.5,<4.0",
    "wandb>=0.24.0,<0.25.0",

-    "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
-    "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
-    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
+    "torch>=2.2.1,<2.11.0", # TODO: Bump dependency
+    "torchcodec>=0.2.1,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bump dependency
+    "torchvision>=0.21.0,<0.26.0", # TODO: Bump dependency

    "draccus==0.10.0", # TODO: Remove ==
    "gymnasium>=1.1.1,<2.0.0",
@@ -360,9 +360,9 @@ ignore_errors = false
 module = "lerobot.cameras.*"
 ignore_errors = false

-# [[tool.mypy.overrides]]
-# module = "lerobot.motors.*"
-# ignore_errors = false
+[[tool.mypy.overrides]]
+module = "lerobot.motors.*"
+ignore_errors = false

 # [[tool.mypy.overrides]]
 # module = "lerobot.robots.*"
--- a/src/lerobot/cameras/init.py
+++ b/src/lerobot/cameras/init.py
@@ -13,5 +13,5 @@
 # limitations under the License.

 from .camera import Camera
-from .configs import CameraConfig, ColorMode, Cv2Rotation
+from .configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
 from .utils import make_cameras_from_configs
--- a/src/lerobot/cameras/camera.py
+++ b/src/lerobot/cameras/camera.py
@@ -150,7 +150,7 @@ class Camera(abc.ABC):
        """
        pass

-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
        """Return the most recent frame captured immediately (Peeking).

        This method is non-blocking and returns whatever is currently in the
--- a/src/lerobot/cameras/configs.py
+++ b/src/lerobot/cameras/configs.py
@@ -25,6 +25,10 @@ class ColorMode(str, Enum):
    RGB = "rgb"
    BGR = "bgr"

+    @classmethod
+    def _missing_(cls, value: object) -> None:
+        raise ValueError(f"`color_mode` is expected to be in {list(cls)}, but {value} is provided.")
+

 class Cv2Rotation(int, Enum):
    NO_ROTATION = 0
@@ -32,6 +36,25 @@ class Cv2Rotation(int, Enum):
    ROTATE_180 = 180
    ROTATE_270 = -90

+    @classmethod
+    def _missing_(cls, value: object) -> None:
+        raise ValueError(f"`rotation` is expected to be in {list(cls)}, but {value} is provided.")
+
+
+# Subset from https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html
+class Cv2Backends(int, Enum):
+    ANY = 0
+    V4L2 = 200
+    DSHOW = 700
+    PVAPI = 800
+    ANDROID = 1000
+    AVFOUNDATION = 1200
+    MSMF = 1400
+
+    @classmethod
+    def _missing_(cls, value: object) -> None:
+        raise ValueError(f"`backend` is expected to be in {list(cls)}, but {value} is provided.")
+

@dataclass(kw_only=True)
 class CameraConfig(draccus.ChoiceRegistry, abc.ABC):  # type: ignore  # TODO: add type stubs for draccus
--- a/src/lerobot/cameras/opencv/camera_opencv.py
+++ b/src/lerobot/cameras/opencv/camera_opencv.py
@@ -32,10 +32,11 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
    os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV

-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError

 from ..camera import Camera
-from ..utils import get_cv2_backend, get_cv2_rotation
+from ..utils import get_cv2_rotation
 from .configuration_opencv import ColorMode, OpenCVCameraConfig

 # NOTE(Steven): The maximum opencv device index depends on your operating system. For instance,
@@ -117,7 +118,7 @@ class OpenCVCamera(Camera):
        self.new_frame_event: Event = Event()

        self.rotation: int | None = get_cv2_rotation(config.rotation)
-        self.backend: int = get_cv2_backend()
+        self.backend: int = config.backend

        if self.height and self.width:
            self.capture_width, self.capture_height = self.width, self.height
@@ -132,6 +133,7 @@ class OpenCVCamera(Camera):
        """Checks if the camera is currently connected and opened."""
        return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()

+    @check_if_already_connected
    def connect(self, warmup: bool = True) -> None:
        """
        Connects to the OpenCV camera specified in the configuration.
@@ -148,8 +150,6 @@ class OpenCVCamera(Camera):
            ConnectionError: If the specified camera index/path is not found or fails to open.
            RuntimeError: If the camera opens but fails to apply requested settings.
        """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} is already connected.")

        # Use 1 thread for OpenCV operations to avoid potential conflicts or
        # blocking in multi-threaded applications, especially during data collection.
@@ -178,6 +178,7 @@ class OpenCVCamera(Camera):

        logger.info(f"{self} connected.")

+    @check_if_not_connected
    def _configure_capture_settings(self) -> None:
        """
        Applies the specified FOURCC, FPS, width, and height settings to the connected camera.
@@ -197,8 +198,6 @@ class OpenCVCamera(Camera):
                          to the requested value.
            DeviceNotConnectedError: If the camera is not connected.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")

        # Set FOURCC first (if specified) as it can affect available FPS/resolution options
        if self.config.fourcc is not None:
@@ -348,6 +347,7 @@ class OpenCVCamera(Camera):

        return frame

+    @check_if_not_connected
    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
        """
        Reads a single frame synchronously from the camera.
@@ -374,9 +374,6 @@ class OpenCVCamera(Camera):
                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
            )

-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")

@@ -490,6 +487,7 @@ class OpenCVCamera(Camera):
            self.latest_timestamp = None
            self.new_frame_event.clear()

+    @check_if_not_connected
    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
        """
        Reads the latest available frame asynchronously.
@@ -512,8 +510,6 @@ class OpenCVCamera(Camera):
            TimeoutError: If no frame becomes available within the specified timeout.
            RuntimeError: If an unexpected error occurs.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")
@@ -533,7 +529,8 @@ class OpenCVCamera(Camera):

        return frame

-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
        """Return the most recent frame captured immediately (Peeking).

        This method is non-blocking and returns whatever is currently in the
@@ -548,8 +545,6 @@ class OpenCVCamera(Camera):
            DeviceNotConnectedError: If the camera is not connected.
            RuntimeError: If the camera is connected but has not captured any frames yet.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")
--- a/src/lerobot/cameras/opencv/configuration_opencv.py
+++ b/src/lerobot/cameras/opencv/configuration_opencv.py
@@ -15,9 +15,9 @@
 from dataclasses import dataclass
 from pathlib import Path

-from ..configs import CameraConfig, ColorMode, Cv2Rotation
+from ..configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation

-__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation"]
+__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation", "Cv2Backends"]


@CameraConfig.register_subclass("opencv")
@@ -50,6 +50,7 @@ class OpenCVCameraConfig(CameraConfig):
        rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
        warmup_s: Time reading frames before returning from connect (in seconds)
        fourcc: FOURCC code for video format (e.g., "MJPG", "YUYV", "I420"). Defaults to None (auto-detect).
+        backend: OpenCV backend identifier (https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html). Defaults to ANY.

    Note:
        - Only 3-channel color output (RGB/BGR) is currently supported.
@@ -62,22 +63,12 @@ class OpenCVCameraConfig(CameraConfig):
    rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
    warmup_s: int = 1
    fourcc: str | None = None
+    backend: Cv2Backends = Cv2Backends.ANY

    def __post_init__(self) -> None:
-        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
-            raise ValueError(
-                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
-            )
-
-        if self.rotation not in (
-            Cv2Rotation.NO_ROTATION,
-            Cv2Rotation.ROTATE_90,
-            Cv2Rotation.ROTATE_180,
-            Cv2Rotation.ROTATE_270,
-        ):
-            raise ValueError(
-                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
-            )
+        self.color_mode = ColorMode(self.color_mode)
+        self.rotation = Cv2Rotation(self.rotation)
+        self.backend = Cv2Backends(self.backend)

        if self.fourcc is not None and (not isinstance(self.fourcc, str) or len(self.fourcc) != 4):
            raise ValueError(
--- a/src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py
+++ b/src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py
@@ -74,7 +74,4 @@ class Reachy2CameraConfig(CameraConfig):
                f"`image_type` is expected to be 'left' or 'right' for teleop camera, and 'rgb' or 'depth' for depth camera, but {self.image_type} is provided."
            )

-        if self.color_mode not in ["rgb", "bgr"]:
-            raise ValueError(
-                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
-            )
+        self.color_mode = ColorMode(self.color_mode)
--- a/src/lerobot/cameras/reachy2_camera/reachy2_camera.py
+++ b/src/lerobot/cameras/reachy2_camera/reachy2_camera.py
@@ -32,6 +32,7 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy

+from lerobot.utils.decorators import check_if_not_connected
 from lerobot.utils.import_utils import _reachy2_sdk_available

 if TYPE_CHECKING or _reachy2_sdk_available:
@@ -123,6 +124,7 @@ class Reachy2Camera(Camera):
        """
        raise NotImplementedError("Camera detection is not implemented for Reachy2 cameras.")

+    @check_if_not_connected
    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
        """
        Reads a single frame synchronously from the camera.
@@ -136,9 +138,6 @@ class Reachy2Camera(Camera):
        """
        start_time = time.perf_counter()

-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
        if self.cam_manager is None:
            raise DeviceNotConnectedError(f"{self} is not connected.")

@@ -184,6 +183,7 @@ class Reachy2Camera(Camera):

        return frame

+    @check_if_not_connected
    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
        """
        Same as read()
@@ -197,12 +197,11 @@ class Reachy2Camera(Camera):
            TimeoutError: If no frame becomes available within the specified timeout.
            RuntimeError: If an unexpected error occurs.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        return self.read()

-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
        """Return the most recent frame captured immediately (Peeking).

        This method is non-blocking and returns whatever is currently in the
@@ -219,8 +218,6 @@ class Reachy2Camera(Camera):
            DeviceNotConnectedError: If the camera is not connected.
            RuntimeError: If the camera is connected but has not captured any frames yet.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        if self.latest_frame is None or self.latest_timestamp is None:
            raise RuntimeError(f"{self} has not captured any frames yet.")
@@ -233,6 +230,7 @@ class Reachy2Camera(Camera):

        return self.latest_frame

+    @check_if_not_connected
    def disconnect(self) -> None:
        """
        Stops the background read thread (if running).
@@ -240,8 +238,6 @@ class Reachy2Camera(Camera):
        Raises:
            DeviceNotConnectedError: If the camera is already disconnected.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} not connected.")

        if self.cam_manager is not None:
            self.cam_manager.disconnect()
--- a/src/lerobot/cameras/realsense/camera_realsense.py
+++ b/src/lerobot/cameras/realsense/camera_realsense.py
@@ -30,7 +30,8 @@ try:
 except Exception as e:
    logging.info(f"Could not import realsense: {e}")

-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError

 from ..camera import Camera
 from ..configs import ColorMode
@@ -152,6 +153,7 @@ class RealSenseCamera(Camera):
        """Checks if the camera pipeline is started and streams are active."""
        return self.rs_pipeline is not None and self.rs_profile is not None

+    @check_if_already_connected
    def connect(self, warmup: bool = True) -> None:
        """
        Connects to the RealSense camera specified in the configuration.
@@ -169,8 +171,6 @@ class RealSenseCamera(Camera):
            ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
            RuntimeError: If the pipeline starts but fails to apply requested settings.
        """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} is already connected.")

        self.rs_pipeline = rs.pipeline()
        rs_config = rs.config()
@@ -290,6 +290,7 @@ class RealSenseCamera(Camera):
            if self.use_depth:
                rs_config.enable_stream(rs.stream.depth)

+    @check_if_not_connected
    def _configure_capture_settings(self) -> None:
        """Sets fps, width, and height from device stream if not already configured.

@@ -299,8 +300,6 @@ class RealSenseCamera(Camera):
        Raises:
            DeviceNotConnectedError: If device is not connected.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")

        if self.rs_profile is None:
            raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
@@ -320,6 +319,7 @@ class RealSenseCamera(Camera):
                self.width, self.height = actual_width, actual_height
                self.capture_width, self.capture_height = actual_width, actual_height

+    @check_if_not_connected
    def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
        """
        Reads a single frame (depth) synchronously from the camera.
@@ -345,9 +345,6 @@ class RealSenseCamera(Camera):
                f"Failed to capture depth frame '.read_depth()'. Depth stream is not enabled for {self}."
            )

-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")

@@ -374,6 +371,7 @@ class RealSenseCamera(Camera):

        return frame

+    @check_if_not_connected
    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 0) -> NDArray[Any]:
        """
        Reads a single frame (color) synchronously from the camera.
@@ -403,9 +401,6 @@ class RealSenseCamera(Camera):
                f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
            )

-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")

@@ -534,6 +529,7 @@ class RealSenseCamera(Camera):
            self.new_frame_event.clear()

    # NOTE(Steven): Missing implementation for depth for now
+    @check_if_not_connected
    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
        """
        Reads the latest available frame data (color) asynchronously.
@@ -556,8 +552,6 @@ class RealSenseCamera(Camera):
            TimeoutError: If no frame data becomes available within the specified timeout.
            RuntimeError: If the background thread died unexpectedly or another error occurs.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")
@@ -578,7 +572,8 @@ class RealSenseCamera(Camera):
        return frame

    # NOTE(Steven): Missing implementation for depth for now
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
        """Return the most recent (color) frame captured immediately (Peeking).

        This method is non-blocking and returns whatever is currently in the
@@ -593,8 +588,6 @@ class RealSenseCamera(Camera):
            DeviceNotConnectedError: If the camera is not connected.
            RuntimeError: If the camera is connected but has not captured any frames yet.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")
--- a/src/lerobot/cameras/realsense/configuration_realsense.py
+++ b/src/lerobot/cameras/realsense/configuration_realsense.py
@@ -60,20 +60,8 @@ class RealSenseCameraConfig(CameraConfig):
    warmup_s: int = 1

    def __post_init__(self) -> None:
-        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
-            raise ValueError(
-                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
-            )
-
-        if self.rotation not in (
-            Cv2Rotation.NO_ROTATION,
-            Cv2Rotation.ROTATE_90,
-            Cv2Rotation.ROTATE_180,
-            Cv2Rotation.ROTATE_270,
-        ):
-            raise ValueError(
-                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
-            )
+        self.color_mode = ColorMode(self.color_mode)
+        self.rotation = Cv2Rotation(self.rotation)

        values = (self.fps, self.width, self.height)
        if any(v is not None for v in values) and any(v is None for v in values):
--- a/src/lerobot/cameras/utils.py
+++ b/src/lerobot/cameras/utils.py
@@ -14,7 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-import platform
 from typing import cast

 from lerobot.utils.import_utils import make_device_from_device_class
@@ -68,14 +67,3 @@ def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
        return int(cv2.ROTATE_90_COUNTERCLOCKWISE)
    else:
        return None
-
-
-def get_cv2_backend() -> int:
-    import cv2
-
-    if platform.system() == "Windows":
-        return int(cv2.CAP_MSMF)  # Use MSMF for Windows instead of AVFOUNDATION
-    # elif platform.system() == "Darwin":  # macOS
-    #     return cv2.CAP_AVFOUNDATION
-    else:  # Linux and others
-        return int(cv2.CAP_ANY)
--- a/src/lerobot/cameras/zmq/camera_zmq.py
+++ b/src/lerobot/cameras/zmq/camera_zmq.py
@@ -34,7 +34,8 @@ import cv2
 import numpy as np
 from numpy.typing import NDArray

-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError

 from ..camera import Camera
 from ..configs import ColorMode
@@ -104,6 +105,7 @@ class ZMQCamera(Camera):
        """Checks if the ZMQ socket is initialized and connected."""
        return self._connected and self.context is not None and self.socket is not None

+    @check_if_already_connected
    def connect(self, warmup: bool = True) -> None:
        """Connect to ZMQ camera server.

@@ -111,8 +113,6 @@ class ZMQCamera(Camera):
            warmup (bool): If True, waits for the camera to provide at least one
                           valid frame before returning. Defaults to True.
        """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} is already connected.")

        logger.info(f"Connecting to {self}...")

@@ -211,6 +211,7 @@ class ZMQCamera(Camera):

        return frame

+    @check_if_not_connected
    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
        """
        Reads a single frame synchronously from the camera.
@@ -228,9 +229,6 @@ class ZMQCamera(Camera):
                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
            )

-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")

@@ -301,6 +299,7 @@ class ZMQCamera(Camera):
            self.latest_timestamp = None
            self.new_frame_event.clear()

+    @check_if_not_connected
    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
        """
        Reads the latest available frame asynchronously.
@@ -317,8 +316,6 @@ class ZMQCamera(Camera):
            TimeoutError: If no frame data becomes available within the specified timeout.
            RuntimeError: If the background thread is not running.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")
@@ -335,6 +332,7 @@ class ZMQCamera(Camera):

        return frame

+    @check_if_not_connected
    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
        """Return the most recent frame captured immediately (Peeking).

@@ -350,8 +348,6 @@ class ZMQCamera(Camera):
            DeviceNotConnectedError: If the camera is not connected.
            RuntimeError: If the camera is connected but has not captured any frames yet.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")
--- a/src/lerobot/cameras/zmq/configuration_zmq.py
+++ b/src/lerobot/cameras/zmq/configuration_zmq.py
@@ -32,10 +32,7 @@ class ZMQCameraConfig(CameraConfig):
    warmup_s: int = 1

    def __post_init__(self) -> None:
-        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
-            raise ValueError(
-                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
-            )
+        self.color_mode = ColorMode(self.color_mode)

        if self.timeout_ms <= 0:
            raise ValueError(f"`timeout_ms` must be positive, but {self.timeout_ms} is provided.")
--- a/src/lerobot/configs/policies.py
+++ b/src/lerobot/configs/policies.py
@@ -45,12 +45,12 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
    Args:
        n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
            current step and additional steps going back).
-        input_shapes: A dictionary defining the shapes of the input data for the policy.
-        output_shapes: A dictionary defining the shapes of the output data for the policy.
-        input_normalization_modes: A dictionary with key representing the modality and the value specifies the
-            normalization mode to apply.
-        output_normalization_modes: Similar dictionary as `input_normalization_modes`, but to unnormalize to
-            the original scale.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
    """

    n_obs_steps: int = 1
--- a/src/lerobot/configs/train.py
+++ b/src/lerobot/configs/train.py
@@ -211,3 +211,15 @@ class TrainRLServerPipelineConfig(TrainPipelineConfig):
    # NOTE: In RL, we don't need an offline dataset
    # TODO: Make `TrainPipelineConfig.dataset` optional
    dataset: DatasetConfig | None = None  # type: ignore[assignment] # because the parent class has made it's type non-optional
+
+    # Algorithm name registered in RLAlgorithmConfig registry
+    algorithm: str = "sac"
+
+    # Data mixer strategy name. Currently supports "online_offline"
+    mixer: str = "online_offline"
+    # Fraction sampled from online replay when using OnlineOfflineMixer
+    online_ratio: float = 0.5
+
+    # RL trainer iterator
+    async_prefetch: bool = True
+    queue_size: int = 2
--- a/src/lerobot/datasets/lerobot_dataset.py
+++ b/src/lerobot/datasets/lerobot_dataset.py
@@ -656,7 +656,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
            repo_id (str): This is the repo id that will be used to fetch the dataset. Locally, the dataset
                will be stored under root/repo_id.
            root (Path | None, optional): Local directory to use for downloading/writing files. You can also
-                set the LEROBOT_HOME environment variable to point to a different location. Defaults to
+                set the HF_LEROBOT_HOME environment variable to point to a different location. Defaults to
                '~/.cache/huggingface/lerobot'.
            episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                their episode_index in this list. Defaults to None.
--- a/src/lerobot/datasets/transforms.py
+++ b/src/lerobot/datasets/transforms.py
@@ -216,16 +216,17 @@ class ImageTransformsConfig:


 def make_transform_from_config(cfg: ImageTransformConfig):
-    if cfg.type == "Identity":
-        return v2.Identity(**cfg.kwargs)
-    elif cfg.type == "ColorJitter":
-        return v2.ColorJitter(**cfg.kwargs)
-    elif cfg.type == "SharpnessJitter":
+    if cfg.type == "SharpnessJitter":
        return SharpnessJitter(**cfg.kwargs)
-    elif cfg.type == "RandomAffine":
-        return v2.RandomAffine(**cfg.kwargs)
-    else:
-        raise ValueError(f"Transform '{cfg.type}' is not valid.")
+
+    transform_cls = getattr(v2, cfg.type, None)
+    if isinstance(transform_cls, type) and issubclass(transform_cls, Transform):
+        return transform_cls(**cfg.kwargs)
+
+    raise ValueError(
+        f"Transform '{cfg.type}' is not valid. It must be a class in "
+        f"torchvision.transforms.v2 or 'SharpnessJitter'."
+    )


 class ImageTransforms(Transform):
--- a/src/lerobot/datasets/utils.py
+++ b/src/lerobot/datasets/utils.py
@@ -122,19 +122,9 @@ def load_nested_dataset(
        raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")

    with SuppressProgressBars():
-        # When no filtering needed, Dataset uses memory-mapped loading for efficiency
-        # PyArrow loads the entire dataset into memory
-        if episodes is None:
-            return Dataset.from_parquet([str(path) for path in paths], features=features)
-
-        arrow_dataset = pa_ds.dataset(paths, format="parquet")
-        filter_expr = pa_ds.field("episode_index").isin(episodes)
-        table = arrow_dataset.to_table(filter=filter_expr)
-
-        if features is not None:
-            table = table.cast(features.arrow_schema)
-
-        return Dataset(table)
+        # We use .from_parquet() memory-mapped loading for efficiency
+        filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
+        return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)


 def get_parquet_num_frames(parquet_path: str | Path) -> int:
--- a/src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py
+++ b/src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py
@@ -529,7 +529,7 @@ if __name__ == "__main__":
        type=str,
        required=True,
        help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset "
-        "(e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
+        "(e.g. `lerobot/pusht`, `<USER>/aloha_sim_insertion_human`).",
    )
    parser.add_argument(
        "--branch",
--- a/src/lerobot/envs/configs.py
+++ b/src/lerobot/envs/configs.py
@@ -205,6 +205,7 @@ class ObservationConfig:

    add_joint_velocity_to_observation: bool = False
    add_current_to_observation: bool = False
+    add_ee_pose_to_observation: bool = False
    display_cameras: bool = False


--- a/src/lerobot/envs/libero.py
+++ b/src/lerobot/envs/libero.py
@@ -112,6 +112,7 @@ class LiberoEnv(gym.Env):
        visualization_height: int = 480,
        init_states: bool = True,
        episode_index: int = 0,
+        n_envs: int = 1,
        camera_name_mapping: dict[str, str] | None = None,
        num_steps_wait: int = 10,
        control_mode: str = "relative",
@@ -145,7 +146,9 @@ class LiberoEnv(gym.Env):
        self.episode_length = episode_length
        # 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._reset_stride = n_envs  # when performing a reset, append `_reset_stride` to `init_state_id`.
+
+        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
@@ -295,7 +298,8 @@ class LiberoEnv(gym.Env):
        self._env.seed(seed)
        raw_obs = self._env.reset()
        if self.init_states and self._init_states is not None:
-            raw_obs = self._env.set_init_state(self._init_states[self._init_state_id])
+            raw_obs = self._env.set_init_state(self._init_states[self.init_state_id % len(self._init_states)])
+            self.init_state_id += self._reset_stride  # Change init_state_id when 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.
@@ -373,6 +377,7 @@ def _make_env_fns(
            init_states=init_states,
            episode_length=episode_length,
            episode_index=episode_index,
+            n_envs=n_envs,
            control_mode=control_mode,
            **local_kwargs,
        )
--- a/src/lerobot/motors/calibration_gui.py
+++ b/src/lerobot/motors/calibration_gui.py
@@ -221,7 +221,7 @@ class RangeFinderGUI:

        self.bus = bus
        self.groups = groups if groups is not None else {"all": list(bus.motors)}
-        self.group_names = list(groups)
+        self.group_names = list(self.groups)
        self.current_group = self.group_names[0]

        if not bus.is_connected:
@@ -230,18 +230,20 @@ class RangeFinderGUI:
        self.calibration = bus.read_calibration()
        self.res_table = bus.model_resolution_table
        self.present_cache = {
-            m: bus.read("Present_Position", m, normalize=False) for motors in groups.values() for m in motors
+            m: bus.read("Present_Position", m, normalize=False)
+            for motors in self.groups.values()
+            for m in motors
        }

        pygame.init()
        self.font = pygame.font.Font(None, FONT_SIZE)

-        label_pad = max(self.font.size(m)[0] for ms in groups.values() for m in ms)
+        label_pad = max(self.font.size(m)[0] for ms in self.groups.values() for m in ms)
        self.label_pad = label_pad
        width = 40 + label_pad + BAR_LEN + 6 + BTN_W + 10 + SAVE_W + 10
        self.controls_bottom = 10 + SAVE_H
        self.base_y = self.controls_bottom + TOP_GAP
-        height = self.base_y + PADDING_Y * len(groups[self.current_group]) + 40
+        height = self.base_y + PADDING_Y * len(self.groups[self.current_group]) + 40

        self.screen = pygame.display.set_mode((width, height))
        pygame.display.set_caption("Motors range finder")
--- a/src/lerobot/motors/damiao/damiao.py
+++ b/src/lerobot/motors/damiao/damiao.py
@@ -23,6 +23,7 @@ from copy import deepcopy
 from functools import cached_property
 from typing import TYPE_CHECKING, Any, TypedDict

+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.import_utils import _can_available

 if TYPE_CHECKING or _can_available:
@@ -36,7 +37,6 @@ else:

 import numpy as np

-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import enter_pressed, move_cursor_up

@@ -155,6 +155,7 @@ class DamiaoMotorsBus(MotorsBusBase):
        """Check if the CAN bus is connected."""
        return self._is_connected and self.canbus is not None

+    @check_if_already_connected
    def connect(self, handshake: bool = True) -> None:
        """
        Open the CAN bus and initialize communication.
@@ -162,10 +163,6 @@ class DamiaoMotorsBus(MotorsBusBase):
        Args:
            handshake: If True, ping all motors to verify they're present
        """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(
-                f"{self.__class__.__name__}('{self.port}') is already connected."
-            )

        try:
            # Auto-detect interface type based on port name
@@ -211,6 +208,9 @@ class DamiaoMotorsBus(MotorsBusBase):
        logger.info("Starting handshake with motors...")

        # Drain any pending messages
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        while self.canbus.recv(timeout=0.01):
            pass

@@ -246,6 +246,7 @@ class DamiaoMotorsBus(MotorsBusBase):
            )
        logger.info("Handshake successful. All motors ready.")

+    @check_if_not_connected
    def disconnect(self, disable_torque: bool = True) -> None:
        """
        Close the CAN bus connection.
@@ -253,8 +254,6 @@ class DamiaoMotorsBus(MotorsBusBase):
        Args:
            disable_torque: If True, disable torque on all motors before disconnecting
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self.__class__.__name__}('{self.port}') is not connected.")

        if disable_torque:
            try:
@@ -283,6 +282,10 @@ class DamiaoMotorsBus(MotorsBusBase):
        recv_id = self._get_motor_recv_id(motor)
        data = [0xFF] * 7 + [command_byte]
        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        self.canbus.send(msg)
        if msg := self._recv_motor_response(expected_recv_id=recv_id):
            self._process_response(motor_name, msg)
@@ -341,6 +344,10 @@ class DamiaoMotorsBus(MotorsBusBase):
        recv_id = self._get_motor_recv_id(motor)
        data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
        msg = can.Message(arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        self.canbus.send(msg)
        return self._recv_motor_response(expected_recv_id=recv_id)

@@ -356,6 +363,10 @@ class DamiaoMotorsBus(MotorsBusBase):
        Returns:
            CAN message if received, None otherwise
        """
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        try:
            start_time = time.time()
            messages_seen = []
@@ -394,10 +405,13 @@ class DamiaoMotorsBus(MotorsBusBase):
        Returns:
            Dictionary mapping recv_id to CAN message
        """
-        responses = {}
+        responses: dict[int, can.Message] = {}
        expected_set = set(expected_recv_ids)
        start_time = time.time()

+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        try:
            while len(responses) < len(expected_recv_ids) and (time.time() - start_time) < timeout:
                # 100us poll timeout
@@ -461,6 +475,9 @@ class DamiaoMotorsBus(MotorsBusBase):
        motor_name = self._get_motor_name(motor)
        motor_type = self._motor_types[motor_name]

+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        data = self._encode_mit_packet(motor_type, kp, kd, position_degrees, velocity_deg_per_sec, torque)
        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
        self.canbus.send(msg)
@@ -488,6 +505,9 @@ class DamiaoMotorsBus(MotorsBusBase):

        recv_id_to_motor: dict[int, str] = {}

+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        # Step 1: Send all MIT control commands
        for motor, (kp, kd, position_degrees, velocity_deg_per_sec, torque) in commands.items():
            motor_id = self._get_motor_id(motor)
@@ -562,10 +582,9 @@ class DamiaoMotorsBus(MotorsBusBase):
        except Exception as e:
            logger.warning(f"Failed to decode response from {motor}: {e}")

+    @check_if_not_connected
    def read(self, data_name: str, motor: str) -> Value:
        """Read a value from a single motor. Positions are always in degrees."""
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        # Refresh motor to get latest state
        msg = self._refresh_motor(motor)
@@ -595,6 +614,7 @@ class DamiaoMotorsBus(MotorsBusBase):
            raise ValueError(f"Unknown data_name: {data_name}")
        return mapping[data_name]

+    @check_if_not_connected
    def write(
        self,
        data_name: str,
@@ -605,8 +625,6 @@ class DamiaoMotorsBus(MotorsBusBase):
        Write a value to a single motor. Positions are always in degrees.
        Can write 'Goal_Position', 'Kp', or 'Kd'.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        if data_name in ("Kp", "Kd"):
            self._gains[motor][data_name.lower()] = float(value)
@@ -656,6 +674,10 @@ class DamiaoMotorsBus(MotorsBusBase):

    def _batch_refresh(self, motors: list[str]) -> None:
        """Internal helper to refresh a list of motors and update cache."""
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        # Send refresh commands
        for motor in motors:
            motor_id = self._get_motor_id(motor)
@@ -678,10 +700,12 @@ class DamiaoMotorsBus(MotorsBusBase):
            else:
                logger.warning(f"Packet drop: {motor} (ID: 0x{recv_id:02X}). Using last known state.")

-    def sync_write(self, data_name: str, values: Value | dict[str, Value]) -> None:
+    @check_if_not_connected
+    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
        """
        Write values to multiple motors simultaneously. Positions are always in degrees.
        """
+
        if data_name in ("Kp", "Kd"):
            key = data_name.lower()
            for motor, val in values.items():
@@ -690,6 +714,8 @@ class DamiaoMotorsBus(MotorsBusBase):
        elif data_name == "Goal_Position":
            # Step 1: Send all MIT control commands
            recv_id_to_motor: dict[int, str] = {}
+            if self.canbus is None:
+                raise RuntimeError("CAN bus is not initialized.")
            for motor, value_degrees in values.items():
                motor_id = self._get_motor_id(motor)
                motor_name = self._get_motor_name(motor)
@@ -732,9 +758,9 @@ class DamiaoMotorsBus(MotorsBusBase):

    def record_ranges_of_motion(
        self,
-        motors: NameOrID | list[NameOrID] | None = None,
+        motors: str | list[str] | None = None,
        display_values: bool = True,
-    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
+    ) -> tuple[dict[str, Value], dict[str, Value]]:
        """
        Interactively record the min/max values of each motor in degrees.

--- a/src/lerobot/motors/dynamixel/dynamixel.py
+++ b/src/lerobot/motors/dynamixel/dynamixel.py
@@ -181,10 +181,10 @@ class DynamixelMotorsBus(SerialMotorsBus):
        for motor, m in self.motors.items():
            calibration[motor] = MotorCalibration(
                id=m.id,
-                drive_mode=drive_modes[motor],
-                homing_offset=offsets[motor],
-                range_min=mins[motor],
-                range_max=maxes[motor],
+                drive_mode=int(drive_modes[motor]),
+                homing_offset=int(offsets[motor]),
+                range_min=int(mins[motor]),
+                range_max=int(maxes[motor]),
            )

        return calibration
@@ -198,7 +198,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
        if cache:
            self.calibration = calibration_dict

-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
        for motor in self._get_motors_list(motors):
            self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)

@@ -206,7 +206,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
        addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
        self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)

-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
        for motor in self._get_motors_list(motors):
            self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)

@@ -235,7 +235,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
        On Dynamixel Motors:
        Present_Position = Actual_Position + Homing_Offset
        """
-        half_turn_homings = {}
+        half_turn_homings: dict[NameOrID, Value] = {}
        for motor, pos in positions.items():
            model = self._get_motor_model(motor)
            max_res = self.model_resolution_table[model] - 1
@@ -258,6 +258,6 @@ class DynamixelMotorsBus(SerialMotorsBus):
            if raise_on_error:
                raise ConnectionError(self.packet_handler.getTxRxResult(comm))

-            return
+            return None

        return {id_: data[0] for id_, data in data_list.items()}
--- a/src/lerobot/motors/feetech/feetech.py
+++ b/src/lerobot/motors/feetech/feetech.py
@@ -126,7 +126,7 @@ class FeetechMotorsBus(SerialMotorsBus):

        self.port_handler = scs.PortHandler(self.port)
        # HACK: monkeypatch
-        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(
+        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(  # type: ignore[method-assign]
            self.port_handler, scs.PortHandler
        )
        self.packet_handler = scs.PacketHandler(protocol_version)
@@ -262,9 +262,9 @@ class FeetechMotorsBus(SerialMotorsBus):
            calibration[motor] = MotorCalibration(
                id=m.id,
                drive_mode=0,
-                homing_offset=offsets[motor],
-                range_min=mins[motor],
-                range_max=maxes[motor],
+                homing_offset=int(offsets[motor]),
+                range_min=int(mins[motor]),
+                range_max=int(maxes[motor]),
            )

        return calibration
@@ -284,7 +284,7 @@ class FeetechMotorsBus(SerialMotorsBus):
        On Feetech Motors:
        Present_Position = Actual_Position - Homing_Offset
        """
-        half_turn_homings = {}
+        half_turn_homings: dict[NameOrID, Value] = {}
        for motor, pos in positions.items():
            model = self._get_motor_model(motor)
            max_res = self.model_resolution_table[model] - 1
@@ -292,7 +292,7 @@ class FeetechMotorsBus(SerialMotorsBus):

        return half_turn_homings

-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
        for motor in self._get_motors_list(motors):
            self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
            self.write("Lock", motor, 0, num_retry=num_retry)
@@ -303,7 +303,7 @@ class FeetechMotorsBus(SerialMotorsBus):
        addr, length = get_address(self.model_ctrl_table, model, "Lock")
        self._write(addr, length, motor, 0, num_retry=num_retry)

-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
        for motor in self._get_motors_list(motors):
            self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
            self.write("Lock", motor, 1, num_retry=num_retry)
@@ -334,7 +334,7 @@ class FeetechMotorsBus(SerialMotorsBus):
    def _broadcast_ping(self) -> tuple[dict[int, int], int]:
        import scservo_sdk as scs

-        data_list = {}
+        data_list: dict[int, int] = {}

        status_length = 6

@@ -414,7 +414,7 @@ class FeetechMotorsBus(SerialMotorsBus):
        if not self._is_comm_success(comm):
            if raise_on_error:
                raise ConnectionError(self.packet_handler.getTxRxResult(comm))
-            return
+            return None

        ids_errors = {id_: status for id_, status in ids_status.items() if self._is_error(status)}
        if ids_errors:
--- a/src/lerobot/motors/motors_bus.py
+++ b/src/lerobot/motors/motors_bus.py
@@ -23,6 +23,7 @@ from __future__ import annotations

 import abc
 import logging
+from collections.abc import Sequence
 from contextlib import contextmanager
 from dataclasses import dataclass
 from enum import Enum
@@ -93,7 +94,7 @@ class MotorsBusBase(abc.ABC):
        pass

    @abc.abstractmethod
-    def sync_write(self, data_name: str, values: Value | dict[str, Value]) -> None:
+    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
        """Write values to multiple motors."""
        pass

@@ -179,15 +180,16 @@ class Motor:


 class PortHandler(Protocol):
-    def __init__(self, port_name):
-        self.is_open: bool
-        self.baudrate: int
-        self.packet_start_time: float
-        self.packet_timeout: float
-        self.tx_time_per_byte: float
-        self.is_using: bool
-        self.port_name: str
-        self.ser: serial.Serial
+    is_open: bool
+    baudrate: int
+    packet_start_time: float
+    packet_timeout: float
+    tx_time_per_byte: float
+    is_using: bool
+    port_name: str
+    ser: serial.Serial
+
+    def __init__(self, port_name: str) -> None: ...

    def openPort(self): ...
    def closePort(self): ...
@@ -240,19 +242,22 @@ class PacketHandler(Protocol):
    def regWriteTxRx(self, port, id, address, length, data): ...
    def syncReadTx(self, port, start_address, data_length, param, param_length): ...
    def syncWriteTxOnly(self, port, start_address, data_length, param, param_length): ...
+    def broadcastPing(self, port): ...


 class GroupSyncRead(Protocol):
-    def __init__(self, port, ph, start_address, data_length):
-        self.port: str
-        self.ph: PortHandler
-        self.start_address: int
-        self.data_length: int
-        self.last_result: bool
-        self.is_param_changed: bool
-        self.param: list
-        self.data_dict: dict
+    port: str
+    ph: PortHandler
+    start_address: int
+    data_length: int
+    last_result: bool
+    is_param_changed: bool
+    param: list
+    data_dict: dict

+    def __init__(
+        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
+    ) -> None: ...
    def makeParam(self): ...
    def addParam(self, id): ...
    def removeParam(self, id): ...
@@ -265,15 +270,17 @@ class GroupSyncRead(Protocol):


 class GroupSyncWrite(Protocol):
-    def __init__(self, port, ph, start_address, data_length):
-        self.port: str
-        self.ph: PortHandler
-        self.start_address: int
-        self.data_length: int
-        self.is_param_changed: bool
-        self.param: list
-        self.data_dict: dict
+    port: str
+    ph: PortHandler
+    start_address: int
+    data_length: int
+    is_param_changed: bool
+    param: list
+    data_dict: dict

+    def __init__(
+        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
+    ) -> None: ...
    def makeParam(self): ...
    def addParam(self, id, data): ...
    def removeParam(self, id): ...
@@ -400,7 +407,7 @@ class SerialMotorsBus(MotorsBusBase):
        else:
            raise TypeError(f"'{motor}' should be int, str.")

-    def _get_motor_model(self, motor: NameOrID) -> int:
+    def _get_motor_model(self, motor: NameOrID) -> str:
        if isinstance(motor, str):
            return self.motors[motor].model
        elif isinstance(motor, int):
@@ -408,17 +415,19 @@ class SerialMotorsBus(MotorsBusBase):
        else:
            raise TypeError(f"'{motor}' should be int, str.")

-    def _get_motors_list(self, motors: str | list[str] | None) -> list[str]:
+    def _get_motors_list(self, motors: NameOrID | Sequence[NameOrID] | None) -> list[str]:
        if motors is None:
            return list(self.motors)
        elif isinstance(motors, str):
            return [motors]
-        elif isinstance(motors, list):
-            return motors.copy()
+        elif isinstance(motors, int):
+            return [self._id_to_name(motors)]
+        elif isinstance(motors, Sequence):
+            return [m if isinstance(m, str) else self._id_to_name(m) for m in motors]
        else:
            raise TypeError(motors)

-    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> list[str]:
+    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> dict[int, Value]:
        if isinstance(values, (int | float)):
            return dict.fromkeys(self.ids, values)
        elif isinstance(values, dict):
@@ -640,18 +649,19 @@ class SerialMotorsBus(MotorsBusBase):
        pass

    @abc.abstractmethod
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
        """Enable torque on selected motors.

        Args:
-            motor (int): Same semantics as :pymeth:`disable_torque`. Defaults to `None`.
+            motors (int | str | list[str] | None, optional): Same semantics as :pymeth:`disable_torque`.
+                Defaults to `None`.
            num_retry (int, optional): Number of additional retry attempts on communication failure.
                Defaults to 0.
        """
        pass

    @contextmanager
-    def torque_disabled(self, motors: int | str | list[str] | None = None):
+    def torque_disabled(self, motors: str | list[str] | None = None):
        """Context-manager that guarantees torque is re-enabled.

        This helper is useful to temporarily disable torque when configuring motors.
@@ -728,24 +738,19 @@ class SerialMotorsBus(MotorsBusBase):
        """
        pass

-    def reset_calibration(self, motors: NameOrID | list[NameOrID] | None = None) -> None:
+    def reset_calibration(self, motors: NameOrID | Sequence[NameOrID] | None = None) -> None:
        """Restore factory calibration for the selected motors.

        Homing offset is set to ``0`` and min/max position limits are set to the full usable range.
        The in-memory :pyattr:`calibration` is cleared.

        Args:
-            motors (NameOrID | list[NameOrID] | None, optional): Selection of motors. `None` (default)
+            motors (NameOrID | Sequence[NameOrID] | None, optional): Selection of motors. `None` (default)
                resets every motor.
        """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)

-        for motor in motors:
+        for motor in motor_names:
            model = self._get_motor_model(motor)
            max_res = self.model_resolution_table[model] - 1
            self.write("Homing_Offset", motor, 0, normalize=False)
@@ -754,7 +759,9 @@ class SerialMotorsBus(MotorsBusBase):

        self.calibration = {}

-    def set_half_turn_homings(self, motors: NameOrID | list[NameOrID] | None = None) -> dict[NameOrID, Value]:
+    def set_half_turn_homings(
+        self, motors: NameOrID | Sequence[NameOrID] | None = None
+    ) -> dict[NameOrID, Value]:
        """Centre each motor range around its current position.

        The function computes and writes a homing offset such that the present position becomes exactly one
@@ -764,17 +771,12 @@ class SerialMotorsBus(MotorsBusBase):
            motors (NameOrID | list[NameOrID] | None, optional): Motors to adjust. Defaults to all motors (`None`).

        Returns:
-            dict[NameOrID, Value]: Mapping *motor → written homing offset*.
+            dict[str, Value]: Mapping *motor name → written homing offset*.
        """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)

-        self.reset_calibration(motors)
-        actual_positions = self.sync_read("Present_Position", motors, normalize=False)
+        self.reset_calibration(motor_names)
+        actual_positions = self.sync_read("Present_Position", motor_names, normalize=False)
        homing_offsets = self._get_half_turn_homings(actual_positions)
        for motor, offset in homing_offsets.items():
            self.write("Homing_Offset", motor, offset)
@@ -786,8 +788,8 @@ class SerialMotorsBus(MotorsBusBase):
        pass

    def record_ranges_of_motion(
-        self, motors: NameOrID | list[NameOrID] | None = None, display_values: bool = True
-    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
+        self, motors: NameOrID | Sequence[NameOrID] | None = None, display_values: bool = True
+    ) -> tuple[dict[str, Value], dict[str, Value]]:
        """Interactively record the min/max encoder values of each motor.

        Move the joints by hand (with torque disabled) while the method streams live positions. Press
@@ -799,30 +801,25 @@ class SerialMotorsBus(MotorsBusBase):
            display_values (bool, optional): When `True` (default) a live table is printed to the console.

        Returns:
-            tuple[dict[NameOrID, Value], dict[NameOrID, Value]]: Two dictionaries *mins* and *maxes* with the
+            tuple[dict[str, Value], dict[str, Value]]: Two dictionaries *mins* and *maxes* with the
                extreme values observed for each motor.
        """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)

-        start_positions = self.sync_read("Present_Position", motors, normalize=False)
+        start_positions = self.sync_read("Present_Position", motor_names, normalize=False)
        mins = start_positions.copy()
        maxes = start_positions.copy()

        user_pressed_enter = False
        while not user_pressed_enter:
-            positions = self.sync_read("Present_Position", motors, normalize=False)
+            positions = self.sync_read("Present_Position", motor_names, normalize=False)
            mins = {motor: min(positions[motor], min_) for motor, min_ in mins.items()}
            maxes = {motor: max(positions[motor], max_) for motor, max_ in maxes.items()}

            if display_values:
                print("\n-------------------------------------------")
                print(f"{'NAME':<15} | {'MIN':>6} | {'POS':>6} | {'MAX':>6}")
-                for motor in motors:
+                for motor in motor_names:
                    print(f"{motor:<15} | {mins[motor]:>6} | {positions[motor]:>6} | {maxes[motor]:>6}")

            if enter_pressed():
@@ -830,9 +827,9 @@ class SerialMotorsBus(MotorsBusBase):

            if display_values and not user_pressed_enter:
                # Move cursor up to overwrite the previous output
-                move_cursor_up(len(motors) + 3)
+                move_cursor_up(len(motor_names) + 3)

-        same_min_max = [motor for motor in motors if mins[motor] == maxes[motor]]
+        same_min_max = [motor for motor in motor_names if mins[motor] == maxes[motor]]
        if same_min_max:
            raise ValueError(f"Some motors have the same min and max values:\n{pformat(same_min_max)}")

@@ -955,12 +952,12 @@ class SerialMotorsBus(MotorsBusBase):
            if raise_on_error:
                raise ConnectionError(self.packet_handler.getTxRxResult(comm))
            else:
-                return
+                return None
        if self._is_error(error):
            if raise_on_error:
                raise RuntimeError(self.packet_handler.getRxPacketError(error))
            else:
-                return
+                return None

        return model_number

@@ -1007,12 +1004,13 @@ class SerialMotorsBus(MotorsBusBase):
        err_msg = f"Failed to read '{data_name}' on {id_=} after {num_retry + 1} tries."
        value, _, _ = self._read(addr, length, id_, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)

-        id_value = self._decode_sign(data_name, {id_: value})
+        decoded = self._decode_sign(data_name, {id_: value})

        if normalize and data_name in self.normalized_data:
-            id_value = self._normalize(id_value)
+            normalized = self._normalize(decoded)
+            return normalized[id_]

-        return id_value[id_]
+        return decoded[id_]

    def _read(
        self,
@@ -1023,7 +1021,7 @@ class SerialMotorsBus(MotorsBusBase):
        num_retry: int = 0,
        raise_on_error: bool = True,
        err_msg: str = "",
-    ) -> tuple[int, int]:
+    ) -> tuple[int, int, int]:
        if length == 1:
            read_fn = self.packet_handler.read1ByteTxRx
        elif length == 2:
@@ -1073,13 +1071,14 @@ class SerialMotorsBus(MotorsBusBase):
        model = self.motors[motor].model
        addr, length = get_address(self.model_ctrl_table, model, data_name)

+        int_value = int(value)
        if normalize and data_name in self.normalized_data:
-            value = self._unnormalize({id_: value})[id_]
+            int_value = self._unnormalize({id_: value})[id_]

-        value = self._encode_sign(data_name, {id_: value})[id_]
+        int_value = self._encode_sign(data_name, {id_: int_value})[id_]

-        err_msg = f"Failed to write '{data_name}' on {id_=} with '{value}' after {num_retry + 1} tries."
-        self._write(addr, length, id_, value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
+        err_msg = f"Failed to write '{data_name}' on {id_=} with '{int_value}' after {num_retry + 1} tries."
+        self._write(addr, length, id_, int_value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)

    def _write(
        self,
@@ -1113,7 +1112,7 @@ class SerialMotorsBus(MotorsBusBase):
    def sync_read(
        self,
        data_name: str,
-        motors: str | list[str] | None = None,
+        motors: NameOrID | Sequence[NameOrID] | None = None,
        *,
        normalize: bool = True,
        num_retry: int = 0,
@@ -1122,7 +1121,7 @@ class SerialMotorsBus(MotorsBusBase):

        Args:
            data_name (str): Register name.
-            motors (str | list[str] | None, optional): Motors to query. `None` (default) reads every motor.
+            motors (NameOrID | Sequence[NameOrID] | None, optional): Motors to query. `None` (default) reads every motor.
            normalize (bool, optional): Normalisation flag.  Defaults to `True`.
            num_retry (int, optional): Retry attempts.  Defaults to `0`.

@@ -1143,16 +1142,17 @@ class SerialMotorsBus(MotorsBusBase):
        addr, length = get_address(self.model_ctrl_table, model, data_name)

        err_msg = f"Failed to sync read '{data_name}' on {ids=} after {num_retry + 1} tries."
-        ids_values, _ = self._sync_read(
+        raw_ids_values, _ = self._sync_read(
            addr, length, ids, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
        )

-        ids_values = self._decode_sign(data_name, ids_values)
+        decoded = self._decode_sign(data_name, raw_ids_values)

        if normalize and data_name in self.normalized_data:
-            ids_values = self._normalize(ids_values)
+            normalized = self._normalize(decoded)
+            return {self._id_to_name(id_): value for id_, value in normalized.items()}

-        return {self._id_to_name(id_): value for id_, value in ids_values.items()}
+        return {self._id_to_name(id_): value for id_, value in decoded.items()}

    def _sync_read(
        self,
@@ -1224,21 +1224,24 @@ class SerialMotorsBus(MotorsBusBase):
            num_retry (int, optional): Retry attempts.  Defaults to `0`.
        """

-        ids_values = self._get_ids_values_dict(values)
-        models = [self._id_to_model(id_) for id_ in ids_values]
+        raw_ids_values = self._get_ids_values_dict(values)
+        models = [self._id_to_model(id_) for id_ in raw_ids_values]
        if self._has_different_ctrl_tables:
            assert_same_address(self.model_ctrl_table, models, data_name)

        model = next(iter(models))
        addr, length = get_address(self.model_ctrl_table, model, data_name)

+        int_ids_values = {id_: int(val) for id_, val in raw_ids_values.items()}
        if normalize and data_name in self.normalized_data:
-            ids_values = self._unnormalize(ids_values)
+            int_ids_values = self._unnormalize(raw_ids_values)

-        ids_values = self._encode_sign(data_name, ids_values)
+        int_ids_values = self._encode_sign(data_name, int_ids_values)

-        err_msg = f"Failed to sync write '{data_name}' with {ids_values=} after {num_retry + 1} tries."
-        self._sync_write(addr, length, ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
+        err_msg = f"Failed to sync write '{data_name}' with ids_values={int_ids_values} after {num_retry + 1} tries."
+        self._sync_write(
+            addr, length, int_ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
+        )

    def _sync_write(
        self,
--- a/src/lerobot/policies/act/configuration_act.py
+++ b/src/lerobot/policies/act/configuration_act.py
@@ -28,7 +28,7 @@ class ACTConfig(PreTrainedConfig):
    Defaults are configured for training on bimanual Aloha tasks like "insertion" or "transfer".

    The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_shapes` and 'output_shapes`.
+    Those are: `input_features` and `output_features`.

    Notes on the inputs and outputs:
        - Either:
@@ -48,21 +48,12 @@ class ACTConfig(PreTrainedConfig):
            This should be no greater than the chunk size. For example, if the chunk size size 100, you may
            set this to 50. This would mean that the model predicts 100 steps worth of actions, runs 50 in the
            environment, and throws the other 50 out.
-        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
-            the input data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
-            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
-            include batch dimension or temporal dimension.
-        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
-            the output data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
-            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
-        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
-            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
-            [-1, 1] range.
-        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
-            original scale. Note that this is also used for normalizing the training targets.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
        vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
        pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
            `None` means no pretrained weights.
--- a/src/lerobot/policies/diffusion/configuration_diffusion.py
+++ b/src/lerobot/policies/diffusion/configuration_diffusion.py
@@ -30,7 +30,7 @@ class DiffusionConfig(PreTrainedConfig):
    Defaults are configured for training with PushT providing proprioceptive and single camera observations.

    The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_shapes` and `output_shapes`.
+    Those are: `input_features` and `output_features`.

    Notes on the inputs and outputs:
        - "observation.state" is required as an input key.
@@ -48,21 +48,12 @@ class DiffusionConfig(PreTrainedConfig):
        horizon: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
        n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
            See `DiffusionPolicy.select_action` for more details.
-        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
-            the input data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
-            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
-            include batch dimension or temporal dimension.
-        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
-            the output data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
-            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
-        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
-            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
-            [-1, 1] range.
-        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
-            original scale. Note that this is also used for normalizing the training targets.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
        vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
        crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
            within the image size. If None, no cropping is done.
@@ -73,7 +64,7 @@ class DiffusionConfig(PreTrainedConfig):
        use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
            The group sizes are set to be about 16 (to be precise, feature_dim // 16).
        spatial_softmax_num_keypoints: Number of keypoints for SpatialSoftmax.
-        use_separate_rgb_encoders_per_camera: Whether to use a separate RGB encoder for each camera view.
+        use_separate_rgb_encoder_per_camera: Whether to use a separate RGB encoder for each camera view.
        down_dims: Feature dimension for each stage of temporal downsampling in the diffusion modeling Unet.
            You may provide a variable number of dimensions, therefore also controlling the degree of
            downsampling.
--- a/src/lerobot/policies/rlt/init.py
+++ b/src/lerobot/policies/rlt/init.py
@@ -0,0 +1,18 @@
+# Copyright 2026 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.policies.rlt.configuration_rlt import RLTConfig
+from lerobot.policies.rlt.modeling_rlt import RLTPolicy
+
+__all__ = ["RLTConfig", "RLTPolicy"]
--- a/src/lerobot/policies/rlt/configuration_rlt.py
+++ b/src/lerobot/policies/rlt/configuration_rlt.py
@@ -0,0 +1,156 @@
+# Copyright 2026 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.
+"""RLT (RL Token) policy configuration.
+
+Reference: "RL Token: Bootstrapping Online RL with Vision-Language-Action Models"
+(Xu et al., Physical Intelligence, 2026)
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
+from lerobot.policies.sac.configuration_sac import ActorLearnerConfig, ConcurrencyConfig
+from lerobot.utils.constants import ACTION, OBS_IMAGE, OBS_STATE
+
+
+@dataclass
+class RLTokenConfig:
+    """Configuration for the RL-token encoder/decoder transformer."""
+
+    input_dim: int = 2048
+    rl_token_dim: int = 2048
+    num_encoder_layers: int = 2
+    num_decoder_layers: int = 2
+    num_heads: int = 8
+    ff_dim: int = 2048
+    dropout: float = 0.0
+
+
+@dataclass
+class RLTActorConfig:
+    """Configuration for the lightweight RL actor MLP."""
+
+    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
+    std: float = 0.1
+
+
+@dataclass
+class RLTCriticConfig:
+    """Configuration for the RLT critic MLP."""
+
+    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
+
+
+@PreTrainedConfig.register_subclass("rlt")
+@dataclass
+class RLTConfig(PreTrainedConfig):
+    """Configuration for the RLT (RL Token) policy.
+
+    RLT adds an RL-token encoder/decoder to a frozen VLA backbone, then trains
+    a lightweight actor-critic head using the RL token as state representation.
+    The frozen VLA also provides reference action chunks that the actor refines.
+    """
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.MEAN_STD,
+            "STATE": NormalizationMode.MIN_MAX,
+            "ACTION": NormalizationMode.MIN_MAX,
+        }
+    )
+
+    dataset_stats: dict[str, dict[str, list[float]]] | None = field(
+        default_factory=lambda: {
+            OBS_IMAGE: {
+                "mean": [0.485, 0.456, 0.406],
+                "std": [0.229, 0.224, 0.225],
+            },
+            OBS_STATE: {"min": [0.0], "max": [1.0]},
+            ACTION: {"min": [0.0], "max": [1.0]},
+        }
+    )
+
+    # ── Device ──
+    device: str = "cuda"
+    storage_device: str = "cpu"
+
+    # ── VLA backbone ──
+    vla_checkpoint: str | None = None
+
+    # ── RL-token ──
+    rl_token: RLTokenConfig = field(default_factory=RLTokenConfig)
+
+    # ── Actor / Critic heads ──
+    actor: RLTActorConfig = field(default_factory=RLTActorConfig)
+    critic: RLTCriticConfig = field(default_factory=RLTCriticConfig)
+
+    # ── Action chunks ──
+    chunk_size: int = 10
+    vla_chunk_size: int = 50
+
+    # ── Training parameters ──
+    online_steps: int = 50000
+    offline_steps: int = 5000
+    online_buffer_capacity: int = 100000
+    offline_buffer_capacity: int = 100000
+    online_step_before_learning: int = 500
+    warmup_steps: int = 500
+    async_prefetch: bool = False
+
+    # ── Algorithm hyperparameters ──
+    utd_ratio: int = 5
+    policy_update_freq: int = 2
+    discount: float = 0.99
+    critic_lr: float = 3e-4
+    actor_lr: float = 3e-4
+    rl_token_lr: float = 1e-4
+    tau: float = 0.005
+    clip_grad_norm: float = 10.0
+    num_critics: int = 2
+    bc_reg_coeff: float = 0.1
+    ref_dropout: float = 0.5
+    chunk_stride: int = 2
+    vla_finetune_weight: float = 0.0
+
+    # ── Distributed ──
+    actor_learner_config: ActorLearnerConfig = field(default_factory=ActorLearnerConfig)
+    concurrency: ConcurrencyConfig = field(default_factory=ConcurrencyConfig)
+
+    def __post_init__(self):
+        super().__post_init__()
+
+    def get_optimizer_preset(self):
+        return None
+
+    def get_scheduler_preset(self):
+        return None
+
+    def validate_features(self) -> None:
+        if ACTION not in self.output_features:
+            raise ValueError("You must provide 'action' in the output features")
+
+    @property
+    def observation_delta_indices(self) -> list | None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list | None:
+        return None
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None
--- a/src/lerobot/policies/rlt/modeling_rlt.py
+++ b/src/lerobot/policies/rlt/modeling_rlt.py
@@ -0,0 +1,318 @@
+# Copyright 2026 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.
+"""RLT (RL Token) policy networks.
+
+Reference: "RL Token: Bootstrapping Online RL with Vision-Language-Action Models"
+(Xu et al., Physical Intelligence, 2026)
+
+Architecture:
+  - RLTokenEncoder: compresses VLA token embeddings into a single compact RL token
+  - RLTokenDecoder: reconstructs VLA embeddings from the RL token (Stage 1 training only)
+  - RLTActor: refines VLA reference action chunks conditioned on (z_rl, proprioception, ref_action)
+  - RLTCritic: Q(x, action_chunk) where x = (z_rl, proprioception)
+  - RLTPolicy: bundles RL-token modules + actor into a PreTrainedPolicy for inference
+"""
+
+from __future__ import annotations
+
+import math
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rlt.configuration_rlt import RLTConfig
+
+# ── Building blocks ──────────────────────────────────────────────────
+
+
+class MLP(nn.Module):
+    """Simple feedforward network with ReLU activations."""
+
+    def __init__(self, input_dim: int, hidden_dims: list[int], output_dim: int):
+        super().__init__()
+        layers: list[nn.Module] = []
+        prev = input_dim
+        for h in hidden_dims:
+            layers.append(nn.Linear(prev, h))
+            layers.append(nn.ReLU())
+            prev = h
+        layers.append(nn.Linear(prev, output_dim))
+        self.net = nn.Sequential(*layers)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.net(x)
+
+
+# ── RL Token Encoder ─────────────────────────────────────────────────
+
+
+class RLTokenEncoder(nn.Module):
+    """Compress VLA token embeddings into a single RL token via a small transformer.
+
+    Appends a learnable ``e_rl`` embedding to the VLA token sequence, processes
+    through transformer encoder layers, and returns the output at the ``e_rl``
+    position as the RL token ``z_rl``.
+
+    Paper Eq. 1: z_rl = g_phi([z_{1:M}, e_rl])_{M+1}
+    """
+
+    def __init__(
+        self,
+        input_dim: int,
+        rl_token_dim: int,
+        num_layers: int,
+        num_heads: int,
+        ff_dim: int,
+        dropout: float = 0.0,
+    ):
+        super().__init__()
+        self.rl_token_dim = rl_token_dim
+
+        self.e_rl = nn.Parameter(torch.randn(1, 1, input_dim) * 0.02)
+
+        if input_dim != rl_token_dim:
+            self.input_proj = nn.Linear(input_dim, rl_token_dim)
+        else:
+            self.input_proj = nn.Identity()
+
+        encoder_layer = nn.TransformerEncoderLayer(
+            d_model=rl_token_dim,
+            nhead=num_heads,
+            dim_feedforward=ff_dim,
+            dropout=dropout,
+            batch_first=True,
+        )
+        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
+
+    def forward(self, z_vla: Tensor) -> Tensor:
+        """
+        Args:
+            z_vla: VLA token embeddings, shape ``(B, M, D)``.
+
+        Returns:
+            RL token ``z_rl``, shape ``(B, rl_token_dim)``.
+        """
+        batch_size = z_vla.shape[0]
+        e_rl = self.e_rl.expand(batch_size, -1, -1)
+        seq = torch.cat([z_vla, e_rl], dim=1)  # (B, M+1, D)
+        seq = self.input_proj(seq)
+        out = self.transformer(seq)
+        z_rl = out[:, -1, :]  # output at e_rl position
+        return z_rl
+
+
+# ── RL Token Decoder ─────────────────────────────────────────────────
+
+
+class RLTokenDecoder(nn.Module):
+    """Autoregressively reconstruct VLA embeddings from z_rl.
+
+    Used only during Stage 1 (offline RL-token training).
+
+    Paper Eq. 2: L_ro = E[sum_i || h(d([z_rl, z_bar_{1:i-1}]))_i - z_bar_i ||^2]
+    """
+
+    def __init__(
+        self,
+        rl_token_dim: int,
+        output_dim: int,
+        num_layers: int,
+        num_heads: int,
+        ff_dim: int,
+        dropout: float = 0.0,
+    ):
+        super().__init__()
+        self.output_dim = output_dim
+
+        if rl_token_dim != output_dim:
+            self.rl_proj = nn.Linear(rl_token_dim, output_dim)
+        else:
+            self.rl_proj = nn.Identity()
+
+        decoder_layer = nn.TransformerDecoderLayer(
+            d_model=output_dim,
+            nhead=num_heads,
+            dim_feedforward=ff_dim,
+            dropout=dropout,
+            batch_first=True,
+        )
+        self.transformer = nn.TransformerDecoder(decoder_layer, num_layers=num_layers)
+        self.output_head = nn.Linear(output_dim, output_dim)
+
+    def forward(self, z_rl: Tensor, z_vla_stopped: Tensor) -> Tensor:
+        """
+        Args:
+            z_rl: RL token, shape ``(B, D_rl)``.
+            z_vla_stopped: Stop-gradient VLA embeddings, shape ``(B, M, D)``.
+
+        Returns:
+            Reconstructed embeddings, shape ``(B, M, D)``.
+        """
+        seq_len = z_vla_stopped.shape[1]
+        z_rl_proj = self.rl_proj(z_rl).unsqueeze(1)
+
+        target = torch.cat([z_rl_proj, z_vla_stopped[:, :-1, :]], dim=1)
+
+        causal_mask = nn.Transformer.generate_square_subsequent_mask(seq_len, device=z_rl.device)
+
+        decoded = self.transformer(
+            tgt=target,
+            memory=z_rl_proj,
+            tgt_mask=causal_mask,
+        )
+        return self.output_head(decoded)  # (B, M, D)
+
+
+# ── Actor ────────────────────────────────────────────────────────────
+
+
+class RLTActor(nn.Module):
+    """Lightweight actor that refines VLA reference action chunks.
+
+    Paper Eq. 4: pi_theta(a_{1:C} | x, a_tilde_{1:C}) = N(mu_theta(x, a_tilde), sigma^2 I)
+
+    The actor is conditioned on both the RL state and the VLA's proposed action
+    chunk, acting as a "VLA-guided action editor".
+    """
+
+    def __init__(self, state_dim: int, action_chunk_dim: int, hidden_dims: list[int], std: float = 0.1):
+        super().__init__()
+        input_dim = state_dim + action_chunk_dim
+        self.net = MLP(input_dim, hidden_dims, action_chunk_dim)
+        self.log_std = math.log(std)
+
+    def forward(self, state: Tensor, ref_action_chunk: Tensor) -> Tensor:
+        """Return the mean action chunk.
+
+        Args:
+            state: RL state ``x = (z_rl, proprioception)``, shape ``(B, state_dim)``.
+            ref_action_chunk: Flattened VLA reference chunk, shape ``(B, C*d)``.
+
+        Returns:
+            Refined action chunk (mean), shape ``(B, C*d)``.
+        """
+        x = torch.cat([state, ref_action_chunk], dim=-1)
+        return self.net(x)
+
+    def sample(self, state: Tensor, ref_action_chunk: Tensor) -> tuple[Tensor, Tensor]:
+        """Sample an action and return (action, log_prob)."""
+        mean = self.forward(state, ref_action_chunk)
+        std = math.exp(self.log_std)
+        noise = torch.randn_like(mean) * std
+        action = mean + noise
+        log_prob = -0.5 * (noise / std).pow(2).sum(dim=-1) - mean.shape[-1] * math.log(
+            std * math.sqrt(2 * math.pi)
+        )
+        return action, log_prob
+
+
+# ── Policy (inference bundle) ────────────────────────────────────────
+
+
+class RLTPolicy(PreTrainedPolicy):
+    """RLT policy — bundles the RL-token encoder and actor for inference.
+
+    The frozen VLA backbone is **not** part of this module; it is loaded
+    separately and its embeddings / reference actions are passed in via the
+    observation dict (populated by the actor process or a preprocessor).
+
+    During training, the :class:`RLTAlgorithm` holds the critic, target networks,
+    and optimizers. This class only contains what is needed for ``select_action``.
+    """
+
+    name = "rlt"
+    config_class = RLTConfig
+
+    def __init__(self, config: RLTConfig, dataset_stats=None):
+        super().__init__(config, dataset_stats)
+        action_dim = config.output_features["action"].shape[0]
+        action_chunk_dim = config.chunk_size * action_dim
+        prop_feature = config.input_features.get("observation.state", None)
+        proprioception_dim = prop_feature.shape[0] if prop_feature is not None else 0
+
+        state_dim = config.rl_token.rl_token_dim + proprioception_dim
+
+        # RL-token encoder (frozen after Stage 1)
+        self.rl_token_encoder = RLTokenEncoder(
+            input_dim=config.rl_token.input_dim,
+            rl_token_dim=config.rl_token.rl_token_dim,
+            num_layers=config.rl_token.num_encoder_layers,
+            num_heads=config.rl_token.num_heads,
+            ff_dim=config.rl_token.ff_dim,
+            dropout=config.rl_token.dropout,
+        )
+
+        # RL-token decoder (used only during Stage 1 training)
+        self.rl_token_decoder = RLTokenDecoder(
+            rl_token_dim=config.rl_token.rl_token_dim,
+            output_dim=config.rl_token.input_dim,
+            num_layers=config.rl_token.num_decoder_layers,
+            num_heads=config.rl_token.num_heads,
+            ff_dim=config.rl_token.ff_dim,
+            dropout=config.rl_token.dropout,
+        )
+
+        # Actor MLP
+        self.actor = RLTActor(
+            state_dim=state_dim,
+            action_chunk_dim=action_chunk_dim,
+            hidden_dims=config.actor.hidden_dims,
+            std=config.actor.std,
+        )
+
+        self._action_dim = action_dim
+        self._action_chunk_dim = action_chunk_dim
+        self._state_dim = state_dim
+        self._proprioception_dim = proprioception_dim
+
+    @torch.no_grad()
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select a refined action chunk given an observation.
+
+        Expects the observation dict to contain:
+          - ``"observation.vla_embeddings"``: VLA internal token embeddings ``(M, D)``
+          - ``"observation.reference_action"``: VLA reference chunk ``(C*d,)``
+          - ``"observation.state"`` (optional): proprioceptive state ``(P,)``
+
+        Returns:
+            Action chunk tensor of shape ``(C*d,)``.
+        """
+        self.eval()
+
+        vla_emb = batch["observation.vla_embeddings"]
+        if vla_emb.dim() == 2:
+            vla_emb = vla_emb.unsqueeze(0)
+
+        z_rl = self.rl_token_encoder(vla_emb)  # (1, D_rl)
+
+        parts = [z_rl]
+        if "observation.state" in batch and self._proprioception_dim > 0:
+            prop = batch["observation.state"]
+            if prop.dim() == 1:
+                prop = prop.unsqueeze(0)
+            parts.append(prop)
+
+        state = torch.cat(parts, dim=-1)
+
+        ref = batch["observation.reference_action"]
+        if ref.dim() == 1:
+            ref = ref.unsqueeze(0)
+
+        action = self.actor(state, ref)
+        return action.squeeze(0)
+
+    def reset(self):
+        pass
--- a/src/lerobot/policies/sac/modeling_sac.py
+++ b/src/lerobot/policies/sac/modeling_sac.py
@@ -15,16 +15,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-import math
 from collections.abc import Callable
 from dataclasses import asdict
-from typing import Literal

-import einops
-import numpy as np
 import torch
 import torch.nn as nn
-import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 from torch.distributions import MultivariateNormal, TanhTransform, Transform, TransformedDistribution

@@ -52,20 +47,13 @@ class SACPolicy(

        # Determine action dimension and initialize all components
        continuous_action_dim = config.output_features[ACTION].shape[0]
-        self._init_encoders()
-        self._init_critics(continuous_action_dim)
+        self.encoder = SACObservationEncoder(config)
        self._init_actor(continuous_action_dim)
-        self._init_temperature()
+        self._init_discrete_critic()

    def get_optim_params(self) -> dict:
        optim_params = {
-            "actor": [
-                p
-                for n, p in self.actor.named_parameters()
-                if not n.startswith("encoder") or not self.shared_encoder
-            ],
-            "critic": self.critic_ensemble.parameters(),
-            "temperature": self.log_alpha,
+            "actor": [self.actor.parameters()],
        }
        if self.config.num_discrete_actions is not None:
            optim_params["discrete_critic"] = self.discrete_critic.parameters()
@@ -83,10 +71,9 @@ class SACPolicy(
    @torch.no_grad()
    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
        """Select action for inference/evaluation"""
-
        observations_features = None
-        if self.shared_encoder and self.actor.encoder.has_images:
-            observations_features = self.actor.encoder.get_cached_image_features(batch)
+        if self.encoder.has_images:
+            observations_features = self.encoder.get_cached_image_features(batch)

        actions, _, _ = self.actor(batch, observations_features)

@@ -97,372 +84,35 @@ class SACPolicy(

        return actions

-    def critic_forward(
-        self,
-        observations: dict[str, Tensor],
-        actions: Tensor,
-        use_target: bool = False,
-        observation_features: Tensor | None = None,
-    ) -> Tensor:
-        """Forward pass through a critic network ensemble
-
-        Args:
-            observations: Dictionary of observations
-            actions: Action tensor
-            use_target: If True, use target critics, otherwise use ensemble critics
-
-        Returns:
-            Tensor of Q-values from all critics
-        """
-
-        critics = self.critic_target if use_target else self.critic_ensemble
-        q_values = critics(observations, actions, observation_features)
-        return q_values
-
-    def discrete_critic_forward(
-        self, observations, use_target=False, observation_features=None
-    ) -> torch.Tensor:
-        """Forward pass through a discrete critic network
-
-        Args:
-            observations: Dictionary of observations
-            use_target: If True, use target critics, otherwise use ensemble critics
-            observation_features: Optional pre-computed observation features to avoid recomputing encoder output
-
-        Returns:
-            Tensor of Q-values from the discrete critic network
-        """
-        discrete_critic = self.discrete_critic_target if use_target else self.discrete_critic
-        q_values = discrete_critic(observations, observation_features)
-        return q_values
-
    def forward(
        self,
        batch: dict[str, Tensor | dict[str, Tensor]],
-        model: Literal["actor", "critic", "temperature", "discrete_critic"] = "critic",
    ) -> dict[str, Tensor]:
-        """Compute the loss for the given model
+        """Actor forward pass."""
+        observations = batch.get("state", batch)
+        observation_features = batch.get("observation_feature") if isinstance(batch, dict) else None
+        actions, log_probs, means = self.actor(observations, observation_features)
+        return {"action": actions, "log_prob": log_probs, "action_mean": means}

-        Args:
-            batch: Dictionary containing:
-                - action: Action tensor
-                - reward: Reward tensor
-                - state: Observations tensor dict
-                - next_state: Next observations tensor dict
-                - done: Done mask tensor
-                - observation_feature: Optional pre-computed observation features
-                - next_observation_feature: Optional pre-computed next observation features
-            model: Which model to compute the loss for ("actor", "critic", "discrete_critic", or "temperature")
-
-        Returns:
-            The computed loss tensor
-        """
-        # Extract common components from batch
-        actions: Tensor = batch[ACTION]
-        observations: dict[str, Tensor] = batch["state"]
-        observation_features: Tensor = batch.get("observation_feature")
-
-        if model == "critic":
-            # Extract critic-specific components
-            rewards: Tensor = batch["reward"]
-            next_observations: dict[str, Tensor] = batch["next_state"]
-            done: Tensor = batch["done"]
-            next_observation_features: Tensor = batch.get("next_observation_feature")
-
-            loss_critic = self.compute_loss_critic(
-                observations=observations,
-                actions=actions,
-                rewards=rewards,
-                next_observations=next_observations,
-                done=done,
-                observation_features=observation_features,
-                next_observation_features=next_observation_features,
-            )
-
-            return {"loss_critic": loss_critic}
-
-        if model == "discrete_critic" and self.config.num_discrete_actions is not None:
-            # Extract critic-specific components
-            rewards: Tensor = batch["reward"]
-            next_observations: dict[str, Tensor] = batch["next_state"]
-            done: Tensor = batch["done"]
-            next_observation_features: Tensor = batch.get("next_observation_feature")
-            complementary_info = batch.get("complementary_info")
-            loss_discrete_critic = self.compute_loss_discrete_critic(
-                observations=observations,
-                actions=actions,
-                rewards=rewards,
-                next_observations=next_observations,
-                done=done,
-                observation_features=observation_features,
-                next_observation_features=next_observation_features,
-                complementary_info=complementary_info,
-            )
-            return {"loss_discrete_critic": loss_discrete_critic}
-        if model == "actor":
-            return {
-                "loss_actor": self.compute_loss_actor(
-                    observations=observations,
-                    observation_features=observation_features,
-                )
-            }
-
-        if model == "temperature":
-            return {
-                "loss_temperature": self.compute_loss_temperature(
-                    observations=observations,
-                    observation_features=observation_features,
-                )
-            }
-
-        raise ValueError(f"Unknown model type: {model}")
-
-    def update_target_networks(self):
-        """Update target networks with exponential moving average"""
-        for target_param, param in zip(
-            self.critic_target.parameters(),
-            self.critic_ensemble.parameters(),
-            strict=True,
-        ):
-            target_param.data.copy_(
-                param.data * self.config.critic_target_update_weight
-                + target_param.data * (1.0 - self.config.critic_target_update_weight)
-            )
-        if self.config.num_discrete_actions is not None:
-            for target_param, param in zip(
-                self.discrete_critic_target.parameters(),
-                self.discrete_critic.parameters(),
-                strict=True,
-            ):
-                target_param.data.copy_(
-                    param.data * self.config.critic_target_update_weight
-                    + target_param.data * (1.0 - self.config.critic_target_update_weight)
-                )
-
-    @property
-    def temperature(self) -> float:
-        """Return the current temperature value, always in sync with log_alpha."""
-        return self.log_alpha.exp().item()
-
-    def compute_loss_critic(
-        self,
-        observations,
-        actions,
-        rewards,
-        next_observations,
-        done,
-        observation_features: Tensor | None = None,
-        next_observation_features: Tensor | None = None,
-    ) -> Tensor:
-        with torch.no_grad():
-            next_action_preds, next_log_probs, _ = self.actor(next_observations, next_observation_features)
-
-            # 2- compute q targets
-            q_targets = self.critic_forward(
-                observations=next_observations,
-                actions=next_action_preds,
-                use_target=True,
-                observation_features=next_observation_features,
-            )
-
-            # subsample critics to prevent overfitting if use high UTD (update to date)
-            # TODO: Get indices before forward pass to avoid unnecessary computation
-            if self.config.num_subsample_critics is not None:
-                indices = torch.randperm(self.config.num_critics)
-                indices = indices[: self.config.num_subsample_critics]
-                q_targets = q_targets[indices]
-
-            # critics subsample size
-            min_q, _ = q_targets.min(dim=0)  # Get values from min operation
-            if self.config.use_backup_entropy:
-                min_q = min_q - (self.temperature * next_log_probs)
-
-            td_target = rewards + (1 - done) * self.config.discount * min_q
-
-        # 3- compute predicted qs
-        if self.config.num_discrete_actions is not None:
-            # NOTE: We only want to keep the continuous action part
-            # In the buffer we have the full action space (continuous + discrete)
-            # We need to split them before concatenating them in the critic forward
-            actions: Tensor = actions[:, :DISCRETE_DIMENSION_INDEX]
-        q_preds = self.critic_forward(
-            observations=observations,
-            actions=actions,
-            use_target=False,
-            observation_features=observation_features,
-        )
-
-        # 4- Calculate loss
-        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
-        td_target_duplicate = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
-        # You compute the mean loss of the batch for each critic and then to compute the final loss you sum them up
-        critics_loss = (
-            F.mse_loss(
-                input=q_preds,
-                target=td_target_duplicate,
-                reduction="none",
-            ).mean(dim=1)
-        ).sum()
-        return critics_loss
-
-    def compute_loss_discrete_critic(
-        self,
-        observations,
-        actions,
-        rewards,
-        next_observations,
-        done,
-        observation_features=None,
-        next_observation_features=None,
-        complementary_info=None,
-    ):
-        # NOTE: We only want to keep the discrete action part
-        # In the buffer we have the full action space (continuous + discrete)
-        # We need to split them before concatenating them in the critic forward
-        actions_discrete: Tensor = actions[:, DISCRETE_DIMENSION_INDEX:].clone()
-        actions_discrete = torch.round(actions_discrete)
-        actions_discrete = actions_discrete.long()
-
-        discrete_penalties: Tensor | None = None
-        if complementary_info is not None:
-            discrete_penalties: Tensor | None = complementary_info.get("discrete_penalty")
-
-        with torch.no_grad():
-            # For DQN, select actions using online network, evaluate with target network
-            next_discrete_qs = self.discrete_critic_forward(
-                next_observations, use_target=False, observation_features=next_observation_features
-            )
-            best_next_discrete_action = torch.argmax(next_discrete_qs, dim=-1, keepdim=True)
-
-            # Get target Q-values from target network
-            target_next_discrete_qs = self.discrete_critic_forward(
-                observations=next_observations,
-                use_target=True,
-                observation_features=next_observation_features,
-            )
-
-            # Use gather to select Q-values for best actions
-            target_next_discrete_q = torch.gather(
-                target_next_discrete_qs, dim=1, index=best_next_discrete_action
-            ).squeeze(-1)
-
-            # Compute target Q-value with Bellman equation
-            rewards_discrete = rewards
-            if discrete_penalties is not None:
-                rewards_discrete = rewards + discrete_penalties
-            target_discrete_q = rewards_discrete + (1 - done) * self.config.discount * target_next_discrete_q
-
-        # Get predicted Q-values for current observations
-        predicted_discrete_qs = self.discrete_critic_forward(
-            observations=observations, use_target=False, observation_features=observation_features
-        )
-
-        # Use gather to select Q-values for taken actions
-        predicted_discrete_q = torch.gather(predicted_discrete_qs, dim=1, index=actions_discrete).squeeze(-1)
-
-        # Compute MSE loss between predicted and target Q-values
-        discrete_critic_loss = F.mse_loss(input=predicted_discrete_q, target=target_discrete_q)
-        return discrete_critic_loss
-
-    def compute_loss_temperature(self, observations, observation_features: Tensor | None = None) -> Tensor:
-        """Compute the temperature loss"""
-        # calculate temperature loss
-        with torch.no_grad():
-            _, log_probs, _ = self.actor(observations, observation_features)
-        temperature_loss = (-self.log_alpha.exp() * (log_probs + self.target_entropy)).mean()
-        return temperature_loss
-
-    def compute_loss_actor(
-        self,
-        observations,
-        observation_features: Tensor | None = None,
-    ) -> Tensor:
-        actions_pi, log_probs, _ = self.actor(observations, observation_features)
-
-        q_preds = self.critic_forward(
-            observations=observations,
-            actions=actions_pi,
-            use_target=False,
-            observation_features=observation_features,
-        )
-        min_q_preds = q_preds.min(dim=0)[0]
-
-        actor_loss = ((self.temperature * log_probs) - min_q_preds).mean()
-        return actor_loss
-
-    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.encoder_actor = (
-            self.encoder_critic if self.shared_encoder else SACObservationEncoder(self.config)
-        )
-
-    def _init_critics(self, continuous_action_dim):
-        """Build critic ensemble, targets, and optional discrete critic."""
-        heads = [
-            CriticHead(
-                input_dim=self.encoder_critic.output_dim + continuous_action_dim,
-                **asdict(self.config.critic_network_kwargs),
-            )
-            for _ in range(self.config.num_critics)
-        ]
-        self.critic_ensemble = CriticEnsemble(encoder=self.encoder_critic, ensemble=heads)
-        target_heads = [
-            CriticHead(
-                input_dim=self.encoder_critic.output_dim + continuous_action_dim,
-                **asdict(self.config.critic_network_kwargs),
-            )
-            for _ in range(self.config.num_critics)
-        ]
-        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:
-            self.critic_ensemble = torch.compile(self.critic_ensemble)
-            self.critic_target = torch.compile(self.critic_target)
-
-        if self.config.num_discrete_actions is not None:
-            self._init_discrete_critics()
-
-    def _init_discrete_critics(self):
-        """Build discrete discrete critic ensemble and target networks."""
-        self.discrete_critic = DiscreteCritic(
-            encoder=self.encoder_critic,
-            input_dim=self.encoder_critic.output_dim,
-            output_dim=self.config.num_discrete_actions,
-            **asdict(self.config.discrete_critic_network_kwargs),
-        )
-        self.discrete_critic_target = DiscreteCritic(
-            encoder=self.encoder_critic,
-            input_dim=self.encoder_critic.output_dim,
-            output_dim=self.config.num_discrete_actions,
-            **asdict(self.config.discrete_critic_network_kwargs),
-        )
-
-        # TODO: (maractingi, azouitine) Compile the discrete critic
-        self.discrete_critic_target.load_state_dict(self.discrete_critic.state_dict())
-
-    def _init_actor(self, continuous_action_dim):
-        """Initialize policy actor network and default target entropy."""
-        # NOTE: The actor select only the continuous action part
+    def _init_actor(self, continuous_action_dim: int) -> None:
        self.actor = Policy(
-            encoder=self.encoder_actor,
-            network=MLP(input_dim=self.encoder_actor.output_dim, **asdict(self.config.actor_network_kwargs)),
+            encoder=self.encoder,
+            network=MLP(input_dim=self.encoder.output_dim, **asdict(self.config.actor_network_kwargs)),
            action_dim=continuous_action_dim,
-            encoder_is_shared=self.shared_encoder,
+            encoder_is_shared=False,
            **asdict(self.config.policy_kwargs),
        )

-        self.target_entropy = self.config.target_entropy
-        if self.target_entropy is None:
-            dim = continuous_action_dim + (1 if self.config.num_discrete_actions is not None else 0)
-            self.target_entropy = -np.prod(dim) / 2
-
-    def _init_temperature(self) -> None:
-        """Set up temperature parameter (log_alpha)."""
-        temp_init = self.config.temperature_init
-        self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
+    def _init_discrete_critic(self) -> None:
+        if self.config.num_discrete_actions is None:
+            self.discrete_critic = None
+            return
+        self.discrete_critic = DiscreteCritic(
+            encoder=self.encoder,
+            input_dim=self.encoder.output_dim,
+            output_dim=self.config.num_discrete_actions,
+            **asdict(self.config.discrete_critic_network_kwargs),
+        )


 class SACObservationEncoder(nn.Module):
--- a/src/lerobot/policies/sarm/compute_rabc_weights.py
+++ b/src/lerobot/policies/sarm/compute_rabc_weights.py
@@ -27,18 +27,18 @@ Usage:
    # Full RA-BC computation with visualizations
    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path pepijn223/sarm_single_uni4
+        --reward-model-path <USER>/sarm_single_uni4

    # Faster computation with stride (compute every 5 frames, interpolate the rest)
    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path pepijn223/sarm_single_uni4 \\
+        --reward-model-path <USER>/sarm_single_uni4 \\
        --stride 5

    # Visualize predictions only (no RA-BC computation)
    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path pepijn223/sarm_single_uni4 \\
+        --reward-model-path <USER>/sarm_single_uni4 \\
        --visualize-only \\
        --num-visualizations 5

@@ -714,12 +714,12 @@ Examples:
    # Full RA-BC computation with visualizations
    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path pepijn223/sarm_single_uni4
+        --reward-model-path <USER>/sarm_single_uni4

    # Visualize predictions only (no RA-BC computation)
    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path pepijn223/sarm_single_uni4 \\
+        --reward-model-path <USER>/sarm_single_uni4 \\
        --visualize-only \\
        --num-visualizations 10
        """,
--- a/src/lerobot/policies/smolvla/modeling_smolvla.py
+++ b/src/lerobot/policies/smolvla/modeling_smolvla.py
@@ -30,7 +30,7 @@ Example of finetuning the smolvla pretrained model (`smolvla_base`):
 ```bash
 lerobot-train \
 --policy.path=lerobot/smolvla_base \
--dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
+--dataset.repo_id=<USER>/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```
@@ -40,7 +40,7 @@ and an action expert.
 ```bash
 lerobot-train \
 --policy.type=smolvla \
--dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
+--dataset.repo_id=<USER>/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```
@@ -378,16 +378,16 @@ class SmolVLAPolicy(PreTrainedPolicy):
        actions_is_pad = batch.get("actions_id_pad")
        loss_dict = {}
        losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions, noise, time)
-        loss_dict["losses_after_forward"] = losses.clone()
+        loss_dict["losses_after_forward"] = losses.clone().mean().item()

        if actions_is_pad is not None:
            in_episode_bound = ~actions_is_pad
            losses = losses * in_episode_bound.unsqueeze(-1)
-            loss_dict["losses_after_in_ep_bound"] = losses.clone()
+            loss_dict["losses_after_in_ep_bound"] = losses.clone().mean().item()

        # Remove padding
        losses = losses[:, :, : self.config.max_action_dim]
-        loss_dict["losses_after_rm_padding"] = losses.clone()
+        loss_dict["losses_after_rm_padding"] = losses.clone().mean().item()

        if reduction == "none":
            # Return per-sample losses (B,) by averaging over time and action dims
--- a/src/lerobot/policies/tdmpc/configuration_tdmpc.py
+++ b/src/lerobot/policies/tdmpc/configuration_tdmpc.py
@@ -30,7 +30,7 @@ class TDMPCConfig(PreTrainedConfig):
    camera observations.

    The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_shapes`, `output_shapes`, and perhaps `max_random_shift_ratio`.
+    Those are: `input_features`, `output_features`, and perhaps `max_random_shift_ratio`.

    Args:
        n_action_repeats: The number of times to repeat the action returned by the planning. (hint: Google
@@ -40,24 +40,12 @@ class TDMPCConfig(PreTrainedConfig):
            is an alternative to using action repeats. If this is set to more than 1, then we require
            `n_action_repeats == 1`, `use_mpc == True` and `n_action_steps <= horizon`. Note that this
            approach of using multiple steps from the plan is not in the original implementation.
-        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
-            the input data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
-            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
-            include batch dimension or temporal dimension.
-        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
-            the output data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
-            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
-        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
-            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
-            [-1, 1] range. Note that here this defaults to None meaning inputs are not normalized. This is to
-            match the original implementation.
-        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
-            original scale. Note that this is also used for normalizing the training targets. NOTE: Clipping
-            to [-1, +1] is used during MPPI/CEM. Therefore, it is recommended that you stick with "min_max"
-            normalization mode here.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
        image_encoder_hidden_dim: Number of channels for the convolutional layers used for image encoding.
        state_encoder_hidden_dim: Hidden dimension for MLP used for state vector encoding.
        latent_dim: Observation's latent embedding dimension.
--- a/src/lerobot/policies/vqbet/configuration_vqbet.py
+++ b/src/lerobot/policies/vqbet/configuration_vqbet.py
@@ -32,7 +32,7 @@ class VQBeTConfig(PreTrainedConfig):
    Defaults are configured for training with PushT providing proprioceptive and single camera observations.

    The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_shapes` and `output_shapes`.
+    Those are: `input_features` and `output_features`.

    Notes on the inputs and outputs:
        - "observation.state" is required as an input key.
@@ -46,21 +46,12 @@ class VQBeTConfig(PreTrainedConfig):
            current step and additional steps going back).
        n_action_pred_token: Total number of current token and future tokens that VQ-BeT predicts.
        action_chunk_size: Action chunk size of each action prediction token.
-        input_shapes: A dictionary defining the shapes of the input data for the policy.
-            The key represents the input data name, and the value is a list indicating the dimensions
-            of the corresponding data. For example, "observation.image" refers to an input from
-            a camera with dimensions [3, 96, 96], indicating it has three color channels and 96x96 resolution.
-            Importantly, shapes doesnt include batch dimension or temporal dimension.
-        output_shapes: A dictionary defining the shapes of the output data for the policy.
-            The key represents the output data name, and the value is a list indicating the dimensions
-            of the corresponding data. For example, "action" refers to an output shape of [14], indicating
-            14-dimensional actions. Importantly, shapes doesnt include batch dimension or temporal dimension.
-        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
-            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
-            [-1, 1] range.
-        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
-            original scale. Note that this is also used for normalizing the training targets.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
        vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
        crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
            within the image size. If None, no cropping is done.
--- a/src/lerobot/processor/init.py
+++ b/src/lerobot/processor/init.py
@@ -44,6 +44,7 @@ from .hil_processor import (
    AddTeleopActionAsComplimentaryDataStep,
    AddTeleopEventsAsInfoStep,
    GripperPenaltyProcessorStep,
+    GymHILAdapterProcessorStep,
    ImageCropResizeProcessorStep,
    InterventionActionProcessorStep,
    RewardClassifierProcessorStep,
@@ -87,6 +88,7 @@ __all__ = [
    "DoneProcessorStep",
    "EnvAction",
    "EnvTransition",
+    "GymHILAdapterProcessorStep",
    "GripperPenaltyProcessorStep",
    "hotswap_stats",
    "IdentityProcessorStep",
--- a/src/lerobot/processor/gym_action_processor.py
+++ b/src/lerobot/processor/gym_action_processor.py
@@ -20,6 +20,7 @@ from lerobot.configs.types import PipelineFeatureType, PolicyFeature

 from .converters import to_tensor
 from .core import EnvAction, EnvTransition, PolicyAction
+from .hil_processor import TELEOP_ACTION_KEY
 from .pipeline import ActionProcessorStep, ProcessorStep, ProcessorStepRegistry


@@ -89,6 +90,13 @@ class Numpy2TorchActionProcessorStep(ProcessorStep):
            torch_action = to_tensor(action, dtype=None)  # Preserve original dtype
            new_transition[TransitionKey.ACTION] = torch_action

+        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+        if TELEOP_ACTION_KEY in complementary_data:
+            teleop_action = complementary_data[TELEOP_ACTION_KEY]
+            if isinstance(teleop_action, EnvAction):
+                complementary_data[TELEOP_ACTION_KEY] = to_tensor(teleop_action)
+            new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
+
        return new_transition

    def transform_features(
--- a/src/lerobot/processor/hil_processor.py
+++ b/src/lerobot/processor/hil_processor.py
@@ -312,9 +312,40 @@ class TimeLimitProcessorStep(TruncatedProcessorStep):
        return features


+@ProcessorStepRegistry.register("gym_hil_adapter_processor")
+class GymHILAdapterProcessorStep(ProcessorStep):
+    """
+    Adapts the output of the `gym-hil` environment to the format expected by `lerobot` processors.
+
+    This step normalizes the `transition` object by:
+    1. Copying `teleop_action` from `info` to `complementary_data`.
+    2. Copying `is_intervention` from `info` (using the string key) to `info` (using the enum key).
+    """
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        info = transition.get(TransitionKey.INFO, {})
+        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+
+        if TELEOP_ACTION_KEY in info:
+            complementary_data[TELEOP_ACTION_KEY] = info[TELEOP_ACTION_KEY]
+
+        if "is_intervention" in info:
+            info[TeleopEvents.IS_INTERVENTION] = info["is_intervention"]
+
+        transition[TransitionKey.INFO] = info
+        transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
+
+        return transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
@dataclass
@ProcessorStepRegistry.register("gripper_penalty_processor")
-class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):
+class GripperPenaltyProcessorStep(ProcessorStep):
    """
    Applies a penalty for inefficient gripper usage.

@@ -329,26 +360,27 @@ class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):
    penalty: float = -0.01
    max_gripper_pos: float = 30.0

-    def complementary_data(self, complementary_data: dict) -> dict:
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
        """
        Calculates the gripper penalty and adds it to the complementary data.

        Args:
-            complementary_data: The incoming complementary data, which should contain
-                                raw joint positions.
+            transition: The incoming environment transition.

        Returns:
-            A new complementary data dictionary with the `discrete_penalty` key added.
+            The modified transition with the penalty added to complementary data.
        """
-        action = self.transition.get(TransitionKey.ACTION)
+        new_transition = transition.copy()
+        action = new_transition.get(TransitionKey.ACTION)
+        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})

        raw_joint_positions = complementary_data.get("raw_joint_positions")
        if raw_joint_positions is None:
-            return complementary_data
+            return new_transition

        current_gripper_pos = raw_joint_positions.get(GRIPPER_KEY, None)
        if current_gripper_pos is None:
-            return complementary_data
+            return new_transition

        # Gripper action is a PolicyAction at this stage
        gripper_action = action[-1].item()
@@ -364,11 +396,12 @@ class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):

        gripper_penalty = self.penalty * int(gripper_penalty_bool)

-        # Create new complementary data with penalty info
+        # Update complementary data with penalty info
        new_complementary_data = dict(complementary_data)
        new_complementary_data[DISCRETE_PENALTY_KEY] = gripper_penalty
+        new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data

-        return new_complementary_data
+        return new_transition

    def get_config(self) -> dict[str, Any]:
        """
--- a/src/lerobot/processor/normalize_processor.py
+++ b/src/lerobot/processor/normalize_processor.py
@@ -131,6 +131,15 @@ class _NormalizationMixin:
        if self.dtype is None:
            self.dtype = torch.float32
        self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)
+        self._reshape_visual_stats()
+
+    def _reshape_visual_stats(self) -> None:
+        """Reshape visual stats from ``[C]`` to ``[C, 1, 1]`` for image broadcasting."""
+        for key, feature in self.features.items():
+            if feature.type == FeatureType.VISUAL and key in self._tensor_stats:
+                for stat_name, stat_tensor in self._tensor_stats[key].items():
+                    if isinstance(stat_tensor, Tensor) and stat_tensor.ndim == 1:
+                        self._tensor_stats[key][stat_name] = stat_tensor.reshape(-1, 1, 1)

    def to(
        self, device: torch.device | str | None = None, dtype: torch.dtype | None = None
@@ -149,6 +158,7 @@ class _NormalizationMixin:
        if dtype is not None:
            self.dtype = dtype
        self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)
+        self._reshape_visual_stats()
        return self

    def state_dict(self) -> dict[str, Tensor]:
@@ -198,6 +208,7 @@ class _NormalizationMixin:
            # Don't load from state_dict, keep the explicitly provided stats
            # But ensure _tensor_stats is properly initialized
            self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)  # type: ignore[assignment]
+            self._reshape_visual_stats()
            return

        # Normal behavior: load stats from state_dict
@@ -208,6 +219,7 @@ class _NormalizationMixin:
            self._tensor_stats.setdefault(key, {})[stat_name] = tensor.to(
                dtype=torch.float32, device=self.device
            )
+        self._reshape_visual_stats()

        # Reconstruct the original stats dict from tensor stats for compatibility with to() method
        # and other functions that rely on self.stats
--- a/src/lerobot/processor/pipeline.py
+++ b/src/lerobot/processor/pipeline.py
@@ -413,7 +413,7 @@ class DataProcessorPipeline(HubMixin, Generic[TInput, TOutput]):
        Args:
            save_directory: The directory where the pipeline will be saved. If None, saves to
                HF_LEROBOT_HOME/processors/{sanitized_pipeline_name}.
-            repo_id: ID of your repository on the Hub. Used only if `push_to_hub=True`.
+            repo_id: ID of your repository on the Hub. Used only if `push_to_hub=true`.
            push_to_hub: Whether or not to push your object to the Hugging Face Hub after saving it.
            card_kwargs: Additional arguments passed to the card template to customize the card.
            config_filename: The name of the JSON configuration file. If None, a name is
--- a/src/lerobot/rl/init.py
+++ b/src/lerobot/rl/init.py
@@ -0,0 +1,13 @@
+# Copyright 2026 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.
--- a/src/lerobot/rl/actor.py
+++ b/src/lerobot/rl/actor.py
@@ -61,7 +61,7 @@ from lerobot.cameras import opencv  # noqa: F401
 from lerobot.configs import parser
 from lerobot.configs.train import TrainRLServerPipelineConfig
 from lerobot.policies.factory import make_policy
-from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.processor import TransitionKey
 from lerobot.rl.process import ProcessSignalHandler
 from lerobot.rl.queue import get_last_item_from_queue
@@ -248,16 +248,16 @@ def act_with_policy(

    logging.info("make_policy")

-    ### Instantiate the policy in both the actor and learner processes
-    ### To avoid sending a SACPolicy object through the port, we create a policy instance
-    ### on both sides, the learner sends the updated parameters every n steps to update the actor's parameters
-    policy: SACPolicy = make_policy(
+    policy = make_policy(
        cfg=cfg.policy,
        env_cfg=cfg.env,
    )
    policy = policy.eval()
    assert isinstance(policy, nn.Module)

+    # TODO: Re-enable processor pipeline once refactoring is validated against main
+    # preprocessor, postprocessor = None, None
+
    obs, info = online_env.reset()
    env_processor.reset()
    action_processor.reset()
@@ -288,7 +288,6 @@ def act_with_policy(

        # Time policy inference and check if it meets FPS requirement
        with policy_timer:
-            # Extract observation from transition for policy
            action = policy.select_action(batch=observation)
        policy_fps = policy_timer.fps_last

@@ -649,12 +648,12 @@ def interactions_stream(
 #  Policy functions


-def update_policy_parameters(policy: SACPolicy, parameters_queue: Queue, device):
+def update_policy_parameters(policy: PreTrainedPolicy, parameters_queue: Queue, device):
+    """Load the latest policy weights from the learner."""
    bytes_state_dict = get_last_item_from_queue(parameters_queue, block=False)
    if bytes_state_dict is not None:
        logging.info("[ACTOR] Load new parameters from Learner.")
        state_dicts = bytes_to_state_dict(bytes_state_dict)
-
        # TODO: check encoder parameter synchronization possible issues:
        # 1. When shared_encoder=True, we're loading stale encoder params from actor's state_dict
        #    instead of the updated encoder params from critic (which is optimized separately)
@@ -664,18 +663,9 @@ def update_policy_parameters(policy: SACPolicy, parameters_queue: Queue, device)
        # - Send critic's encoder state when shared_encoder=True
        # - Skip encoder params entirely when freeze_vision_encoder=True
        # - Ensure discrete_critic gets correct encoder state (currently uses encoder_critic)
-
        # Load actor state dict
-        actor_state_dict = move_state_dict_to_device(state_dicts["policy"], device=device)
-        policy.actor.load_state_dict(actor_state_dict)
-
-        # Load discrete critic if present
-        if hasattr(policy, "discrete_critic") and "discrete_critic" in state_dicts:
-            discrete_critic_state_dict = move_state_dict_to_device(
-                state_dicts["discrete_critic"], device=device
-            )
-            policy.discrete_critic.load_state_dict(discrete_critic_state_dict)
-            logging.info("[ACTOR] Loaded discrete critic parameters from Learner.")
+        state_dicts = move_state_dict_to_device(state_dicts, device=device)
+        policy.load_state_dict(state_dicts)


 #  Utilities functions
--- a/src/lerobot/rl/algorithms/init.py
+++ b/src/lerobot/rl/algorithms/init.py
@@ -0,0 +1,70 @@
+# Copyright 2026 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 torch
+
+from lerobot.rl.algorithms.base import (
+    RLAlgorithm,
+    RLAlgorithmConfig,
+    TrainingStats,
+)
+from lerobot.rl.algorithms.rlt import RLTAlgorithm, RLTAlgorithmConfig
+from lerobot.rl.algorithms.sac import SACAlgorithm, SACAlgorithmConfig
+
+
+def make_algorithm(
+    policy: torch.nn.Module,
+    policy_cfg,
+    *,
+    algorithm_name: str,
+) -> RLAlgorithm:
+    """Construct an :class:`RLAlgorithm` from a policy and its config.
+
+    Algorithm selection is explicit via ``algorithm_name`` (from
+    ``cfg.algorithm``).
+
+    This is fully registry-driven — adding a new algorithm only requires
+    registering an ``RLAlgorithmConfig`` subclass; no changes here.
+
+    The returned algorithm has **no optimizers** yet.  On the learner side,
+    call ``algorithm.make_optimizers()`` afterwards to create them.  On the
+    actor side (inference-only), leave them empty.
+
+    Args:
+        policy: Instantiated policy (e.g. ``SACPolicy``).
+        policy_cfg: The policy's ``PreTrainedConfig`` with the hyper-parameters
+            expected by the algorithm config's ``from_policy_config`` class-method.
+        algorithm_name: Algorithm registry key to instantiate.
+    """
+    known = RLAlgorithmConfig.get_known_choices()
+    if algorithm_name not in known:
+        raise ValueError(f"No RLAlgorithmConfig registered for '{algorithm_name}'. Known: {list(known)}")
+
+    config_cls = RLAlgorithmConfig.get_choice_class(algorithm_name)
+    algo_config = config_cls.from_policy_config(policy_cfg)
+    return algo_config.build_algorithm(policy)
+
+
+__all__ = [
+    "RLAlgorithm",
+    "RLAlgorithmConfig",
+    "TrainingStats",
+    "SACAlgorithm",
+    "SACAlgorithmConfig",
+    "RLTAlgorithm",
+    "RLTAlgorithmConfig",
+    "make_algorithm",
+]
--- a/src/lerobot/rl/algorithms/base.py
+++ b/src/lerobot/rl/algorithms/base.py
@@ -0,0 +1,183 @@
+# Copyright 2026 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.
+"""Base classes for RL algorithms.
+
+Defines the abstract interface that every algorithm must implement, a registry
+for algorithm configs, and a dataclass for training statistics.
+"""
+
+from __future__ import annotations
+
+import abc
+from collections.abc import Iterator
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+import draccus
+import torch
+from torch import Tensor
+from torch.optim import Optimizer
+
+if TYPE_CHECKING:
+    from lerobot.rl.data_sources.data_mixer import DataMixer
+
+BatchType = dict[str, Any]
+
+
+@dataclass
+class TrainingStats:
+    """Returned by ``algorithm.update()`` for logging and checkpointing."""
+
+    # Generic containers for all algorithms
+    losses: dict[str, float] = field(default_factory=dict)
+    grad_norms: dict[str, float] = field(default_factory=dict)
+    extra: dict[str, float] = field(default_factory=dict)
+
+    def to_log_dict(self) -> dict[str, float]:
+        """Flatten all stats into a single dict for logging."""
+
+        d: dict[str, float] = {}
+        for name, val in self.losses.items():
+            d[name] = val
+        for name, val in self.grad_norms.items():
+            d[f"{name}_grad_norm"] = val
+        for name, val in self.extra.items():
+            d[name] = val
+        return d
+
+
+@dataclass
+class RLAlgorithmConfig(draccus.ChoiceRegistry):
+    """Registry for algorithm configs."""
+
+    def build_algorithm(self, policy: torch.nn.Module) -> RLAlgorithm:
+        """Construct the :class:`RLAlgorithm` for this config.
+
+        Must be overridden by every registered config subclass.
+        """
+        raise NotImplementedError(f"{type(self).__name__} must implement build_algorithm()")
+
+    @classmethod
+    def from_policy_config(cls, policy_cfg: Any) -> RLAlgorithmConfig:
+        """Build an algorithm config from a policy config.
+
+        Must be overridden by every registered config subclass.
+        """
+        raise NotImplementedError(f"{cls.__name__} must implement from_policy_config()")
+
+
+class RLAlgorithm(abc.ABC):
+    """Base for all RL algorithms."""
+
+    @abc.abstractmethod
+    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """One complete training step.
+
+        The algorithm calls ``next(batch_iterator)`` as many times as it
+        needs (e.g. ``utd_ratio`` times for SAC) to obtain fresh batches.
+        The iterator is owned by the trainer; the algorithm just consumes
+        from it.
+        """
+        ...
+
+    def supports_offline_phase(self) -> bool:
+        """Whether this algorithm has an offline pretraining phase.
+
+        Algorithms like RLT (RL-token training) or ConRFT (Cal-QL pretraining)
+        return ``True`` here. The learner checks this before the main online
+        loop and routes to :meth:`offline_update` accordingly.
+        """
+        return False
+
+    def offline_update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """One offline training step (called before any online collection).
+
+        Only called when :meth:`supports_offline_phase` returns ``True``.
+        Uses the same iterator protocol as :meth:`update`.
+        """
+        raise NotImplementedError(
+            f"{type(self).__name__} does not implement offline_update(). "
+            "Either override this method or return False from supports_offline_phase()."
+        )
+
+    def transition_to_online(self) -> None:  # noqa: B027
+        """Called once when switching from offline to online phase.
+
+        Use this to freeze modules trained offline, rebuild optimizers for the
+        online phase, reset step counters, etc.
+
+        Default is a no-op; subclasses override when they have an offline phase.
+        """
+
+    def configure_data_iterator(
+        self,
+        data_mixer: DataMixer,
+        batch_size: int,
+        *,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ) -> Iterator[BatchType]:
+        """Create the data iterator this algorithm needs.
+
+        The default implementation uses the standard ``data_mixer.get_iterator()``.
+        Algorithms that need specialised sampling should override this method.
+        """
+        return data_mixer.get_iterator(
+            batch_size=batch_size,
+            async_prefetch=async_prefetch,
+            queue_size=queue_size,
+        )
+
+    def make_optimizers(self) -> dict[str, Optimizer]:
+        """Create, store, and return the optimizers needed for training.
+
+        Called on the **learner** side after construction.  Subclasses must
+        override this with algorithm-specific optimizer setup.
+        """
+        return {}
+
+    def get_optimizers(self) -> dict[str, Optimizer]:
+        """Return optimizers for checkpointing / external scheduling."""
+        return {}
+
+    @property
+    def optimization_step(self) -> int:
+        """Current learner optimization step.
+
+        Part of the stable contract for checkpoint/resume. Algorithms can
+        either use this default storage or override for custom behavior.
+        """
+        return getattr(self, "_optimization_step", 0)
+
+    @optimization_step.setter
+    def optimization_step(self, value: int) -> None:
+        self._optimization_step = int(value)
+
+    def get_weights(self) -> dict[str, Any]:
+        """Policy state-dict to push to actors."""
+        return {}
+
+    @abc.abstractmethod
+    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
+        """Load policy state-dict received from the learner (inverse of ``get_weights``)."""
+
+    @torch.no_grad()
+    def get_observation_features(
+        self, observations: Tensor, next_observations: Tensor
+    ) -> tuple[Tensor | None, Tensor | None]:
+        """Pre-compute observation features (e.g. frozen encoder cache).
+
+        Returns ``(None, None)`` when caching is not applicable.
+        """
+        return None, None
--- a/src/lerobot/rl/algorithms/rlt/init.py
+++ b/src/lerobot/rl/algorithms/rlt/init.py
@@ -0,0 +1,18 @@
+# Copyright 2026 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.rl.algorithms.rlt.configuration_rlt import RLTAlgorithmConfig
+from lerobot.rl.algorithms.rlt.rlt_algorithm import RLTAlgorithm
+
+__all__ = ["RLTAlgorithm", "RLTAlgorithmConfig"]
--- a/src/lerobot/rl/algorithms/rlt/configuration_rlt.py
+++ b/src/lerobot/rl/algorithms/rlt/configuration_rlt.py
@@ -0,0 +1,83 @@
+# Copyright 2026 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.
+"""RLT algorithm configuration."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+import torch
+
+from lerobot.rl.algorithms.base import RLAlgorithmConfig
+
+if TYPE_CHECKING:
+    from lerobot.rl.algorithms.rlt.rlt_algorithm import RLTAlgorithm
+
+
+@RLAlgorithmConfig.register_subclass("rlt")
+@dataclass
+class RLTAlgorithmConfig(RLAlgorithmConfig):
+    """RLT-specific hyper-parameters that control the update loop."""
+
+    # ── Action chunks ──
+    chunk_size: int = 10
+    chunk_stride: int = 2
+
+    # ── Update cadence ──
+    utd_ratio: int = 5
+    policy_update_freq: int = 2
+    clip_grad_norm: float = 10.0
+
+    # ── Learning rates ──
+    actor_lr: float = 3e-4
+    critic_lr: float = 3e-4
+    rl_token_lr: float = 1e-4
+
+    # ── TD learning ──
+    discount: float = 0.99
+    tau: float = 0.005
+    num_critics: int = 2
+
+    # ── Policy constraint (paper Eq. 5) ──
+    bc_reg_coeff: float = 0.1
+    ref_dropout: float = 0.5
+
+    # ── Offline RL-token training ──
+    vla_finetune_weight: float = 0.0
+
+    @classmethod
+    def from_policy_config(cls, policy_cfg) -> RLTAlgorithmConfig:
+        """Build from an existing ``RLTConfig`` (cfg.policy)."""
+        return cls(
+            chunk_size=policy_cfg.chunk_size,
+            chunk_stride=policy_cfg.chunk_stride,
+            utd_ratio=policy_cfg.utd_ratio,
+            policy_update_freq=policy_cfg.policy_update_freq,
+            clip_grad_norm=policy_cfg.clip_grad_norm,
+            actor_lr=policy_cfg.actor_lr,
+            critic_lr=policy_cfg.critic_lr,
+            rl_token_lr=policy_cfg.rl_token_lr,
+            discount=policy_cfg.discount,
+            tau=policy_cfg.tau,
+            num_critics=policy_cfg.num_critics,
+            bc_reg_coeff=policy_cfg.bc_reg_coeff,
+            ref_dropout=policy_cfg.ref_dropout,
+            vla_finetune_weight=policy_cfg.vla_finetune_weight,
+        )
+
+    def build_algorithm(self, policy: torch.nn.Module) -> RLTAlgorithm:
+        from lerobot.rl.algorithms.rlt.rlt_algorithm import RLTAlgorithm
+
+        return RLTAlgorithm(policy=policy, config=self)
--- a/src/lerobot/rl/algorithms/rlt/rlt_algorithm.py
+++ b/src/lerobot/rl/algorithms/rlt/rlt_algorithm.py
@@ -0,0 +1,319 @@
+# Copyright 2026 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.
+"""RLT (RL Token) algorithm.
+
+Implements the two-stage training from "RL Token: Bootstrapping Online RL
+with Vision-Language-Action Models" (Xu et al., Physical Intelligence, 2026).
+
+Stage 1 (offline): Train RL-token encoder/decoder via reconstruction loss.
+Stage 2 (online):  Train actor-critic with chunked TD, BC regularization,
+                   reference-action pass-through, and reference-action dropout.
+"""
+
+from __future__ import annotations
+
+import copy
+from collections.abc import Iterator
+from typing import Any
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from torch import Tensor
+from torch.optim import Optimizer
+
+from lerobot.policies.rlt.modeling_rlt import MLP, RLTPolicy
+from lerobot.policies.utils import get_device_from_parameters
+from lerobot.rl.algorithms.base import (
+    BatchType,
+    RLAlgorithm,
+    TrainingStats,
+)
+from lerobot.rl.algorithms.rlt.configuration_rlt import RLTAlgorithmConfig
+from lerobot.utils.constants import ACTION
+
+
+class RLTCritic(nn.Module):
+    """Q-function over (state, action_chunk) pairs.
+
+    Paper Eq. 3: Q_psi(x, a_{1:C})
+
+    Training-only component — lives on the algorithm side, not in the policy.
+    """
+
+    def __init__(self, state_dim: int, action_chunk_dim: int, hidden_dims: list[int]):
+        super().__init__()
+        self.net = MLP(state_dim + action_chunk_dim, hidden_dims, output_dim=1)
+
+    def forward(self, state: Tensor, action_chunk: Tensor) -> Tensor:
+        x = torch.cat([state, action_chunk], dim=-1)
+        return self.net(x)
+
+
+class RLTAlgorithm(RLAlgorithm):
+    """RL Token: lightweight actor-critic on frozen VLA features.
+
+    Owns the ``RLTPolicy`` (RL-token encoder/decoder + actor), a critic
+    ensemble, and target networks.  All VLA-specific logic (embedding
+    extraction, reference actions) lives in ``_prepare_forward_batch``.
+    """
+
+    def __init__(self, policy: RLTPolicy, config: RLTAlgorithmConfig):
+        self.policy = policy
+        self.config = config
+        self.optimizers: dict[str, Optimizer] = {}
+        self._optimization_step: int = 0
+        self._device = get_device_from_parameters(self.policy)
+        self._is_online = False
+
+        self._init_critics()
+        self._move_to_device()
+
+    # ── Initialization ───────────────────────────────────────────────
+
+    def _init_critics(self) -> None:
+        state_dim = self.policy._state_dim
+        action_chunk_dim = self.policy._action_chunk_dim
+        hidden_dims = self.policy.config.critic.hidden_dims
+
+        self.critics = torch.nn.ModuleList(
+            [RLTCritic(state_dim, action_chunk_dim, hidden_dims) for _ in range(self.config.num_critics)]
+        )
+        self.critic_targets = torch.nn.ModuleList([copy.deepcopy(c) for c in self.critics])
+        for ct in self.critic_targets:
+            ct.requires_grad_(False)
+
+    def _move_to_device(self) -> None:
+        self.critics.to(self._device)
+        self.critic_targets.to(self._device)
+
+    # ── Offline phase (Stage 1): RL-token training ───────────────────
+
+    def supports_offline_phase(self) -> bool:
+        return True
+
+    def offline_update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """Train RL-token encoder/decoder on demonstration data.
+
+        Paper Eq. 2: L_ro = E[ sum_i || h(d([z_rl, z_bar_{1:i-1}]))_i - z_bar_i ||^2 ]
+        """
+        batch = next(batch_iterator)
+
+        vla_embeddings = batch["state"]["observation.vla_embeddings"].to(self._device)
+        z_vla = vla_embeddings.detach()  # stop-gradient on VLA embeddings
+
+        z_rl = self.policy.rl_token_encoder(z_vla)
+        z_reconstructed = self.policy.rl_token_decoder(z_rl, z_vla)
+
+        loss_ro = F.mse_loss(z_reconstructed, z_vla)
+
+        self.optimizers["rl_token"].zero_grad()
+        loss_ro.backward()
+        torch.nn.utils.clip_grad_norm_(
+            list(self.policy.rl_token_encoder.parameters()) + list(self.policy.rl_token_decoder.parameters()),
+            max_norm=self.config.clip_grad_norm,
+        )
+        self.optimizers["rl_token"].step()
+
+        self._optimization_step += 1
+        return TrainingStats(losses={"loss_rl_token": loss_ro.item()})
+
+    def transition_to_online(self) -> None:
+        """Freeze RL-token modules; rebuild optimizers for actor-critic only."""
+        self.policy.rl_token_encoder.requires_grad_(False)
+        self.policy.rl_token_decoder.requires_grad_(False)
+        self._is_online = True
+
+        self.optimizers = {
+            "actor": torch.optim.Adam(self.policy.actor.parameters(), lr=self.config.actor_lr),
+            "critic": torch.optim.Adam(self.critics.parameters(), lr=self.config.critic_lr),
+        }
+        self._optimization_step = 0
+
+    # ── Online phase (Stage 2): Actor-Critic ─────────────────────────
+
+    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """One full RLT update step with UTD critic warm-up.
+
+        Pulls ``utd_ratio`` batches. First ``utd_ratio - 1`` are critic-only;
+        the last batch also updates the actor (every ``policy_update_freq`` steps).
+        """
+        for _ in range(self.config.utd_ratio - 1):
+            batch = next(batch_iterator)
+            fb = self._prepare_forward_batch(batch)
+            self._critic_step(fb)
+            self._update_target_networks()
+
+        batch = next(batch_iterator)
+        fb = self._prepare_forward_batch(batch)
+        critic_loss = self._critic_step(fb)
+
+        stats = TrainingStats(losses={"loss_critic": critic_loss})
+
+        if self._optimization_step % self.config.policy_update_freq == 0:
+            actor_loss, bc_loss, q_val = self._actor_step(fb)
+            stats.losses["loss_actor"] = actor_loss
+            stats.extra["bc_loss"] = bc_loss
+            stats.extra["q_value_mean"] = q_val
+
+        self._update_target_networks()
+        self._optimization_step += 1
+        return stats
+
+    def _prepare_forward_batch(self, batch: BatchType) -> dict[str, Any]:
+        """Convert a replay batch into algorithm-ready tensors.
+
+        Extracts RL-token from VLA embeddings, builds RL state, reads
+        reference action from complementary_info.
+        """
+        obs = batch["state"]
+        next_obs = batch["next_state"]
+        device = self._device
+
+        vla_emb = obs["observation.vla_embeddings"].to(device)
+        next_vla_emb = next_obs["observation.vla_embeddings"].to(device)
+
+        with torch.no_grad():
+            z_rl = self.policy.rl_token_encoder(vla_emb)
+            z_rl_next = self.policy.rl_token_encoder(next_vla_emb)
+
+        parts = [z_rl]
+        next_parts = [z_rl_next]
+        if "observation.state" in obs and self.policy._proprioception_dim > 0:
+            prop = obs["observation.state"].to(device)
+            next_prop = next_obs["observation.state"].to(device)
+            parts.append(prop)
+            next_parts.append(next_prop)
+
+        state = torch.cat(parts, dim=-1)
+        next_state = torch.cat(next_parts, dim=-1)
+
+        action = batch[ACTION].to(device)
+        reward = batch["reward"].to(device)
+        done = batch["done"].to(device)
+
+        ref_action = None
+        comp_info = batch.get("complementary_info")
+        if comp_info is not None and "reference_action" in comp_info:
+            ref_action = comp_info["reference_action"].to(device)
+
+        return {
+            "state": state,
+            "next_state": next_state,
+            "action": action,
+            "reward": reward,
+            "done": done,
+            "reference_action": ref_action,
+        }
+
+    def _critic_step(self, fb: dict[str, Any]) -> float:
+        """Paper Eq. 3: chunked TD with clipped double-Q target."""
+        state = fb["state"]
+        next_state = fb["next_state"]
+        action = fb["action"]
+        reward = fb["reward"]
+        done = fb["done"]
+
+        with torch.no_grad():
+            ref = fb.get("reference_action")
+            if ref is None:
+                ref = torch.zeros_like(action)
+            next_action = self.policy.actor(next_state, ref)
+
+            target_qs = [ct(next_state, next_action) for ct in self.critic_targets]
+            min_target_q = torch.min(torch.cat(target_qs, dim=-1), dim=-1, keepdim=True).values
+
+            discount_chunk = self.config.discount**self.config.chunk_size
+            td_target = reward.unsqueeze(-1) + (1 - done.unsqueeze(-1)) * discount_chunk * min_target_q
+
+        q_preds = [c(state, action) for c in self.critics]
+        loss = sum(F.mse_loss(q, td_target) for q in q_preds)
+
+        self.optimizers["critic"].zero_grad()
+        loss.backward()
+        torch.nn.utils.clip_grad_norm_(self.critics.parameters(), max_norm=self.config.clip_grad_norm)
+        self.optimizers["critic"].step()
+        return loss.item()
+
+    def _actor_step(self, fb: dict[str, Any]) -> tuple[float, float, float]:
+        """Paper Eq. 5: maximize Q while staying near VLA reference.
+
+        L_pi(theta) = E[ -Q(x, a) + beta * ||a - a_tilde||^2 ]
+        With reference-action dropout applied to the actor's ref input.
+        """
+        state = fb["state"]
+        ref = fb.get("reference_action")
+        if ref is None:
+            ref = torch.zeros(state.shape[0], self.policy._action_chunk_dim, device=self._device)
+
+        # Reference-action dropout (paper Section IV-B)
+        mask = (torch.rand(ref.shape[0], 1, device=self._device) > self.config.ref_dropout).float()
+        ref_input = ref * mask
+
+        action = self.policy.actor(state, ref_input)
+
+        q_value = self.critics[0](state, action)
+
+        bc_loss = F.mse_loss(action, ref)
+
+        loss = -q_value.mean() + self.config.bc_reg_coeff * bc_loss
+
+        self.optimizers["actor"].zero_grad()
+        loss.backward()
+        torch.nn.utils.clip_grad_norm_(self.policy.actor.parameters(), max_norm=self.config.clip_grad_norm)
+        self.optimizers["actor"].step()
+
+        return loss.item(), bc_loss.item(), q_value.mean().item()
+
+    def _update_target_networks(self) -> None:
+        tau = self.config.tau
+        for critic, target in zip(self.critics, self.critic_targets, strict=True):
+            for p, tp in zip(critic.parameters(), target.parameters(), strict=True):
+                tp.data.copy_(tau * p.data + (1 - tau) * tp.data)
+
+    # ── Optimizer management ─────────────────────────────────────────
+
+    def make_optimizers(self) -> dict[str, Optimizer]:
+        """Create optimizers. Initially for RL-token (Stage 1)."""
+        self.optimizers = {
+            "rl_token": torch.optim.Adam(
+                list(self.policy.rl_token_encoder.parameters())
+                + list(self.policy.rl_token_decoder.parameters()),
+                lr=self.config.rl_token_lr,
+            ),
+            "actor": torch.optim.Adam(self.policy.actor.parameters(), lr=self.config.actor_lr),
+            "critic": torch.optim.Adam(self.critics.parameters(), lr=self.config.critic_lr),
+        }
+        return self.optimizers
+
+    def get_optimizers(self) -> dict[str, Optimizer]:
+        return self.optimizers
+
+    # ── Weight sync ──────────────────────────────────────────────────
+
+    def get_weights(self) -> dict[str, Any]:
+        """Push actor + RL-token encoder to actors (small footprint)."""
+        weights = {
+            "actor": self.policy.actor.state_dict(),
+            "rl_token_encoder": self.policy.rl_token_encoder.state_dict(),
+        }
+        return {k: {kk: vv.cpu() for kk, vv in v.items()} for k, v in weights.items()}
+
+    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
+        if "actor" in weights:
+            self.policy.actor.load_state_dict({k: v.to(device) for k, v in weights["actor"].items()})
+        if "rl_token_encoder" in weights:
+            self.policy.rl_token_encoder.load_state_dict(
+                {k: v.to(device) for k, v in weights["rl_token_encoder"].items()}
+            )
--- a/src/lerobot/rl/algorithms/sac/init.py
+++ b/src/lerobot/rl/algorithms/sac/init.py
@@ -0,0 +1,18 @@
+# Copyright 2026 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.rl.algorithms.sac.configuration_sac import SACAlgorithmConfig
+from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
+
+__all__ = ["SACAlgorithm", "SACAlgorithmConfig"]
--- a/src/lerobot/rl/algorithms/sac/configuration_sac.py
+++ b/src/lerobot/rl/algorithms/sac/configuration_sac.py
@@ -0,0 +1,81 @@
+# Copyright 2026 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.
+"""SAC algorithm configuration."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING
+
+import torch
+
+from lerobot.policies.sac.configuration_sac import CriticNetworkConfig
+from lerobot.rl.algorithms.base import RLAlgorithmConfig
+
+if TYPE_CHECKING:
+    from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
+
+
+@RLAlgorithmConfig.register_subclass("sac")
+@dataclass
+class SACAlgorithmConfig(RLAlgorithmConfig):
+    """SAC-specific hyper-parameters that control the update loop."""
+
+    utd_ratio: int = 1
+    policy_update_freq: int = 1
+    clip_grad_norm: float = 40.0
+    actor_lr: float = 3e-4
+    critic_lr: float = 3e-4
+    temperature_lr: float = 3e-4
+    discount: float = 0.99
+    temperature_init: float = 1.0
+    target_entropy: float | None = None
+    use_backup_entropy: bool = True
+    critic_target_update_weight: float = 0.005
+    num_critics: int = 2
+    num_subsample_critics: int | None = None
+    num_discrete_actions: int | None = None
+    shared_encoder: bool = True
+    critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
+    discrete_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
+    use_torch_compile: bool = True
+
+    @classmethod
+    def from_policy_config(cls, policy_cfg) -> SACAlgorithmConfig:
+        """Build from an existing ``SACConfig`` (cfg.policy) for backwards compat."""
+        return cls(
+            utd_ratio=policy_cfg.utd_ratio,
+            policy_update_freq=policy_cfg.policy_update_freq,
+            clip_grad_norm=policy_cfg.grad_clip_norm,
+            actor_lr=policy_cfg.actor_lr,
+            critic_lr=policy_cfg.critic_lr,
+            temperature_lr=policy_cfg.temperature_lr,
+            discount=policy_cfg.discount,
+            temperature_init=policy_cfg.temperature_init,
+            target_entropy=policy_cfg.target_entropy,
+            use_backup_entropy=policy_cfg.use_backup_entropy,
+            critic_target_update_weight=policy_cfg.critic_target_update_weight,
+            num_critics=policy_cfg.num_critics,
+            num_subsample_critics=policy_cfg.num_subsample_critics,
+            num_discrete_actions=policy_cfg.num_discrete_actions,
+            shared_encoder=policy_cfg.shared_encoder,
+            critic_network_kwargs=policy_cfg.critic_network_kwargs,
+            discrete_critic_network_kwargs=policy_cfg.discrete_critic_network_kwargs,
+            use_torch_compile=policy_cfg.use_torch_compile,
+        )
+
+    def build_algorithm(self, policy: torch.nn.Module) -> SACAlgorithm:
+        from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
+
+        return SACAlgorithm(policy=policy, config=self)
--- a/src/lerobot/rl/algorithms/sac/sac_algorithm.py
+++ b/src/lerobot/rl/algorithms/sac/sac_algorithm.py
@@ -0,0 +1,409 @@
+# Copyright 2026 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.
+"""SAC (Soft Actor-Critic) algorithm.
+
+This module encapsulates all SAC-specific training logic (critic, actor,
+temperature, and discrete-critic updates) behind the ``RLAlgorithm`` interface.
+"""
+
+from __future__ import annotations
+
+import math
+from collections.abc import Iterator
+from dataclasses import asdict
+from typing import Any
+
+import einops
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from torch import Tensor
+from torch.optim import Optimizer
+
+from lerobot.policies.sac.modeling_sac import (
+    DISCRETE_DIMENSION_INDEX,
+    CriticEnsemble,
+    CriticHead,
+    DiscreteCritic,
+    SACObservationEncoder,
+    SACPolicy,
+)
+from lerobot.policies.utils import get_device_from_parameters
+from lerobot.rl.algorithms.base import (
+    BatchType,
+    RLAlgorithm,
+    TrainingStats,
+)
+from lerobot.rl.algorithms.sac.configuration_sac import SACAlgorithmConfig
+from lerobot.utils.constants import ACTION
+from lerobot.utils.transition import move_state_dict_to_device
+
+
+class SACAlgorithm(RLAlgorithm):
+    """Soft Actor-Critic with optional discrete-critic head.
+
+    Owns the ``SACPolicy`` and its optimizers.  All loss methods call
+    ``self.policy(batch_dict)`` rather than reaching into ``self.policy.actor``
+    directly, so any policy that returns ``{"action", "log_prob"}`` from its
+    ``forward()`` is compatible.
+    """
+
+    def __init__(
+        self,
+        policy: SACPolicy,
+        config: SACAlgorithmConfig,
+    ):
+        self.policy = policy
+        self.config = config
+        self.optimizers: dict[str, Optimizer] = {}
+        self._optimization_step: int = 0
+
+        self._device = get_device_from_parameters(self.policy)
+        self._init_critic_encoder()
+        self._init_critics()
+        self._init_temperature()
+        self._move_to_device()
+
+    def _init_critic_encoder(self) -> None:
+        """Build or share the encoder used by critics."""
+        if self.config.shared_encoder:
+            self.critic_encoder = self.policy.encoder
+            self.policy.actor.encoder_is_shared = True
+        else:
+            self.critic_encoder = SACObservationEncoder(self.policy.config)
+
+    def _init_critics(self) -> None:
+        """Build critic ensemble, targets, and optional discrete critic."""
+        action_dim = self.policy.config.output_features[ACTION].shape[0]
+        input_dim = self.critic_encoder.output_dim + action_dim
+
+        heads = [
+            CriticHead(input_dim=input_dim, **asdict(self.config.critic_network_kwargs))
+            for _ in range(self.config.num_critics)
+        ]
+        self.critic_ensemble = CriticEnsemble(encoder=self.critic_encoder, ensemble=heads)
+
+        target_heads = [
+            CriticHead(input_dim=input_dim, **asdict(self.config.critic_network_kwargs))
+            for _ in range(self.config.num_critics)
+        ]
+        self.critic_target = CriticEnsemble(encoder=self.critic_encoder, ensemble=target_heads)
+        self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
+
+        if self.config.use_torch_compile:
+            self.critic_ensemble = torch.compile(self.critic_ensemble)
+            self.critic_target = torch.compile(self.critic_target)
+
+        if self.config.num_discrete_actions is not None:
+            self._init_discrete_critic_target()
+
+    def _init_discrete_critic_target(self) -> None:
+        """Build only the target discrete critic."""
+        input_dim = self.critic_encoder.output_dim
+        self.discrete_critic_target = DiscreteCritic(
+            encoder=self.critic_encoder,
+            input_dim=input_dim,
+            output_dim=self.config.num_discrete_actions,
+            **asdict(self.config.discrete_critic_network_kwargs),
+        )
+        # TODO: (kmeftah) Compile the discrete critic
+        self.discrete_critic_target.load_state_dict(self.policy.discrete_critic.state_dict())
+
+    def _init_temperature(self) -> None:
+        """Set up temperature parameter (log_alpha) and default target entropy."""
+        temp_init = self.config.temperature_init
+        self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
+
+        action_dim = self.policy.config.output_features[ACTION].shape[0]
+        self.target_entropy = self.config.target_entropy
+        if self.target_entropy is None:
+            dim = action_dim + (1 if self.config.num_discrete_actions is not None else 0)
+            self.target_entropy = -np.prod(dim) / 2
+
+    def _move_to_device(self) -> None:
+        """Move algorithm-owned modules to the policy device."""
+        self.critic_ensemble.to(self._device)
+        self.critic_target.to(self._device)
+        self.log_alpha = nn.Parameter(self.log_alpha.data.to(self._device))
+        if hasattr(self, "discrete_critic_target"):
+            self.discrete_critic_target.to(self._device)
+
+    @property
+    def temperature(self) -> float:
+        return self.log_alpha.exp().item()
+
+    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """Run one full SAC update with UTD critic warm-up.
+
+        Pulls ``utd_ratio`` batches from ``batch_iterator``.  The first
+        ``utd_ratio - 1`` batches are used for critic-only warm-up steps;
+        the last batch drives the full update (critic + actor + temperature).
+        """
+        for _ in range(self.config.utd_ratio - 1):
+            batch = next(batch_iterator)
+            forward_batch = self._prepare_forward_batch(batch)
+
+            loss_critic = self._compute_loss_critic(forward_batch)
+            self.optimizers["critic"].zero_grad()
+            loss_critic.backward()
+            torch.nn.utils.clip_grad_norm_(
+                self.critic_ensemble.parameters(),
+                max_norm=self.config.clip_grad_norm,
+            ).item()
+            self.optimizers["critic"].step()
+
+            if self.config.num_discrete_actions is not None:
+                loss_discrete = self._compute_loss_discrete_critic(forward_batch)
+                self.optimizers["discrete_critic"].zero_grad()
+                loss_discrete.backward()
+                torch.nn.utils.clip_grad_norm_(
+                    self.policy.discrete_critic.parameters(),
+                    max_norm=self.config.clip_grad_norm,
+                ).item()
+                self.optimizers["discrete_critic"].step()
+            self._update_target_networks()
+
+        batch = next(batch_iterator)
+        forward_batch = self._prepare_forward_batch(batch)
+
+        loss_critic = self._compute_loss_critic(forward_batch)
+        self.optimizers["critic"].zero_grad()
+        loss_critic.backward()
+        critic_grad_norm = torch.nn.utils.clip_grad_norm_(
+            self.critic_ensemble.parameters(),
+            max_norm=self.config.clip_grad_norm,
+        ).item()
+        self.optimizers["critic"].step()
+
+        critic_loss_val = loss_critic.item()
+        stats = TrainingStats(
+            losses={"loss_critic": critic_loss_val},
+            grad_norms={"critic": critic_grad_norm},
+        )
+
+        if self.config.num_discrete_actions is not None:
+            loss_discrete = self._compute_loss_discrete_critic(forward_batch)
+            self.optimizers["discrete_critic"].zero_grad()
+            loss_discrete.backward()
+            dc_grad = torch.nn.utils.clip_grad_norm_(
+                self.policy.discrete_critic.parameters(),
+                max_norm=self.config.clip_grad_norm,
+            ).item()
+            self.optimizers["discrete_critic"].step()
+            stats.losses["loss_discrete_critic"] = loss_discrete.item()
+            stats.grad_norms["discrete_critic"] = dc_grad
+
+        if self._optimization_step % self.config.policy_update_freq == 0:
+            for _ in range(self.config.policy_update_freq):
+                actor_loss = self._compute_loss_actor(forward_batch)
+                self.optimizers["actor"].zero_grad()
+                actor_loss.backward()
+                actor_grad = torch.nn.utils.clip_grad_norm_(
+                    self.policy.actor.parameters(),
+                    max_norm=self.config.clip_grad_norm,
+                ).item()
+                self.optimizers["actor"].step()
+
+                temp_loss = self._compute_loss_temperature(forward_batch)
+                self.optimizers["temperature"].zero_grad()
+                temp_loss.backward()
+                temp_grad = torch.nn.utils.clip_grad_norm_(
+                    [self.log_alpha],
+                    max_norm=self.config.clip_grad_norm,
+                ).item()
+                self.optimizers["temperature"].step()
+
+            stats.losses["loss_actor"] = actor_loss.item()
+            stats.losses["loss_temperature"] = temp_loss.item()
+            stats.grad_norms["actor"] = actor_grad
+            stats.grad_norms["temperature"] = temp_grad
+            stats.extra["temperature"] = self.temperature
+
+        self._update_target_networks()
+
+        self._optimization_step += 1
+        return stats
+
+    def _compute_loss_critic(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        actions = batch[ACTION]
+        rewards = batch["reward"]
+        next_observations = batch["next_state"]
+        done = batch["done"]
+        obs_features = batch.get("observation_feature")
+        next_obs_features = batch.get("next_observation_feature")
+
+        with torch.no_grad():
+            next_output = self.policy({"state": next_observations, "observation_feature": next_obs_features})
+            next_actions = next_output["action"]
+            next_log_probs = next_output["log_prob"]
+
+            q_targets = self.critic_target(next_observations, next_actions, next_obs_features)
+
+            if self.config.num_subsample_critics is not None:
+                indices = torch.randperm(self.config.num_critics)
+                indices = indices[: self.config.num_subsample_critics]
+                q_targets = q_targets[indices]
+
+            min_q, _ = q_targets.min(dim=0)
+            if self.config.use_backup_entropy:
+                min_q = min_q - (self.temperature * next_log_probs)
+
+            td_target = rewards + (1 - done) * self.config.discount * min_q
+
+        if self.config.num_discrete_actions is not None:
+            actions = actions[:, :DISCRETE_DIMENSION_INDEX]
+
+        q_preds = self.critic_ensemble(observations, actions, obs_features)
+
+        td_target_dup = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
+        critics_loss = (F.mse_loss(input=q_preds, target=td_target_dup, reduction="none").mean(dim=1)).sum()
+        return critics_loss
+
+    def _compute_loss_discrete_critic(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        actions = batch[ACTION]
+        rewards = batch["reward"]
+        next_observations = batch["next_state"]
+        done = batch["done"]
+        obs_features = batch.get("observation_feature")
+        next_obs_features = batch.get("next_observation_feature")
+        complementary_info = batch.get("complementary_info")
+
+        actions_discrete: Tensor = actions[:, DISCRETE_DIMENSION_INDEX:].clone()
+        actions_discrete = torch.round(actions_discrete).long()
+
+        discrete_penalties: Tensor | None = None
+        if complementary_info is not None:
+            discrete_penalties = complementary_info.get("discrete_penalty")
+
+        with torch.no_grad():
+            next_discrete_qs = self.policy.discrete_critic(next_observations, next_obs_features)
+            best_next_action = torch.argmax(next_discrete_qs, dim=-1, keepdim=True)
+
+            target_next_qs = self.discrete_critic_target(next_observations, next_obs_features)
+            target_next_q = torch.gather(target_next_qs, dim=1, index=best_next_action).squeeze(-1)
+
+            rewards_disc = rewards
+            if discrete_penalties is not None:
+                rewards_disc = rewards + discrete_penalties
+            target_q = rewards_disc + (1 - done) * self.config.discount * target_next_q
+
+        predicted_qs = self.policy.discrete_critic(observations, obs_features)
+        predicted_q = torch.gather(predicted_qs, dim=1, index=actions_discrete).squeeze(-1)
+
+        return F.mse_loss(input=predicted_q, target=target_q)
+
+    def _compute_loss_actor(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        obs_features = batch.get("observation_feature")
+
+        output = self.policy({"state": observations, "observation_feature": obs_features})
+        actions_pi = output["action"]
+        log_probs = output["log_prob"]
+
+        q_preds = self.critic_ensemble(observations, actions_pi, obs_features)
+        min_q = q_preds.min(dim=0)[0]
+
+        return ((self.temperature * log_probs) - min_q).mean()
+
+    def _compute_loss_temperature(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        obs_features = batch.get("observation_feature")
+
+        with torch.no_grad():
+            output = self.policy({"state": observations, "observation_feature": obs_features})
+            log_probs = output["log_prob"]
+
+        return (-self.log_alpha.exp() * (log_probs + self.target_entropy)).mean()
+
+    def _update_target_networks(self) -> None:
+        tau = self.config.critic_target_update_weight
+        for target_p, p in zip(
+            self.critic_target.parameters(), self.critic_ensemble.parameters(), strict=True
+        ):
+            target_p.data.copy_(p.data * tau + target_p.data * (1.0 - tau))
+        if self.config.num_discrete_actions is not None:
+            for target_p, p in zip(
+                self.discrete_critic_target.parameters(),
+                self.policy.discrete_critic.parameters(),
+                strict=True,
+            ):
+                target_p.data.copy_(p.data * tau + target_p.data * (1.0 - tau))
+
+    def _prepare_forward_batch(self, batch: BatchType) -> dict[str, Any]:
+        """Build the dict expected by loss computation from a sampled batch."""
+        observations = batch["state"]
+        next_observations = batch["next_state"]
+
+        observation_features, next_observation_features = self.get_observation_features(
+            observations, next_observations
+        )
+        forward_batch: dict[str, Any] = {
+            ACTION: batch[ACTION],
+            "reward": batch["reward"],
+            "state": observations,
+            "next_state": next_observations,
+            "done": batch["done"],
+            "observation_feature": observation_features,
+            "next_observation_feature": next_observation_features,
+        }
+        if "complementary_info" in batch:
+            forward_batch["complementary_info"] = batch["complementary_info"]
+        return forward_batch
+
+    def make_optimizers(self) -> dict[str, Optimizer]:
+        """Create Adam optimizers for the SAC components and store them."""
+        actor_params = [
+            p
+            for n, p in self.policy.actor.named_parameters()
+            if not self.config.shared_encoder or not n.startswith("encoder")
+        ]
+        self.optimizers = {
+            "actor": torch.optim.Adam(actor_params, lr=self.config.actor_lr),
+            "critic": torch.optim.Adam(self.critic_ensemble.parameters(), lr=self.config.critic_lr),
+            "temperature": torch.optim.Adam([self.log_alpha], lr=self.config.temperature_lr),
+        }
+        if self.config.num_discrete_actions is not None:
+            self.optimizers["discrete_critic"] = torch.optim.Adam(
+                self.policy.discrete_critic.parameters(), lr=self.config.critic_lr
+            )
+        return self.optimizers
+
+    def get_optimizers(self) -> dict[str, Optimizer]:
+        return self.optimizers
+
+    def get_weights(self) -> dict[str, Any]:
+        """Policy state-dict to push to actors (includes actor + discrete critic)."""
+        return move_state_dict_to_device(self.policy.state_dict(), device="cpu")
+
+    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
+        """Load policy state-dict received from the learner."""
+        state = move_state_dict_to_device(weights, device=device)
+        self.policy.load_state_dict(state)
+
+    @torch.no_grad()
+    def get_observation_features(
+        self, observations: Tensor, next_observations: Tensor
+    ) -> tuple[Tensor | None, Tensor | None]:
+        if not self.config.shared_encoder:
+            return None, None
+        if self.policy.config.vision_encoder_name is None or not self.policy.config.freeze_vision_encoder:
+            return None, None
+        if not self.policy.encoder.has_images:
+            return None, None
+        observation_features = self.policy.encoder.get_cached_image_features(observations)
+        next_observation_features = self.policy.encoder.get_cached_image_features(next_observations)
+        return observation_features, next_observation_features
--- a/src/lerobot/rl/data_sources/init.py
+++ b/src/lerobot/rl/data_sources/init.py
@@ -0,0 +1,17 @@
+# Copyright 2026 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.rl.data_sources.data_mixer import BatchType, DataMixer, OnlineOfflineMixer
+
+__all__ = ["BatchType", "DataMixer", "OnlineOfflineMixer"]
--- a/src/lerobot/rl/data_sources/data_mixer.py
+++ b/src/lerobot/rl/data_sources/data_mixer.py
@@ -0,0 +1,94 @@
+# Copyright 2026 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 abc
+from typing import Any
+
+from lerobot.rl.buffer import ReplayBuffer, concatenate_batch_transitions
+
+BatchType = dict[str, Any]
+
+
+class DataMixer(abc.ABC):
+    """Abstract interface for all data mixing strategies.
+
+    Subclasses must implement ``sample(batch_size)`` and may override
+    ``get_iterator`` for specialised iteration.
+    """
+
+    @abc.abstractmethod
+    def sample(self, batch_size: int) -> BatchType:
+        """Draw one batch of ``batch_size`` transitions."""
+        ...
+
+    def get_iterator(
+        self,
+        batch_size: int,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ):
+        """Infinite iterator that yields batches.
+
+        The default implementation repeatedly calls ``self.sample()``.
+        Subclasses with underlying buffer iterators (async prefetch)
+        should override this for better throughput.
+        """
+        while True:
+            yield self.sample(batch_size)
+
+
+class OnlineOfflineMixer(DataMixer):
+    """Mixes transitions from an online and an optional offline replay buffer.
+
+    When both buffers are present, each batch is constructed by sampling
+    ``ceil(batch_size * online_ratio)`` from the online buffer and the
+    remainder from the offline buffer, then concatenating.
+
+    This mixer assumes both online and offline buffers are present.
+    """
+
+    def __init__(
+        self,
+        online_buffer: ReplayBuffer,
+        offline_buffer: ReplayBuffer | None = None,
+        online_ratio: float = 1.0,
+    ):
+        if not 0.0 <= online_ratio <= 1.0:
+            raise ValueError(f"online_ratio must be in [0, 1], got {online_ratio}")
+        self.online_buffer = online_buffer
+        self.offline_buffer = offline_buffer
+        self.online_ratio = online_ratio
+
+    def sample(self, batch_size: int) -> BatchType:
+        if self.offline_buffer is None:
+            return self.online_buffer.sample(batch_size)
+
+        n_online = max(1, int(batch_size * self.online_ratio))
+        n_offline = batch_size - n_online
+
+        online_batch = self.online_buffer.sample(n_online)
+        offline_batch = self.offline_buffer.sample(n_offline)
+        return concatenate_batch_transitions(online_batch, offline_batch)
+
+    def get_iterator(
+        self,
+        batch_size: int,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ):
+        """Yield batches from online/offline mixed sampling."""
+        while True:
+            yield self.sample(batch_size)
--- a/src/lerobot/rl/gym_manipulator.py
+++ b/src/lerobot/rl/gym_manipulator.py
@@ -36,6 +36,7 @@ from lerobot.processor import (
    DeviceProcessorStep,
    EnvTransition,
    GripperPenaltyProcessorStep,
+    GymHILAdapterProcessorStep,
    ImageCropResizeProcessorStep,
    InterventionActionProcessorStep,
    MapDeltaActionToRobotActionStep,
@@ -379,6 +380,7 @@ def make_processors(
        ]

        env_pipeline_steps = [
+            GymHILAdapterProcessorStep(),
            Numpy2TorchActionProcessorStep(),
            VanillaObservationProcessorStep(),
            AddBatchDimensionProcessorStep(),
@@ -412,7 +414,10 @@ def make_processors(
        if cfg.processor.observation.add_current_to_observation:
            env_pipeline_steps.append(MotorCurrentProcessorStep(robot=env.robot))

-    if kinematics_solver is not None:
+    add_ee_pose = (
+        cfg.processor.observation is not None and cfg.processor.observation.add_ee_pose_to_observation
+    )
+    if kinematics_solver is not None and add_ee_pose:
        env_pipeline_steps.append(
            ForwardKinematicsJointsToEEObservation(
                kinematics=kinematics_solver,
@@ -435,7 +440,12 @@ def make_processors(
        )

    # Add gripper penalty processor if gripper config exists and enabled
-    if cfg.processor.gripper is not None and cfg.processor.gripper.use_gripper:
+    # Only add if max_gripper_pos is explicitly configured (required for normalization)
+    if (
+        cfg.processor.gripper is not None
+        and cfg.processor.gripper.use_gripper
+        and cfg.processor.max_gripper_pos is not None
+    ):
        env_pipeline_steps.append(
            GripperPenaltyProcessorStep(
                penalty=cfg.processor.gripper.gripper_penalty,
@@ -600,7 +610,14 @@ def control_loop(

    dataset = None
    if cfg.mode == "record":
-        action_features = teleop_device.action_features
+        if teleop_device:
+            action_features = teleop_device.action_features
+        else:
+            action_features = {
+                "dtype": "float32",
+                "shape": (4,),
+                "names": ["delta_x", "delta_y", "delta_z", "gripper"],
+            }
        features = {
            ACTION: action_features,
            REWARD: {"dtype": "float32", "shape": (1,), "names": None},
@@ -648,7 +665,7 @@ def control_loop(
        # Create a neutral action (no movement)
        neutral_action = torch.tensor([0.0, 0.0, 0.0], dtype=torch.float32)
        if use_gripper:
-            neutral_action = torch.cat([neutral_action, torch.tensor([1.0])])  # Gripper stay
+            neutral_action = torch.cat([neutral_action, torch.tensor([0.0])])  # Gripper stay

        # Use the new step function
        transition = step_env_and_process_transition(
@@ -717,6 +734,8 @@ def control_loop(
        precise_sleep(max(dt - (time.perf_counter() - step_start_time), 0.0))

    if dataset is not None and cfg.dataset.push_to_hub:
+        logging.info("Finalizing dataset before pushing to hub")
+        dataset.finalize()
        logging.info("Pushing dataset to hub")
        dataset.push_to_hub()

--- a/src/lerobot/rl/learner.py
+++ b/src/lerobot/rl/learner.py
@@ -65,9 +65,11 @@ from lerobot.configs.train import TrainRLServerPipelineConfig
 from lerobot.datasets.factory import make_dataset
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.policies.factory import make_policy
-from lerobot.policies.sac.modeling_sac import SACPolicy
-from lerobot.rl.buffer import ReplayBuffer, concatenate_batch_transitions
+from lerobot.rl.algorithms import make_algorithm
+from lerobot.rl.buffer import ReplayBuffer
+from lerobot.rl.data_sources import OnlineOfflineMixer
 from lerobot.rl.process import ProcessSignalHandler
+from lerobot.rl.trainer import RLTrainer
 from lerobot.rl.wandb_utils import WandBLogger
 from lerobot.robots import so_follower  # noqa: F401
 from lerobot.teleoperators import gamepad, so_leader  # noqa: F401
@@ -93,7 +95,7 @@ from lerobot.utils.train_utils import (
    save_checkpoint,
    update_last_checkpoint,
 )
-from lerobot.utils.transition import move_state_dict_to_device, move_transition_to_device
+from lerobot.utils.transition import move_transition_to_device
 from lerobot.utils.utils import (
    format_big_number,
    get_safe_torch_device,
@@ -264,8 +266,8 @@ def add_actor_information_and_train(
    - Transfers transitions from the actor to the replay buffer.
    - Logs received interaction messages.
    - Ensures training begins only when the replay buffer has a sufficient number of transitions.
-    - Samples batches from the replay buffer and performs multiple critic updates.
-    - Periodically updates the actor, critic, and temperature optimizers.
+    - Delegates training updates to an ``RLAlgorithm`` (currently ``SACAlgorithm``).
+    - Periodically pushes updated weights to actors.
    - Logs training statistics, including loss values and optimization frequency.

    NOTE: This function doesn't have a single responsibility, it should be split into multiple functions
@@ -284,17 +286,15 @@ def add_actor_information_and_train(
    # of 7%
    device = get_safe_torch_device(try_device=cfg.policy.device, log=True)
    storage_device = get_safe_torch_device(try_device=cfg.policy.storage_device)
-    clip_grad_norm_value = cfg.policy.grad_clip_norm
    online_step_before_learning = cfg.policy.online_step_before_learning
-    utd_ratio = cfg.policy.utd_ratio
    fps = cfg.env.fps
    log_freq = cfg.log_freq
    save_freq = cfg.save_freq
-    policy_update_freq = cfg.policy.policy_update_freq
    policy_parameters_push_frequency = cfg.policy.actor_learner_config.policy_parameters_push_frequency
    saving_checkpoint = cfg.save_checkpoint
    online_steps = cfg.policy.online_steps
-    async_prefetch = cfg.policy.async_prefetch
+    async_prefetch = cfg.async_prefetch
+    queue_size = cfg.queue_size

    # Initialize logging for multiprocessing
    if not use_threads(cfg):
@@ -306,7 +306,7 @@ def add_actor_information_and_train(

    logging.info("Initializing policy")

-    policy: SACPolicy = make_policy(
+    policy = make_policy(
        cfg=cfg.policy,
        env_cfg=cfg.env,
    )
@@ -315,19 +315,24 @@ def add_actor_information_and_train(

    policy.train()

-    push_actor_policy_to_queue(parameters_queue=parameters_queue, policy=policy)
+    algorithm = make_algorithm(
+        policy=policy,
+        policy_cfg=cfg.policy,
+        algorithm_name=cfg.algorithm,
+    )

+    # TODO: Re-enable processor pipeline once refactoring is validated against main
+    preprocessor, postprocessor = None, None
+
+    # Push initial policy weights to actors (same path as periodic push)
+    state_bytes = state_to_bytes(algorithm.get_weights())
+    parameters_queue.put(state_bytes)
    last_time_policy_pushed = time.time()

-    optimizers, lr_scheduler = make_optimizers_and_scheduler(cfg=cfg, policy=policy)
-
-    # If we are resuming, we need to load the training state
-    resume_optimization_step, resume_interaction_step = load_training_state(cfg=cfg, optimizers=optimizers)
-
    log_training_info(cfg=cfg, policy=policy)

    replay_buffer = initialize_replay_buffer(cfg, device, storage_device)
-    batch_size = cfg.batch_size
+    total_batch_size = cfg.batch_size
    offline_replay_buffer = None

    if cfg.dataset is not None:
@@ -336,20 +341,70 @@ def add_actor_information_and_train(
            device=device,
            storage_device=storage_device,
        )
-        batch_size: int = batch_size // 2  # We will sample from both replay buffer
+
+    # DataMixer: online-only or online/offline 50-50 mix
+    data_mixer = OnlineOfflineMixer(
+        online_buffer=replay_buffer,
+        offline_buffer=offline_replay_buffer,
+        online_ratio=cfg.online_ratio,
+    )
+    # RLTrainer owns the iterator, preprocessor, and creates optimizers.
+    trainer = RLTrainer(
+        algorithm=algorithm,
+        data_mixer=data_mixer,
+        batch_size=total_batch_size,
+        preprocessor=preprocessor,
+        action_dim=cfg.policy.output_features["action"].shape[0],
+        async_prefetch=async_prefetch,
+        queue_size=queue_size,
+    )
+
+    # If we are resuming, we need to load the training state
+    optimizers = algorithm.get_optimizers()
+    resume_optimization_step, resume_interaction_step = load_training_state(cfg=cfg, optimizers=optimizers)

    logging.info("Starting learner thread")
    interaction_message = None
    optimization_step = resume_optimization_step if resume_optimization_step is not None else 0
+    algorithm.optimization_step = optimization_step
    interaction_step_shift = resume_interaction_step if resume_interaction_step is not None else 0

    dataset_repo_id = None
    if cfg.dataset is not None:
        dataset_repo_id = cfg.dataset.repo_id

-    # Initialize iterators
-    online_iterator = None
-    offline_iterator = None
+    # ── Offline phase (e.g. RLT RL-token training, ConRFT Cal-QL pretraining) ──
+    offline_steps = getattr(cfg.policy, "offline_steps", 0)
+    if algorithm.supports_offline_phase() and offline_steps > 0 and offline_replay_buffer is not None:
+        logging.info(f"[LEARNER] Starting offline phase ({offline_steps} steps)")
+        offline_mixer = OnlineOfflineMixer(
+            online_buffer=offline_replay_buffer,
+            offline_buffer=None,
+            online_ratio=1.0,
+        )
+        offline_iterator = algorithm.configure_data_iterator(
+            data_mixer=offline_mixer,
+            batch_size=total_batch_size,
+            async_prefetch=async_prefetch,
+            queue_size=queue_size,
+        )
+        for step in range(offline_steps):
+            if shutdown_event is not None and shutdown_event.is_set():
+                logging.info("[LEARNER] Shutdown during offline phase. Exiting...")
+                return
+
+            stats = algorithm.offline_update(offline_iterator)
+
+            if step % log_freq == 0:
+                logging.info(f"[LEARNER] Offline step {step}/{offline_steps}: {stats.to_log_dict()}")
+                if wandb_logger:
+                    log_dict = stats.to_log_dict()
+                    log_dict["offline_step"] = step
+                    wandb_logger.log_dict(d=log_dict, mode="train", custom_step_key="offline_step")
+
+        algorithm.transition_to_online()
+        optimizers = algorithm.get_optimizers()
+        logging.info("[LEARNER] Offline phase complete, transitioned to online")

    # NOTE: THIS IS THE MAIN LOOP OF THE LEARNER
    while True:
@@ -380,180 +435,22 @@ def add_actor_information_and_train(
        if len(replay_buffer) < online_step_before_learning:
            continue

-        if online_iterator is None:
-            online_iterator = replay_buffer.get_iterator(
-                batch_size=batch_size, async_prefetch=async_prefetch, queue_size=2
-            )
-
-        if offline_replay_buffer is not None and offline_iterator is None:
-            offline_iterator = offline_replay_buffer.get_iterator(
-                batch_size=batch_size, async_prefetch=async_prefetch, queue_size=2
-            )
-
        time_for_one_optimization_step = time.time()
-        for _ in range(utd_ratio - 1):
-            # Sample from the iterators
-            batch = next(online_iterator)

-            if dataset_repo_id is not None:
-                batch_offline = next(offline_iterator)
-                batch = concatenate_batch_transitions(
-                    left_batch_transitions=batch, right_batch_transition=batch_offline
-                )
-
-            actions = batch[ACTION]
-            rewards = batch["reward"]
-            observations = batch["state"]
-            next_observations = batch["next_state"]
-            done = batch["done"]
-            check_nan_in_transition(observations=observations, actions=actions, next_state=next_observations)
-
-            observation_features, next_observation_features = get_observation_features(
-                policy=policy, observations=observations, next_observations=next_observations
-            )
-
-            # Create a batch dictionary with all required elements for the forward method
-            forward_batch = {
-                ACTION: actions,
-                "reward": rewards,
-                "state": observations,
-                "next_state": next_observations,
-                "done": done,
-                "observation_feature": observation_features,
-                "next_observation_feature": next_observation_features,
-                "complementary_info": batch["complementary_info"],
-            }
-
-            # Use the forward method for critic loss
-            critic_output = policy.forward(forward_batch, model="critic")
-
-            # Main critic optimization
-            loss_critic = critic_output["loss_critic"]
-            optimizers["critic"].zero_grad()
-            loss_critic.backward()
-            critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-                parameters=policy.critic_ensemble.parameters(), max_norm=clip_grad_norm_value
-            )
-            optimizers["critic"].step()
-
-            # Discrete critic optimization (if available)
-            if policy.config.num_discrete_actions is not None:
-                discrete_critic_output = policy.forward(forward_batch, model="discrete_critic")
-                loss_discrete_critic = discrete_critic_output["loss_discrete_critic"]
-                optimizers["discrete_critic"].zero_grad()
-                loss_discrete_critic.backward()
-                discrete_critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-                    parameters=policy.discrete_critic.parameters(), max_norm=clip_grad_norm_value
-                )
-                optimizers["discrete_critic"].step()
-
-            # Update target networks (main and discrete)
-            policy.update_target_networks()
-
-        # Sample for the last update in the UTD ratio
-        batch = next(online_iterator)
-
-        if dataset_repo_id is not None:
-            batch_offline = next(offline_iterator)
-            batch = concatenate_batch_transitions(
-                left_batch_transitions=batch, right_batch_transition=batch_offline
-            )
-
-        actions = batch[ACTION]
-        rewards = batch["reward"]
-        observations = batch["state"]
-        next_observations = batch["next_state"]
-        done = batch["done"]
-
-        check_nan_in_transition(observations=observations, actions=actions, next_state=next_observations)
-
-        observation_features, next_observation_features = get_observation_features(
-            policy=policy, observations=observations, next_observations=next_observations
-        )
-
-        # Create a batch dictionary with all required elements for the forward method
-        forward_batch = {
-            ACTION: actions,
-            "reward": rewards,
-            "state": observations,
-            "next_state": next_observations,
-            "done": done,
-            "observation_feature": observation_features,
-            "next_observation_feature": next_observation_features,
-        }
-
-        critic_output = policy.forward(forward_batch, model="critic")
-
-        loss_critic = critic_output["loss_critic"]
-        optimizers["critic"].zero_grad()
-        loss_critic.backward()
-        critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-            parameters=policy.critic_ensemble.parameters(), max_norm=clip_grad_norm_value
-        ).item()
-        optimizers["critic"].step()
-
-        # Initialize training info dictionary
-        training_infos = {
-            "loss_critic": loss_critic.item(),
-            "critic_grad_norm": critic_grad_norm,
-        }
-
-        # Discrete critic optimization (if available)
-        if policy.config.num_discrete_actions is not None:
-            discrete_critic_output = policy.forward(forward_batch, model="discrete_critic")
-            loss_discrete_critic = discrete_critic_output["loss_discrete_critic"]
-            optimizers["discrete_critic"].zero_grad()
-            loss_discrete_critic.backward()
-            discrete_critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-                parameters=policy.discrete_critic.parameters(), max_norm=clip_grad_norm_value
-            ).item()
-            optimizers["discrete_critic"].step()
-
-            # Add discrete critic info to training info
-            training_infos["loss_discrete_critic"] = loss_discrete_critic.item()
-            training_infos["discrete_critic_grad_norm"] = discrete_critic_grad_norm
-
-        # Actor and temperature optimization (at specified frequency)
-        if optimization_step % policy_update_freq == 0:
-            for _ in range(policy_update_freq):
-                # Actor optimization
-                actor_output = policy.forward(forward_batch, model="actor")
-                loss_actor = actor_output["loss_actor"]
-                optimizers["actor"].zero_grad()
-                loss_actor.backward()
-                actor_grad_norm = torch.nn.utils.clip_grad_norm_(
-                    parameters=policy.actor.parameters(), max_norm=clip_grad_norm_value
-                ).item()
-                optimizers["actor"].step()
-
-                # Add actor info to training info
-                training_infos["loss_actor"] = loss_actor.item()
-                training_infos["actor_grad_norm"] = actor_grad_norm
-
-                # Temperature optimization
-                temperature_output = policy.forward(forward_batch, model="temperature")
-                loss_temperature = temperature_output["loss_temperature"]
-                optimizers["temperature"].zero_grad()
-                loss_temperature.backward()
-                temp_grad_norm = torch.nn.utils.clip_grad_norm_(
-                    parameters=[policy.log_alpha], max_norm=clip_grad_norm_value
-                ).item()
-                optimizers["temperature"].step()
-
-                # Add temperature info to training info
-                training_infos["loss_temperature"] = loss_temperature.item()
-                training_infos["temperature_grad_norm"] = temp_grad_norm
-                training_infos["temperature"] = policy.temperature
+        # One training step (trainer owns data_mixer iterator; algorithm owns UTD loop)
+        stats = trainer.training_step()

        # Push policy to actors if needed
        if time.time() - last_time_policy_pushed > policy_parameters_push_frequency:
-            push_actor_policy_to_queue(parameters_queue=parameters_queue, policy=policy)
+            state_dicts = algorithm.get_weights()
+            state_bytes = state_to_bytes(state_dicts)
+            parameters_queue.put(state_bytes)
            last_time_policy_pushed = time.time()

-        # Update target networks (main and discrete)
-        policy.update_target_networks()
+        training_infos = stats.to_log_dict()

        # Log training metrics at specified intervals
+        optimization_step = algorithm.optimization_step
        if optimization_step % log_freq == 0:
            training_infos["replay_buffer_size"] = len(replay_buffer)
            if offline_replay_buffer is not None:
@@ -581,7 +478,6 @@ def add_actor_information_and_train(
                custom_step_key="Optimization step",
            )

-        optimization_step += 1
        if optimization_step % log_freq == 0:
            logging.info(f"[LEARNER] Number of optimization step: {optimization_step}")

@@ -598,6 +494,8 @@ def add_actor_information_and_train(
                offline_replay_buffer=offline_replay_buffer,
                dataset_repo_id=dataset_repo_id,
                fps=fps,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
            )


@@ -682,6 +580,8 @@ def save_training_checkpoint(
    offline_replay_buffer: ReplayBuffer | None = None,
    dataset_repo_id: str | None = None,
    fps: int = 30,
+    preprocessor=None,
+    postprocessor=None,
 ) -> None:
    """
    Save training checkpoint and associated data.
@@ -705,6 +605,8 @@ def save_training_checkpoint(
        offline_replay_buffer: Optional offline replay buffer to save
        dataset_repo_id: Repository ID for dataset
        fps: Frames per second for dataset
+        preprocessor: Optional preprocessor pipeline to save
+        postprocessor: Optional postprocessor pipeline to save
    """
    logging.info(f"Checkpoint policy after step {optimization_step}")
    _num_digits = max(6, len(str(online_steps)))
@@ -721,6 +623,8 @@ def save_training_checkpoint(
        policy=policy,
        optimizer=optimizers,
        scheduler=None,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
    )

    # Save interaction step manually
@@ -758,58 +662,6 @@ def save_training_checkpoint(
    logging.info("Resume training")


-def make_optimizers_and_scheduler(cfg: TrainRLServerPipelineConfig, policy: nn.Module):
-    """
-    Creates and returns optimizers for the actor, critic, and temperature components of a reinforcement learning policy.
-
-    This function sets up Adam optimizers for:
-    - The **actor network**, ensuring that only relevant parameters are optimized.
-    - The **critic ensemble**, which evaluates the value function.
-    - The **temperature parameter**, which controls the entropy in soft actor-critic (SAC)-like methods.
-
-    It also initializes a learning rate scheduler, though currently, it is set to `None`.
-
-    NOTE:
-    - If the encoder is shared, its parameters are excluded from the actor's optimization process.
-    - The policy's log temperature (`log_alpha`) is wrapped in a list to ensure proper optimization as a standalone tensor.
-
-    Args:
-        cfg: Configuration object containing hyperparameters.
-        policy (nn.Module): The policy model containing the actor, critic, and temperature components.
-
-    Returns:
-        Tuple[Dict[str, torch.optim.Optimizer], Optional[torch.optim.lr_scheduler._LRScheduler]]:
-        A tuple containing:
-        - `optimizers`: A dictionary mapping component names ("actor", "critic", "temperature") to their respective Adam optimizers.
-        - `lr_scheduler`: Currently set to `None` but can be extended to support learning rate scheduling.
-
-    """
-    optimizer_actor = torch.optim.Adam(
-        params=[
-            p
-            for n, p in policy.actor.named_parameters()
-            if not policy.config.shared_encoder or not n.startswith("encoder")
-        ],
-        lr=cfg.policy.actor_lr,
-    )
-    optimizer_critic = torch.optim.Adam(params=policy.critic_ensemble.parameters(), lr=cfg.policy.critic_lr)
-
-    if cfg.policy.num_discrete_actions is not None:
-        optimizer_discrete_critic = torch.optim.Adam(
-            params=policy.discrete_critic.parameters(), lr=cfg.policy.critic_lr
-        )
-    optimizer_temperature = torch.optim.Adam(params=[policy.log_alpha], lr=cfg.policy.critic_lr)
-    lr_scheduler = None
-    optimizers = {
-        "actor": optimizer_actor,
-        "critic": optimizer_critic,
-        "temperature": optimizer_temperature,
-    }
-    if cfg.policy.num_discrete_actions is not None:
-        optimizers["discrete_critic"] = optimizer_discrete_critic
-    return optimizers, lr_scheduler
-
-
 # Training setup functions


@@ -1014,33 +866,6 @@ def initialize_offline_replay_buffer(
 # Utilities/Helpers functions


-def get_observation_features(
-    policy: SACPolicy, observations: torch.Tensor, next_observations: torch.Tensor
-) -> tuple[torch.Tensor | None, torch.Tensor | None]:
-    """
-    Get observation features from the policy encoder. It act as cache for the observation features.
-    when the encoder is frozen, the observation features are not updated.
-    We can save compute by caching the observation features.
-
-    Args:
-        policy: The policy model
-        observations: The current observations
-        next_observations: The next observations
-
-    Returns:
-        tuple: observation_features, next_observation_features
-    """
-
-    if policy.config.vision_encoder_name is None or not policy.config.freeze_vision_encoder:
-        return None, None
-
-    with torch.no_grad():
-        observation_features = policy.actor.encoder.get_cached_image_features(observations)
-        next_observation_features = policy.actor.encoder.get_cached_image_features(next_observations)
-
-    return observation_features, next_observation_features
-
-
 def use_threads(cfg: TrainRLServerPipelineConfig) -> bool:
    return cfg.policy.concurrency.learner == "threads"

@@ -1091,23 +916,6 @@ def check_nan_in_transition(
    return nan_detected


-def push_actor_policy_to_queue(parameters_queue: Queue, policy: nn.Module):
-    logging.debug("[LEARNER] Pushing actor policy to the queue")
-
-    # Create a dictionary to hold all the state dicts
-    state_dicts = {"policy": move_state_dict_to_device(policy.actor.state_dict(), device="cpu")}
-
-    # Add discrete critic if it exists
-    if hasattr(policy, "discrete_critic") and policy.discrete_critic is not None:
-        state_dicts["discrete_critic"] = move_state_dict_to_device(
-            policy.discrete_critic.state_dict(), device="cpu"
-        )
-        logging.debug("[LEARNER] Including discrete critic in state dict push")
-
-    state_bytes = state_to_bytes(state_dicts)
-    parameters_queue.put(state_bytes)
-
-
 def process_interaction_message(
    message, interaction_step_shift: int, wandb_logger: WandBLogger | None = None
 ):
--- a/src/lerobot/rl/trainer.py
+++ b/src/lerobot/rl/trainer.py
@@ -0,0 +1,132 @@
+# Copyright 2026 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
+
+from collections.abc import Iterator
+from typing import Any
+
+import torch
+
+from lerobot.rl.algorithms.base import (
+    BatchType,
+    RLAlgorithm,
+    TrainingStats,
+)
+from lerobot.rl.data_sources.data_mixer import DataMixer
+from lerobot.utils.constants import ACTION
+
+
+def preprocess_rl_batch(preprocessor: Any, batch: BatchType, *, action_dim: int | None = None) -> BatchType:
+    """Apply a policy preprocessor to an RL batch."""
+    observations = batch["state"]
+    next_observations = batch["next_state"]
+    actions = batch[ACTION]
+
+    extra_action = None
+    if action_dim is not None and actions.shape[-1] > action_dim:
+        extra_action = actions[..., action_dim:]
+        actions = actions[..., :action_dim]
+
+    obs_action = {**observations, ACTION: actions}
+    obs_action = preprocessor(obs_action)
+    batch["state"] = {k: v for k, v in obs_action.items() if k.startswith("observation.")}
+    batch[ACTION] = obs_action[ACTION]
+
+    if extra_action is not None:
+        batch[ACTION] = torch.cat([batch[ACTION], extra_action], dim=-1)
+
+    next_obs = {**next_observations}
+    next_obs = preprocessor(next_obs)
+    batch["next_state"] = {k: v for k, v in next_obs.items() if k.startswith("observation.")}
+
+    return batch
+
+
+class _PreprocessedIterator:
+    """Iterator wrapper that preprocesses each sampled RL batch."""
+
+    __slots__ = ("_raw", "_preprocessor", "_action_dim")
+
+    def __init__(
+        self, raw_iterator: Iterator[BatchType], preprocessor: Any, action_dim: int | None = None
+    ) -> None:
+        self._raw = raw_iterator
+        self._preprocessor = preprocessor
+        self._action_dim = action_dim
+
+    def __iter__(self) -> _PreprocessedIterator:
+        return self
+
+    def __next__(self) -> BatchType:
+        batch = next(self._raw)
+        return preprocess_rl_batch(self._preprocessor, batch, action_dim=self._action_dim)
+
+
+class RLTrainer:
+    """Unified training step orchestrator.
+
+    Holds the algorithm, a DataMixer, and an optional preprocessor.
+    """
+
+    def __init__(
+        self,
+        algorithm: RLAlgorithm,
+        data_mixer: DataMixer,
+        batch_size: int,
+        *,
+        preprocessor: Any | None = None,
+        action_dim: int | None = None,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ):
+        self.algorithm = algorithm
+        self.data_mixer = data_mixer
+        self.batch_size = batch_size
+        self._preprocessor = preprocessor
+        self._action_dim = action_dim
+        self.async_prefetch = async_prefetch
+        self.queue_size = queue_size
+
+        self._iterator: Iterator[BatchType] | None = None
+
+        self.algorithm.make_optimizers()
+
+    def _build_data_iterator(self) -> Iterator[BatchType]:
+        """Create a fresh algorithm-configured iterator (optionally preprocessed)."""
+        raw = self.algorithm.configure_data_iterator(
+            data_mixer=self.data_mixer,
+            batch_size=self.batch_size,
+            async_prefetch=self.async_prefetch,
+            queue_size=self.queue_size,
+        )
+        if self._preprocessor is not None:
+            return _PreprocessedIterator(raw, self._preprocessor, self._action_dim)
+        return raw
+
+    def reset_data_iterator(self) -> None:
+        """Discard the current iterator so it will be rebuilt lazily next step."""
+        self._iterator = None
+
+    def set_data_mixer(self, data_mixer: DataMixer, *, reset: bool = True) -> None:
+        """Swap the active data mixer, optionally resetting the iterator."""
+        self.data_mixer = data_mixer
+        if reset:
+            self.reset_data_iterator()
+
+    def training_step(self) -> TrainingStats:
+        """Run one training step (algorithm-agnostic)."""
+        if self._iterator is None:
+            self._iterator = self._build_data_iterator()
+        return self.algorithm.update(self._iterator)
--- a/src/lerobot/rl/wandb_utils.py
+++ b/src/lerobot/rl/wandb_utils.py
@@ -26,8 +26,21 @@ from lerobot.configs.train import TrainPipelineConfig
 from lerobot.utils.constants import PRETRAINED_MODEL_DIR


-def cfg_to_group(cfg: TrainPipelineConfig, return_list: bool = False) -> list[str] | str:
+def cfg_to_group(
+    cfg: TrainPipelineConfig, return_list: bool = False, truncate_tags: bool = False, max_tag_length: int = 64
+) -> list[str] | str:
    """Return a group name for logging. Optionally returns group name as list."""
+
+    def _maybe_truncate(tag: str) -> str:
+        """Truncate tag to max_tag_length characters if required.
+
+        wandb rejects tags longer than 64 characters.
+        See: https://github.com/wandb/wandb/blob/main/wandb/sdk/wandb_settings.py
+        """
+        if len(tag) <= max_tag_length:
+            return tag
+        return tag[:max_tag_length]
+
    lst = [
        f"policy:{cfg.policy.type}",
        f"seed:{cfg.seed}",
@@ -36,6 +49,8 @@ def cfg_to_group(cfg: TrainPipelineConfig, return_list: bool = False) -> list[st
        lst.append(f"dataset:{cfg.dataset.repo_id}")
    if cfg.env is not None:
        lst.append(f"env:{cfg.env.type}")
+    if truncate_tags:
+        lst = [_maybe_truncate(tag) for tag in lst]
    return lst if return_list else "-".join(lst)


@@ -83,7 +98,7 @@ class WandBLogger:
            entity=self.cfg.entity,
            name=self.job_name,
            notes=self.cfg.notes,
-            tags=cfg_to_group(cfg, return_list=True),
+            tags=cfg_to_group(cfg, return_list=True, truncate_tags=True),
            dir=self.log_dir,
            config=cfg.to_dict(),
            # TODO(rcadene): try set to True
--- a/src/lerobot/robots/bi_openarm_follower/bi_openarm_follower.py
+++ b/src/lerobot/robots/bi_openarm_follower/bi_openarm_follower.py
@@ -19,6 +19,7 @@ from functools import cached_property

 from lerobot.processor import RobotAction, RobotObservation
 from lerobot.robots.openarm_follower import OpenArmFollower, OpenArmFollowerConfig
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected

 from ..robot import Robot
 from .config_bi_openarm_follower import BiOpenArmFollowerConfig
@@ -112,6 +113,7 @@ class BiOpenArmFollower(Robot):
    def is_connected(self) -> bool:
        return self.left_arm.is_connected and self.right_arm.is_connected

+    @check_if_already_connected
    def connect(self, calibrate: bool = True) -> None:
        self.left_arm.connect(calibrate)
        self.right_arm.connect(calibrate)
@@ -133,6 +135,7 @@ class BiOpenArmFollower(Robot):
            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
        )

+    @check_if_not_connected
    def get_observation(self) -> RobotObservation:
        obs_dict = {}

@@ -146,6 +149,7 @@ class BiOpenArmFollower(Robot):

        return obs_dict

+    @check_if_not_connected
    def send_action(
        self,
        action: RobotAction,
@@ -170,6 +174,7 @@ class BiOpenArmFollower(Robot):

        return {**prefixed_sent_action_left, **prefixed_sent_action_right}

+    @check_if_not_connected
    def disconnect(self):
        self.left_arm.disconnect()
        self.right_arm.disconnect()
--- a/src/lerobot/robots/bi_so_follower/bi_so_follower.py
+++ b/src/lerobot/robots/bi_so_follower/bi_so_follower.py
@@ -19,6 +19,7 @@ from functools import cached_property

 from lerobot.processor import RobotAction, RobotObservation
 from lerobot.robots.so_follower import SOFollower, SOFollowerRobotConfig
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected

 from ..robot import Robot
 from .config_bi_so_follower import BiSOFollowerConfig
@@ -96,6 +97,7 @@ class BiSOFollower(Robot):
    def is_connected(self) -> bool:
        return self.left_arm.is_connected and self.right_arm.is_connected

+    @check_if_already_connected
    def connect(self, calibrate: bool = True) -> None:
        self.left_arm.connect(calibrate)
        self.right_arm.connect(calibrate)
@@ -116,6 +118,7 @@ class BiSOFollower(Robot):
        self.left_arm.setup_motors()
        self.right_arm.setup_motors()

+    @check_if_not_connected
    def get_observation(self) -> RobotObservation:
        obs_dict = {}

@@ -129,6 +132,7 @@ class BiSOFollower(Robot):

        return obs_dict

+    @check_if_not_connected
    def send_action(self, action: RobotAction) -> RobotAction:
        # Remove "left_" prefix
        left_action = {
@@ -148,6 +152,7 @@ class BiSOFollower(Robot):

        return {**prefixed_sent_action_left, **prefixed_sent_action_right}

+    @check_if_not_connected
    def disconnect(self):
        self.left_arm.disconnect()
        self.right_arm.disconnect()
--- a/src/lerobot/robots/hope_jr/hope_jr_arm.py
+++ b/src/lerobot/robots/hope_jr/hope_jr_arm.py
@@ -140,7 +140,7 @@ class HopeJrArm(Robot):
        # Capture images from cameras
        for cam_key, cam in self.cameras.items():
            start = time.perf_counter()
-            obs_dict[cam_key] = cam.async_read()
+            obs_dict[cam_key] = cam.read_latest()
            dt_ms = (time.perf_counter() - start) * 1e3
            logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")

--- a/src/lerobot/robots/hope_jr/hope_jr_hand.py
+++ b/src/lerobot/robots/hope_jr/hope_jr_hand.py
@@ -171,7 +171,7 @@ class HopeJrHand(Robot):
        # Capture images from cameras
        for cam_key, cam in self.cameras.items():
            start = time.perf_counter()
-            obs_dict[cam_key] = cam.async_read()
+            obs_dict[cam_key] = cam.read_latest()
            dt_ms = (time.perf_counter() - start) * 1e3
            logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")

--- a/src/lerobot/robots/koch_follower/koch_follower.py
+++ b/src/lerobot/robots/koch_follower/koch_follower.py
@@ -193,7 +193,7 @@ class KochFollower(Robot):
        # Capture images from cameras
        for cam_key, cam in self.cameras.items():
            start = time.perf_counter()
-            obs_dict[cam_key] = cam.async_read()
+            obs_dict[cam_key] = cam.read_latest()
            dt_ms = (time.perf_counter() - start) * 1e3
            logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")

--- a/src/lerobot/robots/lekiwi/lekiwi.py
+++ b/src/lerobot/robots/lekiwi/lekiwi.py
@@ -360,7 +360,7 @@ class LeKiwi(Robot):
        # Capture images from cameras
        for cam_key, cam in self.cameras.items():
            start = time.perf_counter()
-            obs_dict[cam_key] = cam.async_read()
+            obs_dict[cam_key] = cam.read_latest()
            dt_ms = (time.perf_counter() - start) * 1e3
            logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")

--- a/src/lerobot/robots/omx_follower/omx_follower.py
+++ b/src/lerobot/robots/omx_follower/omx_follower.py
@@ -176,7 +176,7 @@ class OmxFollower(Robot):
        # Capture images from cameras
        for cam_key, cam in self.cameras.items():
            start = time.perf_counter()
-            obs_dict[cam_key] = cam.async_read()
+            obs_dict[cam_key] = cam.read_latest()
            dt_ms = (time.perf_counter() - start) * 1e3
            logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")

--- a/src/lerobot/robots/openarm_follower/openarm_follower.py
+++ b/src/lerobot/robots/openarm_follower/openarm_follower.py
@@ -23,7 +23,7 @@ from lerobot.cameras.utils import make_cameras_from_configs
 from lerobot.motors import Motor, MotorCalibration, MotorNormMode
 from lerobot.motors.damiao import DamiaoMotorsBus
 from lerobot.processor import RobotAction, RobotObservation
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected

 from ..robot import Robot
 from ..utils import ensure_safe_goal_position
@@ -119,6 +119,7 @@ class OpenArmFollower(Robot):
        """Check if robot is connected."""
        return self.bus.is_connected and all(cam.is_connected for cam in self.cameras.values())

+    @check_if_already_connected
    def connect(self, calibrate: bool = True) -> None:
        """
        Connect to the robot and optionally calibrate.
@@ -126,8 +127,6 @@ class OpenArmFollower(Robot):
        We assume that at connection time, the arms are in a safe rest position,
        and torque can be safely disabled to run calibration if needed.
        """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} already connected")

        # Connect to CAN bus
        logger.info(f"Connecting arm on {self.config.port}...")
@@ -219,6 +218,7 @@ class OpenArmFollower(Robot):
            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
        )

+    @check_if_not_connected
    def get_observation(self) -> RobotObservation:
        """
        Get current observation from robot including position, velocity, and torque.
@@ -228,9 +228,6 @@ class OpenArmFollower(Robot):
        """
        start = time.perf_counter()

-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
        obs_dict: dict[str, Any] = {}

        states = self.bus.sync_read_all_states()
@@ -244,7 +241,7 @@ class OpenArmFollower(Robot):
        # Capture images from cameras
        for cam_key, cam in self.cameras.items():
            start = time.perf_counter()
-            obs_dict[cam_key] = cam.async_read()
+            obs_dict[cam_key] = cam.read_latest()
            dt_ms = (time.perf_counter() - start) * 1e3
            logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")

@@ -253,6 +250,7 @@ class OpenArmFollower(Robot):

        return obs_dict

+    @check_if_not_connected
    def send_action(
        self,
        action: RobotAction,
@@ -272,8 +270,6 @@ class OpenArmFollower(Robot):
        Returns:
            The action actually sent (potentially clipped)
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        goal_pos = {key.removesuffix(".pos"): val for key, val in action.items() if key.endswith(".pos")}

@@ -333,10 +329,9 @@ class OpenArmFollower(Robot):

        return {f"{motor}.pos": val for motor, val in goal_pos.items()}

+    @check_if_not_connected
    def disconnect(self):
        """Disconnect from robot."""
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        # Disconnect CAN bus
        self.bus.disconnect(self.config.disable_torque_on_disconnect)
--- a/src/lerobot/robots/reachy2/robot_reachy2.py
+++ b/src/lerobot/robots/reachy2/robot_reachy2.py
@@ -180,7 +180,7 @@ class Reachy2Robot(Robot):

        # Capture images from cameras
        for cam_key, cam in self.cameras.items():
-            obs_dict[cam_key] = cam.async_read()
+            obs_dict[cam_key] = cam.read_latest()

        return obs_dict

--- a/src/lerobot/robots/so_follower/config_so_follower.py
+++ b/src/lerobot/robots/so_follower/config_so_follower.py
@@ -40,7 +40,7 @@ class SOFollowerConfig:
    cameras: dict[str, CameraConfig] = field(default_factory=dict)

    # Set to `True` for backward compatibility with previous policies/dataset
-    use_degrees: bool = False
+    use_degrees: bool = True


@RobotConfig.register_subclass("so101_follower")
--- a/src/lerobot/robots/so_follower/so_follower.py
+++ b/src/lerobot/robots/so_follower/so_follower.py
@@ -187,7 +187,7 @@ class SOFollower(Robot):
        # Capture images from cameras
        for cam_key, cam in self.cameras.items():
            start = time.perf_counter()
-            obs_dict[cam_key] = cam.async_read()
+            obs_dict[cam_key] = cam.read_latest()
            dt_ms = (time.perf_counter() - start) * 1e3
            logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")

--- a/src/lerobot/robots/unitree_g1/unitree_g1.py
+++ b/src/lerobot/robots/unitree_g1/unitree_g1.py
@@ -324,7 +324,7 @@ class UnitreeG1(Robot):

        # Cameras - read images from ZMQ cameras
        for cam_name, cam in self._cameras.items():
-            obs[cam_name] = cam.async_read()
+            obs[cam_name] = cam.read_latest()

        return obs

--- a/src/lerobot/scripts/lerobot_dataset_viz.py
+++ b/src/lerobot/scripts/lerobot_dataset_viz.py
@@ -47,16 +47,14 @@ local$ rerun lerobot_pusht_episode_0.rrd
 ```

 - Visualize data stored on a distant machine through streaming:
-(You need to forward the websocket port to the distant machine, with
-`ssh -L 9087:localhost:9087 username@remote-host`)
 ```
 distant$ lerobot-dataset-viz \
    --repo-id lerobot/pusht \
    --episode-index 0 \
    --mode distant \
-    --ws-port 9087
+    --grpc-port 9876

-local$ rerun ws://localhost:9087
+local$ rerun rerun+http://IP:GRPC_PORT/proxy
 ```

 """
@@ -75,6 +73,7 @@ import tqdm

 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.utils.constants import ACTION, DONE, OBS_STATE, REWARD
+from lerobot.utils.utils import init_logging


 def to_hwc_uint8_numpy(chw_float32_torch: torch.Tensor) -> np.ndarray:
@@ -93,10 +92,11 @@ def visualize_dataset(
    num_workers: int = 0,
    mode: str = "local",
    web_port: int = 9090,
-    ws_port: int = 9087,
+    grpc_port: int = 9876,
    save: bool = False,
    output_dir: Path | None = None,
    display_compressed_images: bool = False,
+    **kwargs,
 ) -> Path | None:
    if save:
        assert output_dir is not None, (
@@ -126,7 +126,9 @@ def visualize_dataset(
    gc.collect()

    if mode == "distant":
-        rr.serve_web_viewer(open_browser=False, web_port=web_port)
+        server_uri = rr.serve_grpc(grpc_port=grpc_port)
+        logging.info(f"Connect to a Rerun Server: rerun rerun+http://IP:{grpc_port}/proxy")
+        rr.serve_web_viewer(open_browser=False, web_port=web_port, connect_to=server_uri)

    logging.info("Logging to Rerun")

@@ -226,7 +228,7 @@ def main():
            "Mode of viewing between 'local' or 'distant'. "
            "'local' requires data to be on a local machine. It spawns a viewer to visualize the data locally. "
            "'distant' creates a server on the distant machine where the data is stored. "
-            "Visualize the data by connecting to the server with `rerun ws://localhost:PORT` on the local machine."
+            "Visualize the data by connecting to the server with `rerun rerun+http://IP:GRPC_PORT/proxy` on the local machine."
        ),
    )
    parser.add_argument(
@@ -238,8 +240,13 @@ def main():
    parser.add_argument(
        "--ws-port",
        type=int,
-        default=9087,
-        help="Web socket port for rerun.io when `--mode distant` is set.",
+        help="deprecated, please use --grpc-port instead.",
+    )
+    parser.add_argument(
+        "--grpc-port",
+        type=int,
+        default=9876,
+        help="gRPC port for rerun.io when `--mode distant` is set.",
    )
    parser.add_argument(
        "--save",
@@ -265,9 +272,7 @@ def main():

    parser.add_argument(
        "--display-compressed-images",
-        type=bool,
-        required=True,
-        default=False,
+        action="store_true",
        help="If set, display compressed images in Rerun instead of uncompressed ones.",
    )

@@ -277,6 +282,14 @@ def main():
    root = kwargs.pop("root")
    tolerance_s = kwargs.pop("tolerance_s")

+    if kwargs["ws_port"] is not None:
+        logging.warning(
+            "--ws-port is deprecated and will be removed in future versions. Please use --grpc-port instead."
+        )
+        logging.warning("Setting grpc_port to ws_port value.")
+        kwargs["grpc_port"] = kwargs.pop("ws_port")
+
+    init_logging()
    logging.info("Loading dataset")
    dataset = LeRobotDataset(repo_id, episodes=[args.episode_index], root=root, tolerance_s=tolerance_s)

--- a/src/lerobot/scripts/lerobot_edit_dataset.py
+++ b/src/lerobot/scripts/lerobot_edit_dataset.py
@@ -24,96 +24,112 @@ When new_repo_id is specified, creates a new dataset.
 Usage Examples:

 Delete episodes 0, 2, and 5 from a dataset:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht \
        --operation.type delete_episodes \
        --operation.episode_indices "[0, 2, 5]"

 Delete episodes and save to a new dataset:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht \
        --new_repo_id lerobot/pusht_filtered \
        --operation.type delete_episodes \
        --operation.episode_indices "[0, 2, 5]"

 Split dataset by fractions:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht \
        --operation.type split \
        --operation.splits '{"train": 0.8, "val": 0.2}'

 Split dataset by episode indices:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht \
        --operation.type split \
        --operation.splits '{"train": [0, 1, 2, 3], "val": [4, 5]}'

 Split into more than two splits:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht \
        --operation.type split \
        --operation.splits '{"train": 0.6, "val": 0.2, "test": 0.2}'

 Merge multiple datasets:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht_merged \
        --operation.type merge \
        --operation.repo_ids "['lerobot/pusht_train', 'lerobot/pusht_val']"

 Remove camera feature:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht \
        --operation.type remove_feature \
        --operation.feature_names "['observation.images.top']"

 Modify tasks - set a single task for all episodes (WARNING: modifies in-place):
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht \
        --operation.type modify_tasks \
        --operation.new_task "Pick up the cube and place it"

 Modify tasks - set different tasks for specific episodes (WARNING: modifies in-place):
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht \
        --operation.type modify_tasks \
        --operation.episode_tasks '{"0": "Task A", "1": "Task B", "2": "Task A"}'

 Modify tasks - set default task with overrides for specific episodes (WARNING: modifies in-place):
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht \
        --operation.type modify_tasks \
        --operation.new_task "Default task" \
        --operation.episode_tasks '{"5": "Special task for episode 5"}'

 Convert image dataset to video format and save locally:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht_image \
        --operation.type convert_image_to_video \
        --operation.output_dir /path/to/output/pusht_video

 Convert image dataset to video format and save with new repo_id:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht_image \
        --new_repo_id lerobot/pusht_video \
        --operation.type convert_image_to_video

 Convert image dataset to video format and push to hub:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --repo_id lerobot/pusht_image \
        --new_repo_id lerobot/pusht_video \
        --operation.type convert_image_to_video \
        --push_to_hub true

+Show dataset information:
+    lerobot-edit-dataset \
+        --repo_id lerobot/pusht_image \
+        --operation.type info \
+        --operation.show_features true
+
+Show dataset information without feature details:
+    lerobot-edit-dataset \
+        --repo_id lerobot/pusht_image \
+        --operation.type info \
+        --operation.show_features false
+
 Using JSON config file:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
        --config_path path/to/edit_config.json
 """

+import abc
 import logging
 import shutil
+import sys
 from dataclasses import dataclass
 from pathlib import Path

+import draccus
+
 from lerobot.configs import parser
 from lerobot.datasets.dataset_tools import (
    convert_image_to_video_dataset,
@@ -129,39 +145,46 @@ from lerobot.utils.utils import init_logging


@dataclass
-class DeleteEpisodesConfig:
-    type: str = "delete_episodes"
+class OperationConfig(draccus.ChoiceRegistry, abc.ABC):
+    @property
+    def type(self) -> str:
+        return self.get_choice_name(self.__class__)
+
+
+@OperationConfig.register_subclass("delete_episodes")
+@dataclass
+class DeleteEpisodesConfig(OperationConfig):
    episode_indices: list[int] | None = None


+@OperationConfig.register_subclass("split")
@dataclass
-class SplitConfig:
-    type: str = "split"
+class SplitConfig(OperationConfig):
    splits: dict[str, float | list[int]] | None = None


+@OperationConfig.register_subclass("merge")
@dataclass
-class MergeConfig:
-    type: str = "merge"
+class MergeConfig(OperationConfig):
    repo_ids: list[str] | None = None


+@OperationConfig.register_subclass("remove_feature")
@dataclass
-class RemoveFeatureConfig:
-    type: str = "remove_feature"
+class RemoveFeatureConfig(OperationConfig):
    feature_names: list[str] | None = None


+@OperationConfig.register_subclass("modify_tasks")
@dataclass
-class ModifyTasksConfig:
-    type: str = "modify_tasks"
+class ModifyTasksConfig(OperationConfig):
    new_task: str | None = None
    episode_tasks: dict[str, str] | None = None


+@OperationConfig.register_subclass("convert_image_to_video")
@dataclass
-class ConvertImageToVideoConfig:
-    type: str = "convert_image_to_video"
+class ConvertImageToVideoConfig(OperationConfig):
    output_dir: str | None = None
    vcodec: str = "libsvtav1"
    pix_fmt: str = "yuv420p"
@@ -174,17 +197,17 @@ class ConvertImageToVideoConfig:
    max_frames_per_batch: int | None = None


+@OperationConfig.register_subclass("info")
+@dataclass
+class InfoConfig(OperationConfig):
+    type: str = "info"
+    show_features: bool = False
+
+
@dataclass
 class EditDatasetConfig:
    repo_id: str
-    operation: (
-        DeleteEpisodesConfig
-        | SplitConfig
-        | MergeConfig
-        | RemoveFeatureConfig
-        | ModifyTasksConfig
-        | ConvertImageToVideoConfig
-    )
+    operation: OperationConfig
    root: str | None = None
    new_repo_id: str | None = None
    push_to_hub: bool = False
@@ -433,6 +456,49 @@ def handle_convert_image_to_video(cfg: EditDatasetConfig) -> None:
        logging.info("Dataset saved locally (not pushed to hub)")


+def _get_dataset_size(repo_path):
+    import os
+
+    total = 0
+    with os.scandir(repo_path) as it:
+        for entry in it:
+            if entry.is_file():
+                total += entry.stat().st_size
+            elif entry.is_dir():
+                total += _get_dataset_size(entry.path)
+    return total
+
+
+def handle_info(cfg: EditDatasetConfig):
+    if not isinstance(cfg.operation, InfoConfig):
+        raise ValueError("Operation config must be InfoConfig")
+
+    dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
+    sys.stdout.write(f"======Info {dataset.meta.repo_id}\n")
+    sys.stdout.write(f"Repository ID: {dataset.meta.repo_id} \n")
+    sys.stdout.write(f"Total episode: {dataset.meta.total_episodes} \n")
+    sys.stdout.write(f"Total task: {dataset.meta.total_tasks} \n")
+    sys.stdout.write(f"Total frame(Actual Count): {dataset.meta.total_frames}({len(dataset)}) \n")
+    sys.stdout.write(
+        f"Average frame per episode: {dataset.meta.total_frames / dataset.meta.total_episodes:.1f}\n"
+    )
+    sys.stdout.write(
+        f"Average episode time(sec): {(dataset.meta.total_frames / dataset.meta.total_episodes) / dataset.meta.fps:.1f}\n"
+    )
+    sys.stdout.write(f"FPS: {dataset.meta.fps}\n")
+
+    total_file_size = _get_dataset_size(dataset.root)
+    sys.stdout.write(f"Size: {total_file_size / (1024 * 1024):.1f} MB\n")
+    if cfg.operation.show_features:
+        import json
+
+        feature_dump_str = json.dumps(
+            dataset.meta.features, ensure_ascii=False, indent=4, sort_keys=True, separators=(",", ": ")
+        )
+        sys.stdout.write("Features:\n")
+        sys.stdout.write(f"{feature_dump_str}\n")
+
+
@parser.wrap()
 def edit_dataset(cfg: EditDatasetConfig) -> None:
    operation_type = cfg.operation.type
@@ -449,11 +515,11 @@ def edit_dataset(cfg: EditDatasetConfig) -> None:
        handle_modify_tasks(cfg)
    elif operation_type == "convert_image_to_video":
        handle_convert_image_to_video(cfg)
+    elif operation_type == "info":
+        handle_info(cfg)
    else:
-        raise ValueError(
-            f"Unknown operation type: {operation_type}\n"
-            f"Available operations: delete_episodes, split, merge, remove_feature, modify_tasks, convert_image_to_video"
-        )
+        available = ", ".join(OperationConfig.get_known_choices())
+        raise ValueError(f"Unknown operation: {operation_type}\nAvailable operations: {available}")


 def main() -> None:
--- a/src/lerobot/scripts/lerobot_record.py
+++ b/src/lerobot/scripts/lerobot_record.py
@@ -398,7 +398,14 @@ def record_loop(
            )

        dt_s = time.perf_counter() - start_loop_t
-        precise_sleep(max(1 / fps - dt_s, 0.0))
+
+        sleep_time_s: float = 1 / fps - dt_s
+        if sleep_time_s < 0:
+            logging.warning(
+                f"Record loop is running slower ({1 / dt_s:.1f} Hz) than the target FPS ({fps} Hz). Dataset frames might be dropped and robot control might be unstable. Common causes are: 1) Camera FPS not keeping up 2) Policy inference taking too long 3) CPU starvation"
+            )
+
+        precise_sleep(max(sleep_time_s, 0.0))

        timestamp = time.perf_counter() - start_episode_t

--- a/src/lerobot/scripts/lerobot_replay.py
+++ b/src/lerobot/scripts/lerobot_replay.py
@@ -22,7 +22,7 @@ lerobot-replay \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=black \
-    --dataset.repo_id=aliberts/record-test \
+    --dataset.repo_id=<USER>/record-test \
    --dataset.episode=0
 ```

--- a/src/lerobot/scripts/lerobot_setup_can.py
+++ b/src/lerobot/scripts/lerobot_setup_can.py
@@ -45,7 +45,7 @@ from dataclasses import dataclass, field

 import draccus

-from lerobot.utils.import_utils import is_package_available
+from lerobot.utils.import_utils import _can_available

 MOTOR_NAMES = {
    0x01: "joint_1",
@@ -336,7 +336,7 @@ def run_speed(cfg: CANSetupConfig):

@draccus.wrap()
 def setup_can(cfg: CANSetupConfig):
-    if not is_package_available("can"):
+    if not _can_available:
        print("Error: python-can not installed. Install with: pip install python-can")
        sys.exit(1)

--- a/src/lerobot/teleoperators/bi_openarm_leader/bi_openarm_leader.py
+++ b/src/lerobot/teleoperators/bi_openarm_leader/bi_openarm_leader.py
@@ -19,6 +19,7 @@ from functools import cached_property

 from lerobot.processor import RobotAction
 from lerobot.teleoperators.openarm_leader import OpenArmLeaderConfig
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected

 from ..openarm_leader import OpenArmLeader
 from ..teleoperator import Teleoperator
@@ -88,6 +89,7 @@ class BiOpenArmLeader(Teleoperator):
    def is_connected(self) -> bool:
        return self.left_arm.is_connected and self.right_arm.is_connected

+    @check_if_already_connected
    def connect(self, calibrate: bool = True) -> None:
        self.left_arm.connect(calibrate)
        self.right_arm.connect(calibrate)
@@ -109,6 +111,7 @@ class BiOpenArmLeader(Teleoperator):
            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
        )

+    @check_if_not_connected
    def get_action(self) -> RobotAction:
        action_dict = {}

@@ -126,6 +129,7 @@ class BiOpenArmLeader(Teleoperator):
        # TODO: Implement force feedback
        raise NotImplementedError

+    @check_if_not_connected
    def disconnect(self) -> None:
        self.left_arm.disconnect()
        self.right_arm.disconnect()
--- a/src/lerobot/teleoperators/bi_so_leader/bi_so_leader.py
+++ b/src/lerobot/teleoperators/bi_so_leader/bi_so_leader.py
@@ -18,7 +18,7 @@ import logging
 from functools import cached_property

 from lerobot.teleoperators.so_leader import SOLeaderTeleopConfig
-from lerobot.utils.decorators import check_if_not_connected
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected

 from ..so_leader import SOLeader
 from ..teleoperator import Teleoperator
@@ -72,6 +72,7 @@ class BiSOLeader(Teleoperator):
    def is_connected(self) -> bool:
        return self.left_arm.is_connected and self.right_arm.is_connected

+    @check_if_already_connected
    def connect(self, calibrate: bool = True) -> None:
        self.left_arm.connect(calibrate)
        self.right_arm.connect(calibrate)
@@ -110,6 +111,7 @@ class BiSOLeader(Teleoperator):
        # TODO: Implement force feedback
        raise NotImplementedError

+    @check_if_not_connected
    def disconnect(self) -> None:
        self.left_arm.disconnect()
        self.right_arm.disconnect()
--- a/src/lerobot/teleoperators/openarm_leader/openarm_leader.py
+++ b/src/lerobot/teleoperators/openarm_leader/openarm_leader.py
@@ -21,7 +21,7 @@ from typing import Any
 from lerobot.motors import Motor, MotorCalibration, MotorNormMode
 from lerobot.motors.damiao import DamiaoMotorsBus
 from lerobot.processor import RobotAction
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected

 from ..teleoperator import Teleoperator
 from .config_openarm_leader import OpenArmLeaderConfig
@@ -84,6 +84,7 @@ class OpenArmLeader(Teleoperator):
        """Check if teleoperator is connected."""
        return self.bus.is_connected

+    @check_if_already_connected
    def connect(self, calibrate: bool = True) -> None:
        """
        Connect to the teleoperator.
@@ -91,8 +92,6 @@ class OpenArmLeader(Teleoperator):
        For manual control, we disable torque after connecting so the
        arm can be moved by hand.
        """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} already connected")

        # Connect to CAN bus
        logger.info(f"Connecting arm on {self.config.port}...")
@@ -183,6 +182,7 @@ class OpenArmLeader(Teleoperator):
            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
        )

+    @check_if_not_connected
    def get_action(self) -> RobotAction:
        """
        Get current action from the leader arm.
@@ -193,8 +193,6 @@ class OpenArmLeader(Teleoperator):
        Reads all motor states (pos/vel/torque) in one CAN refresh cycle.
        """
        start = time.perf_counter()
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        action_dict: dict[str, Any] = {}

@@ -214,10 +212,9 @@ class OpenArmLeader(Teleoperator):
    def send_feedback(self, feedback: dict[str, float]) -> None:
        raise NotImplementedError("Feedback is not yet implemented for OpenArm leader.")

+    @check_if_not_connected
    def disconnect(self) -> None:
        """Disconnect from teleoperator."""
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        # Disconnect CAN bus
        # For manual control, ensure torque is disabled before disconnecting
--- a/src/lerobot/teleoperators/so_leader/config_so_leader.py
+++ b/src/lerobot/teleoperators/so_leader/config_so_leader.py
@@ -28,7 +28,7 @@ class SOLeaderConfig:
    port: str

    # Whether to use degrees for angles
-    use_degrees: bool = False
+    use_degrees: bool = True


@TeleoperatorConfig.register_subclass("so101_leader")
--- a/src/lerobot/utils/robot_utils.py
+++ b/src/lerobot/utils/robot_utils.py
@@ -16,14 +16,14 @@ import platform
 import time


-def precise_sleep(seconds: float, spin_threshold: float = 0.010, sleep_margin: float = 0.003):
+def precise_sleep(seconds: float, spin_threshold: float = 0.010, sleep_margin: float = 0.005):
    """
    Wait for `seconds` with better precision than time.sleep alone at the expense of more CPU usage.

    Parameters:
      - seconds: duration to wait
      - spin_threshold: if remaining <= spin_threshold -> spin; otherwise sleep (seconds). Default 10ms
-      - sleep_margin: when sleeping leave this much time before deadline to avoid oversleep. Default 3ms
+      - sleep_margin: when sleeping leave this much time before deadline to avoid oversleep. Default 5ms

    Note:
        The default parameters are chosen to prioritize timing accuracy over CPU usage for the common 30 FPS use case.
--- a/src/lerobot/utils/train_utils.py
+++ b/src/lerobot/utils/train_utils.py
@@ -95,6 +95,7 @@ def save_checkpoint(
        optimizer (Optimizer | None, optional): The optimizer to save the state from. Defaults to None.
        scheduler (LRScheduler | None, optional): The scheduler to save the state from. Defaults to None.
        preprocessor: The preprocessor/pipeline to save. Defaults to None.
+        postprocessor: The postprocessor/pipeline to save. Defaults to None.
    """
    pretrained_dir = checkpoint_dir / PRETRAINED_MODEL_DIR
    policy.save_pretrained(pretrained_dir)
--- a/tests/datasets/test_image_transforms.py
+++ b/tests/datasets/test_image_transforms.py
@@ -390,6 +390,30 @@ def test_sharpness_jitter_invalid_range_max_smaller():
        SharpnessJitter((2.0, 0.1))


+def test_make_transform_from_config_with_v2_resize(img_tensor_factory):
+    img_tensor = img_tensor_factory()
+    tf_cfg = ImageTransformConfig(type="Resize", kwargs={"size": (32, 32)})
+    tf = make_transform_from_config(tf_cfg)
+    assert isinstance(tf, v2.Resize)
+    output = tf(img_tensor)
+    assert output.shape[-2:] == (32, 32)
+
+
+def test_make_transform_from_config_with_v2_identity(img_tensor_factory):
+    img_tensor = img_tensor_factory()
+    tf_cfg = ImageTransformConfig(type="Identity", kwargs={})
+    tf = make_transform_from_config(tf_cfg)
+    assert isinstance(tf, v2.Identity)
+    output = tf(img_tensor)
+    assert output.shape == img_tensor.shape
+
+
+def test_make_transform_from_config_invalid_type():
+    tf_cfg = ImageTransformConfig(type="NotARealTransform", kwargs={})
+    with pytest.raises(ValueError, match="not valid"):
+        make_transform_from_config(tf_cfg)
+
+
 def test_save_all_transforms(img_tensor_factory, tmp_path):
    img_tensor = img_tensor_factory()
    tf_cfg = ImageTransformsConfig(enable=True)
--- a/tests/optim/test_schedulers.py
+++ b/tests/optim/test_schedulers.py
@@ -11,6 +11,8 @@
 # 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 torch
+from packaging.version import Version
 from torch.optim.lr_scheduler import LambdaLR

 from lerobot.optim.schedulers import (
@@ -38,6 +40,10 @@ def test_diffuser_scheduler(optimizer):
        "last_epoch": 1,
        "lr_lambdas": [None],
    }
+
+    if Version(torch.__version__) >= Version("2.8"):
+        expected_state_dict["_is_initial"] = False
+
    assert scheduler.state_dict() == expected_state_dict


@@ -56,6 +62,10 @@ def test_vqbet_scheduler(optimizer):
        "last_epoch": 1,
        "lr_lambdas": [None],
    }
+
+    if Version(torch.__version__) >= Version("2.8"):
+        expected_state_dict["_is_initial"] = False
+
    assert scheduler.state_dict() == expected_state_dict


@@ -76,6 +86,10 @@ def test_cosine_decay_with_warmup_scheduler(optimizer):
        "last_epoch": 1,
        "lr_lambdas": [None],
    }
+
+    if Version(torch.__version__) >= Version("2.8"):
+        expected_state_dict["_is_initial"] = False
+
    assert scheduler.state_dict() == expected_state_dict


--- a/tests/policies/test_sac_policy.py
+++ b/tests/policies/test_sac_policy.py
@@ -14,8 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-import math
-
 import pytest
 import torch
 from torch import Tensor, nn
@@ -23,6 +21,7 @@ from torch import Tensor, nn
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.policies.sac.configuration_sac import SACConfig
 from lerobot.policies.sac.modeling_sac import MLP, SACPolicy
+from lerobot.rl.algorithms.sac import SACAlgorithm, SACAlgorithmConfig
 from lerobot.utils.constants import ACTION, OBS_IMAGE, OBS_STATE
 from lerobot.utils.random_utils import seeded_context, set_seed

@@ -138,41 +137,6 @@ def create_observation_batch_with_visual_input(batch_size: int = 8, state_dim: i
    }


-def make_optimizers(policy: SACPolicy, has_discrete_action: bool = False) -> dict[str, torch.optim.Optimizer]:
-    """Create optimizers for the SAC policy."""
-    optimizer_actor = torch.optim.Adam(
-        # Handle the case of shared encoder where the encoder weights are not optimized with the actor gradient
-        params=[
-            p
-            for n, p in policy.actor.named_parameters()
-            if not policy.config.shared_encoder or not n.startswith("encoder")
-        ],
-        lr=policy.config.actor_lr,
-    )
-    optimizer_critic = torch.optim.Adam(
-        params=policy.critic_ensemble.parameters(),
-        lr=policy.config.critic_lr,
-    )
-    optimizer_temperature = torch.optim.Adam(
-        params=[policy.log_alpha],
-        lr=policy.config.critic_lr,
-    )
-
-    optimizers = {
-        "actor": optimizer_actor,
-        "critic": optimizer_critic,
-        "temperature": optimizer_temperature,
-    }
-
-    if has_discrete_action:
-        optimizers["discrete_critic"] = torch.optim.Adam(
-            params=policy.discrete_critic.parameters(),
-            lr=policy.config.critic_lr,
-        )
-
-    return optimizers
-
-
 def create_default_config(
    state_dim: int, continuous_action_dim: int, has_discrete_action: bool = False
 ) -> SACConfig:
@@ -212,7 +176,6 @@ def create_config_with_visual_input(
        "std": torch.randn(3, 1, 1),
    }

-    # Let make tests a little bit faster
    config.state_encoder_hidden_dim = 32
    config.latent_dim = 32

@@ -220,75 +183,112 @@ def create_config_with_visual_input(
    return config


-@pytest.mark.parametrize("batch_size,state_dim,action_dim", [(2, 6, 6), (1, 10, 10)])
-def test_sac_policy_with_default_config(batch_size: int, state_dim: int, action_dim: int):
-    batch = create_default_train_batch(batch_size=batch_size, action_dim=action_dim, state_dim=state_dim)
-    config = create_default_config(state_dim=state_dim, continuous_action_dim=action_dim)
-
+def _make_algorithm(config: SACConfig) -> tuple[SACAlgorithm, SACPolicy]:
+    """Helper to create policy + algorithm pair for tests that need critics."""
    policy = SACPolicy(config=config)
    policy.train()
+    algo_config = SACAlgorithmConfig.from_policy_config(config)
+    algorithm = SACAlgorithm(policy=policy, config=algo_config)
+    algorithm.make_optimizers()
+    return algorithm, policy

-    optimizers = make_optimizers(policy)
-
-    cirtic_loss = policy.forward(batch, model="critic")["loss_critic"]
-    assert cirtic_loss.item() is not None
-    assert cirtic_loss.shape == ()
-    cirtic_loss.backward()
-    optimizers["critic"].step()
-
-    actor_loss = policy.forward(batch, model="actor")["loss_actor"]
-    assert actor_loss.item() is not None
-    assert actor_loss.shape == ()
-
-    actor_loss.backward()
-    optimizers["actor"].step()
-
-    temperature_loss = policy.forward(batch, model="temperature")["loss_temperature"]
-    assert temperature_loss.item() is not None
-    assert temperature_loss.shape == ()
-
-    temperature_loss.backward()
-    optimizers["temperature"].step()

+@pytest.mark.parametrize("batch_size,state_dim,action_dim", [(2, 6, 6), (1, 10, 10)])
+def test_sac_policy_select_action(batch_size: int, state_dim: int, action_dim: int):
+    config = create_default_config(state_dim=state_dim, continuous_action_dim=action_dim)
+    policy = SACPolicy(config=config)
    policy.eval()
+
    with torch.no_grad():
        observation_batch = create_observation_batch(batch_size=batch_size, state_dim=state_dim)
        selected_action = policy.select_action(observation_batch)
-        assert selected_action.shape == (batch_size, action_dim)
+        # squeeze(0) removes batch dim when batch_size==1
+        assert selected_action.shape[-1] == action_dim
+
+
+def test_sac_policy_select_action_with_discrete():
+    """select_action should return continuous + discrete actions."""
+    config = create_default_config(state_dim=10, continuous_action_dim=6)
+    config.num_discrete_actions = 3
+    policy = SACPolicy(config=config)
+    policy.eval()
+
+    with torch.no_grad():
+        observation_batch = create_observation_batch(batch_size=1, state_dim=10)
+        # Squeeze to unbatched (single observation)
+        observation_batch = {k: v.squeeze(0) for k, v in observation_batch.items()}
+        selected_action = policy.select_action(observation_batch)
+        assert selected_action.shape[-1] == 7  # 6 continuous + 1 discrete


@pytest.mark.parametrize("batch_size,state_dim,action_dim", [(2, 6, 6), (1, 10, 10)])
-def test_sac_policy_with_visual_input(batch_size: int, state_dim: int, action_dim: int):
-    config = create_config_with_visual_input(state_dim=state_dim, continuous_action_dim=action_dim)
+def test_sac_policy_forward(batch_size: int, state_dim: int, action_dim: int):
+    config = create_default_config(state_dim=state_dim, continuous_action_dim=action_dim)
    policy = SACPolicy(config=config)
+    policy.eval()
+
+    batch = create_default_train_batch(batch_size=batch_size, action_dim=action_dim, state_dim=state_dim)
+    with torch.no_grad():
+        output = policy.forward(batch)
+        assert "action" in output
+        assert "log_prob" in output
+        assert "action_mean" in output
+        assert output["action"].shape == (batch_size, action_dim)
+
+
+@pytest.mark.parametrize("batch_size,state_dim,action_dim", [(2, 6, 6), (1, 10, 10)])
+def test_sac_training_through_algorithm(batch_size: int, state_dim: int, action_dim: int):
+    config = create_default_config(state_dim=state_dim, continuous_action_dim=action_dim)
+    algorithm, policy = _make_algorithm(config)
+
+    batch = create_default_train_batch(batch_size=batch_size, action_dim=action_dim, state_dim=state_dim)
+    forward_batch = algorithm._prepare_forward_batch(batch)
+
+    critic_loss = algorithm._compute_loss_critic(forward_batch)
+    assert critic_loss.item() is not None
+    assert critic_loss.shape == ()
+    algorithm.optimizers["critic"].zero_grad()
+    critic_loss.backward()
+    algorithm.optimizers["critic"].step()
+
+    actor_loss = algorithm._compute_loss_actor(forward_batch)
+    assert actor_loss.item() is not None
+    assert actor_loss.shape == ()
+    algorithm.optimizers["actor"].zero_grad()
+    actor_loss.backward()
+    algorithm.optimizers["actor"].step()
+
+    temp_loss = algorithm._compute_loss_temperature(forward_batch)
+    assert temp_loss.item() is not None
+    assert temp_loss.shape == ()
+    algorithm.optimizers["temperature"].zero_grad()
+    temp_loss.backward()
+    algorithm.optimizers["temperature"].step()
+
+
+@pytest.mark.parametrize("batch_size,state_dim,action_dim", [(2, 6, 6), (1, 10, 10)])
+def test_sac_training_with_visual_input(batch_size: int, state_dim: int, action_dim: int):
+    config = create_config_with_visual_input(state_dim=state_dim, continuous_action_dim=action_dim)
+    algorithm, policy = _make_algorithm(config)

    batch = create_train_batch_with_visual_input(
        batch_size=batch_size, state_dim=state_dim, action_dim=action_dim
    )
+    forward_batch = algorithm._prepare_forward_batch(batch)

-    policy.train()
+    critic_loss = algorithm._compute_loss_critic(forward_batch)
+    assert critic_loss.item() is not None
+    assert critic_loss.shape == ()
+    algorithm.optimizers["critic"].zero_grad()
+    critic_loss.backward()
+    algorithm.optimizers["critic"].step()

-    optimizers = make_optimizers(policy)
-
-    cirtic_loss = policy.forward(batch, model="critic")["loss_critic"]
-    assert cirtic_loss.item() is not None
-    assert cirtic_loss.shape == ()
-    cirtic_loss.backward()
-    optimizers["critic"].step()
-
-    actor_loss = policy.forward(batch, model="actor")["loss_actor"]
+    actor_loss = algorithm._compute_loss_actor(forward_batch)
    assert actor_loss.item() is not None
    assert actor_loss.shape == ()
-
+    algorithm.optimizers["actor"].zero_grad()
    actor_loss.backward()
-    optimizers["actor"].step()
-
-    temperature_loss = policy.forward(batch, model="temperature")["loss_temperature"]
-    assert temperature_loss.item() is not None
-    assert temperature_loss.shape == ()
-
-    temperature_loss.backward()
-    optimizers["temperature"].step()
+    algorithm.optimizers["actor"].step()

    policy.eval()
    with torch.no_grad():
@@ -296,207 +296,181 @@ def test_sac_policy_with_visual_input(batch_size: int, state_dim: int, action_di
            batch_size=batch_size, state_dim=state_dim
        )
        selected_action = policy.select_action(observation_batch)
-        assert selected_action.shape == (batch_size, action_dim)
+        assert selected_action.shape[-1] == action_dim


-# Let's check best candidates for pretrained encoders
@pytest.mark.parametrize(
    "batch_size,state_dim,action_dim,vision_encoder_name",
    [(1, 6, 6, "helper2424/resnet10"), (1, 6, 6, "facebook/convnext-base-224")],
 )
@pytest.mark.skipif(not TRANSFORMERS_AVAILABLE, reason="Transformers are not installed")
-def test_sac_policy_with_pretrained_encoder(
+def test_sac_training_with_pretrained_encoder(
    batch_size: int, state_dim: int, action_dim: int, vision_encoder_name: str
 ):
    config = create_config_with_visual_input(state_dim=state_dim, continuous_action_dim=action_dim)
    config.vision_encoder_name = vision_encoder_name
-    policy = SACPolicy(config=config)
-    policy.train()
+    algorithm, policy = _make_algorithm(config)

    batch = create_train_batch_with_visual_input(
        batch_size=batch_size, state_dim=state_dim, action_dim=action_dim
    )
+    forward_batch = algorithm._prepare_forward_batch(batch)

-    optimizers = make_optimizers(policy)
+    critic_loss = algorithm._compute_loss_critic(forward_batch)
+    assert critic_loss.item() is not None
+    assert critic_loss.shape == ()
+    algorithm.optimizers["critic"].zero_grad()
+    critic_loss.backward()
+    algorithm.optimizers["critic"].step()

-    cirtic_loss = policy.forward(batch, model="critic")["loss_critic"]
-    assert cirtic_loss.item() is not None
-    assert cirtic_loss.shape == ()
-    cirtic_loss.backward()
-    optimizers["critic"].step()
-
-    actor_loss = policy.forward(batch, model="actor")["loss_actor"]
+    actor_loss = algorithm._compute_loss_actor(forward_batch)
    assert actor_loss.item() is not None
    assert actor_loss.shape == ()


-def test_sac_policy_with_shared_encoder():
+def test_sac_training_with_shared_encoder():
    batch_size = 2
    action_dim = 10
    state_dim = 10
    config = create_config_with_visual_input(state_dim=state_dim, continuous_action_dim=action_dim)
    config.shared_encoder = True

-    policy = SACPolicy(config=config)
-    policy.train()
+    algorithm, policy = _make_algorithm(config)

    batch = create_train_batch_with_visual_input(
        batch_size=batch_size, state_dim=state_dim, action_dim=action_dim
    )
+    forward_batch = algorithm._prepare_forward_batch(batch)

-    policy.train()
+    critic_loss = algorithm._compute_loss_critic(forward_batch)
+    assert critic_loss.shape == ()
+    algorithm.optimizers["critic"].zero_grad()
+    critic_loss.backward()
+    algorithm.optimizers["critic"].step()

-    optimizers = make_optimizers(policy)
-
-    cirtic_loss = policy.forward(batch, model="critic")["loss_critic"]
-    assert cirtic_loss.item() is not None
-    assert cirtic_loss.shape == ()
-    cirtic_loss.backward()
-    optimizers["critic"].step()
-
-    actor_loss = policy.forward(batch, model="actor")["loss_actor"]
-    assert actor_loss.item() is not None
+    actor_loss = algorithm._compute_loss_actor(forward_batch)
    assert actor_loss.shape == ()
-
+    algorithm.optimizers["actor"].zero_grad()
    actor_loss.backward()
-    optimizers["actor"].step()
+    algorithm.optimizers["actor"].step()


-def test_sac_policy_with_discrete_critic():
+def test_sac_training_with_discrete_critic():
    batch_size = 2
    continuous_action_dim = 9
-    full_action_dim = continuous_action_dim + 1  # the last action is discrete
+    full_action_dim = continuous_action_dim + 1
    state_dim = 10
    config = create_config_with_visual_input(
        state_dim=state_dim, continuous_action_dim=continuous_action_dim, has_discrete_action=True
    )
+    config.num_discrete_actions = 5

-    num_discrete_actions = 5
-    config.num_discrete_actions = num_discrete_actions
-
-    policy = SACPolicy(config=config)
-    policy.train()
+    algorithm, policy = _make_algorithm(config)

    batch = create_train_batch_with_visual_input(
        batch_size=batch_size, state_dim=state_dim, action_dim=full_action_dim
    )
+    forward_batch = algorithm._prepare_forward_batch(batch)

-    policy.train()
+    critic_loss = algorithm._compute_loss_critic(forward_batch)
+    assert critic_loss.shape == ()
+    algorithm.optimizers["critic"].zero_grad()
+    critic_loss.backward()
+    algorithm.optimizers["critic"].step()

-    optimizers = make_optimizers(policy, has_discrete_action=True)
-
-    cirtic_loss = policy.forward(batch, model="critic")["loss_critic"]
-    assert cirtic_loss.item() is not None
-    assert cirtic_loss.shape == ()
-    cirtic_loss.backward()
-    optimizers["critic"].step()
-
-    discrete_critic_loss = policy.forward(batch, model="discrete_critic")["loss_discrete_critic"]
-    assert discrete_critic_loss.item() is not None
+    discrete_critic_loss = algorithm._compute_loss_discrete_critic(forward_batch)
    assert discrete_critic_loss.shape == ()
+    algorithm.optimizers["discrete_critic"].zero_grad()
    discrete_critic_loss.backward()
-    optimizers["discrete_critic"].step()
+    algorithm.optimizers["discrete_critic"].step()

-    actor_loss = policy.forward(batch, model="actor")["loss_actor"]
-    assert actor_loss.item() is not None
+    actor_loss = algorithm._compute_loss_actor(forward_batch)
    assert actor_loss.shape == ()
-
+    algorithm.optimizers["actor"].zero_grad()
    actor_loss.backward()
-    optimizers["actor"].step()
+    algorithm.optimizers["actor"].step()

    policy.eval()
    with torch.no_grad():
        observation_batch = create_observation_batch_with_visual_input(
            batch_size=batch_size, state_dim=state_dim
        )
-        selected_action = policy.select_action(observation_batch)
-        assert selected_action.shape == (batch_size, full_action_dim)
-
-        discrete_actions = selected_action[:, -1].long()
-        discrete_action_values = set(discrete_actions.tolist())
-
-        assert all(action in range(num_discrete_actions) for action in discrete_action_values), (
-            f"Discrete action {discrete_action_values} is not in range({num_discrete_actions})"
-        )
+        # Policy.select_action now handles both continuous + discrete
+        selected_action = policy.select_action({k: v.squeeze(0) for k, v in observation_batch.items()})
+        assert selected_action.shape[-1] == continuous_action_dim + 1


-def test_sac_policy_with_default_entropy():
+def test_sac_algorithm_target_entropy():
    config = create_default_config(continuous_action_dim=10, state_dim=10)
-    policy = SACPolicy(config=config)
-    assert policy.target_entropy == -5.0
+    _, policy = _make_algorithm(config)
+    algo_config = SACAlgorithmConfig.from_policy_config(config)
+    algorithm = SACAlgorithm(policy=policy, config=algo_config)
+    assert algorithm.target_entropy == -5.0


-def test_sac_policy_default_target_entropy_with_discrete_action():
+def test_sac_algorithm_target_entropy_with_discrete_action():
    config = create_config_with_visual_input(state_dim=10, continuous_action_dim=6, has_discrete_action=True)
+    config.num_discrete_actions = 5
+    algo_config = SACAlgorithmConfig.from_policy_config(config)
    policy = SACPolicy(config=config)
-    assert policy.target_entropy == -3.0
+    algorithm = SACAlgorithm(policy=policy, config=algo_config)
+    assert algorithm.target_entropy == -3.5


-def test_sac_policy_with_predefined_entropy():
-    config = create_default_config(state_dim=10, continuous_action_dim=6)
-    config.target_entropy = -3.5
+def test_sac_algorithm_temperature():
+    import math

-    policy = SACPolicy(config=config)
-    assert policy.target_entropy == pytest.approx(-3.5)
-
-
-def test_sac_policy_update_temperature():
-    """Test that temperature property is always in sync with log_alpha."""
    config = create_default_config(continuous_action_dim=10, state_dim=10)
+    algo_config = SACAlgorithmConfig.from_policy_config(config)
    policy = SACPolicy(config=config)
+    algorithm = SACAlgorithm(policy=policy, config=algo_config)

-    assert policy.temperature == pytest.approx(1.0)
-    policy.log_alpha.data = torch.tensor([math.log(0.1)])
-    # Temperature property automatically reflects log_alpha changes
-    assert policy.temperature == pytest.approx(0.1)
+    assert algorithm.temperature == pytest.approx(1.0)
+    algorithm.log_alpha.data = torch.tensor([math.log(0.1)])
+    assert algorithm.temperature == pytest.approx(0.1)


-def test_sac_policy_update_target_network():
+def test_sac_algorithm_update_target_network():
    config = create_default_config(state_dim=10, continuous_action_dim=6)
    config.critic_target_update_weight = 1.0
-
+    algo_config = SACAlgorithmConfig.from_policy_config(config)
    policy = SACPolicy(config=config)
-    policy.train()
+    algorithm = SACAlgorithm(policy=policy, config=algo_config)

-    for p in policy.critic_ensemble.parameters():
+    for p in algorithm.critic_ensemble.parameters():
        p.data = torch.ones_like(p.data)

-    policy.update_target_networks()
-    for p in policy.critic_target.parameters():
-        assert torch.allclose(p.data, torch.ones_like(p.data)), (
-            f"Target network {p.data} is not equal to {torch.ones_like(p.data)}"
-        )
+    algorithm._update_target_networks()
+    for p in algorithm.critic_target.parameters():
+        assert torch.allclose(p.data, torch.ones_like(p.data))


@pytest.mark.parametrize("num_critics", [1, 3])
-def test_sac_policy_with_critics_number_of_heads(num_critics: int):
+def test_sac_algorithm_with_critics_number_of_heads(num_critics: int):
    batch_size = 2
    action_dim = 10
    state_dim = 10
    config = create_config_with_visual_input(state_dim=state_dim, continuous_action_dim=action_dim)
    config.num_critics = num_critics

-    policy = SACPolicy(config=config)
-    policy.train()
+    algorithm, policy = _make_algorithm(config)

-    assert len(policy.critic_ensemble.critics) == num_critics
+    assert len(algorithm.critic_ensemble.critics) == num_critics

    batch = create_train_batch_with_visual_input(
        batch_size=batch_size, state_dim=state_dim, action_dim=action_dim
    )
+    forward_batch = algorithm._prepare_forward_batch(batch)

-    policy.train()
-
-    optimizers = make_optimizers(policy)
-
-    cirtic_loss = policy.forward(batch, model="critic")["loss_critic"]
-    assert cirtic_loss.item() is not None
-    assert cirtic_loss.shape == ()
-    cirtic_loss.backward()
-    optimizers["critic"].step()
+    critic_loss = algorithm._compute_loss_critic(forward_batch)
+    assert critic_loss.shape == ()
+    algorithm.optimizers["critic"].zero_grad()
+    critic_loss.backward()
+    algorithm.optimizers["critic"].step()


 def test_sac_policy_save_and_load(tmp_path):
+    """Test that the policy can be saved and loaded from pretrained."""
    root = tmp_path / "test_sac_save_and_load"

    state_dim = 10
@@ -510,34 +484,41 @@ def test_sac_policy_save_and_load(tmp_path):
    loaded_policy = SACPolicy.from_pretrained(root, config=config)
    loaded_policy.eval()

-    batch = create_default_train_batch(batch_size=1, state_dim=10, action_dim=10)
+    assert policy.state_dict().keys() == loaded_policy.state_dict().keys()
+    for k in policy.state_dict():
+        assert torch.allclose(policy.state_dict()[k], loaded_policy.state_dict()[k], atol=1e-6)

    with torch.no_grad():
        with seeded_context(12):
-            # Collect policy values before saving
-            cirtic_loss = policy.forward(batch, model="critic")["loss_critic"]
-            actor_loss = policy.forward(batch, model="actor")["loss_actor"]
-            temperature_loss = policy.forward(batch, model="temperature")["loss_temperature"]
-
            observation_batch = create_observation_batch(batch_size=batch_size, state_dim=state_dim)
            actions = policy.select_action(observation_batch)

        with seeded_context(12):
-            # Collect policy values after loading
-            loaded_cirtic_loss = loaded_policy.forward(batch, model="critic")["loss_critic"]
-            loaded_actor_loss = loaded_policy.forward(batch, model="actor")["loss_actor"]
-            loaded_temperature_loss = loaded_policy.forward(batch, model="temperature")["loss_temperature"]
-
            loaded_observation_batch = create_observation_batch(batch_size=batch_size, state_dim=state_dim)
            loaded_actions = loaded_policy.select_action(loaded_observation_batch)

-        assert policy.state_dict().keys() == loaded_policy.state_dict().keys()
-        for k in policy.state_dict():
-            assert torch.allclose(policy.state_dict()[k], loaded_policy.state_dict()[k], atol=1e-6)
-
-        # Compare values before and after saving and loading
-        # They should be the same
-        assert torch.allclose(cirtic_loss, loaded_cirtic_loss)
-        assert torch.allclose(actor_loss, loaded_actor_loss)
-        assert torch.allclose(temperature_loss, loaded_temperature_loss)
        assert torch.allclose(actions, loaded_actions)
+
+
+def test_sac_policy_save_and_load_with_discrete_critic(tmp_path):
+    """Discrete critic should be saved/loaded as part of the policy."""
+    root = tmp_path / "test_sac_save_and_load_discrete"
+
+    state_dim = 10
+    action_dim = 6
+
+    config = create_default_config(state_dim=state_dim, continuous_action_dim=action_dim)
+    config.num_discrete_actions = 3
+    policy = SACPolicy(config=config)
+    policy.eval()
+    policy.save_pretrained(root)
+
+    loaded_policy = SACPolicy.from_pretrained(root, config=config)
+    loaded_policy.eval()
+
+    assert loaded_policy.discrete_critic is not None
+    dc_keys = [k for k in loaded_policy.state_dict() if k.startswith("discrete_critic.")]
+    assert len(dc_keys) > 0
+
+    for k in policy.state_dict():
+        assert torch.allclose(policy.state_dict()[k], loaded_policy.state_dict()[k], atol=1e-6)
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Khalil Meftah	519234a5d8	feat: add offline training in learner	2026-03-22 23:00:07 +01:00
Khalil Meftah	d9371b9a34	feat: add RLT algorithm	2026-03-22 22:59:35 +01:00
Khalil Meftah	17f47b9cbc	feat: add RLT policy RL-token encoder-decoder and actor	2026-03-22 22:57:43 +01:00
Khalil Meftah	05395c8b10	Add offline phase hooks to RLAlgorithm base	2026-03-22 22:52:56 +01:00
Khalil Meftah	f495054321	disable processor in actor for sac/hilserl	2026-03-19 13:42:46 +01:00
Khalil Meftah	2345c779ee	disable processor for sac/hilserl	2026-03-19 13:12:21 +01:00
Khalil Meftah	aaf8576411	chore: rename losses	2026-03-19 12:36:02 +01:00
Khalil Meftah	d3e6f14d4f	fix: move algorithm-owned modules to the policy device	2026-03-18 15:27:41 +01:00
Khalil Meftah	1f5487eea8	refactor: decouple policy from algorithm	2026-03-11 16:49:14 +01:00
Khalil Meftah	8d50be9faa	refactor: RL stack refactoring — RLAlgorithm, RLTrainer, DataMixer, and SAC restructuring - Add RLAlgorithm base class and RLAlgorithmConfig with draccus.ChoiceRegistry - Add RLTrainer for unified training orchestration with iterator pattern - Add DataMixer and OnlineOfflineMixer for online/offline data mixing - Restructure SAC algorithm with batch iterator and factory pattern - Add observation normalization pre/post processors - Add comprehensive tests for all new components	2026-03-03 16:50:00 +01:00
Khalil	2dd366436e	Fix gym-hil integration with the new LeRobot pipeline. (#2482 ) * Add GymHILAdapterProcessorStep for gym-hil environment integration * Fix action features in control loop for None teleop device with gym-hil * Finalize dataset before pushing to hub for visualization on the hub * Fix neutral action for gripper * fix pre-commit	2026-02-19 14:35:02 +01:00
Steven Palma	5f15232271	chore: remove usernames + use entrypoints in docs, comments & sample commands (#2988 )	2026-02-18 22:46:12 +01:00
Steven Palma	bc38261321	feat(robots): use read_latest() camera (#2987 ) * feat(robots): use read_latest() camera * fix(test): add read_latest reachy cam mock	2026-02-18 20:05:15 +01:00
Caroline Pascal	aaf3707058	fix(filtering): fixing episodes filtering in load_nested_dataset to always use .from_parquet() (#2982 )	2026-02-18 19:16:53 +01:00
Steven Palma	89bd58a9a2	chore(scripts): warn if we don't respect the target FPS (#2986 )	2026-02-18 18:22:35 +01:00
Steven Palma	b22e0315b0	fix(utils): more conservative sleep_margin default value in precise_sleep (#2985 )	2026-02-18 17:32:25 +01:00
HUANG TZU-CHUN	fcbf550952	fix(docs): update environment variable name to HF_LEROBOT_HOME in docstring (#2973 ) Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>	2026-02-18 11:27:40 +01:00
Sota Nakamura	af036ce57e	fix(scripts): serve grpc for a web viewer (#2881 ) * serve grpc for a web viewer * add help * remove ip detection * fix comment * pass grpc_port * fix(CLI): fixing CLI display-compressed-images argument 1/2 Co-authored-by: HUANG TZU-CHUN <tzu.chun.huang.tw@gmail.com> Signed-off-by: Caroline Pascal <caroline8.pascal@gmail.com> * fix(CLI): fixing CLI display-compressed-images argument 2/2 Co-authored-by: HUANG TZU-CHUN <tzu.chun.huang.tw@gmail.com> Signed-off-by: Caroline Pascal <caroline8.pascal@gmail.com> --------- Signed-off-by: Caroline Pascal <caroline8.pascal@gmail.com> Co-authored-by: Caroline Pascal <caroline8.pascal@gmail.com> Co-authored-by: HUANG TZU-CHUN <tzu.chun.huang.tw@gmail.com> Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>	2026-02-18 01:05:51 +01:00
Vladislav Sovrasov	1c388c0002	(Chore) Bump upper bound for torch version (#2897 ) * Bump upper torch version bound * Apply suggestion from @Copilot Signed-off-by: Vladislav Sovrasov <vladislav.sovrasov@intel.com> * Update ref state dicts for schedulers * Support older than 2.8 torch versions * Fix precommit --------- Signed-off-by: Vladislav Sovrasov <vladislav.sovrasov@intel.com>	2026-02-17 23:37:46 +01:00
masato-ka	51d3822d75	feat(datasets): Add info operation to lerobot-edit-dataset command (#2917 ) * Add New featrue to lerobot_edit_datset.py that show dataset information. * Fix to draccus error when happen give only --operation.type=info * Updating test and documents regarding lerobot-edit-dataset info function. * Updating documents regarding lerobot-edit-dataset extract function. option name in document is mistake. * feat(datasets): Update to align formatting with pre-commit.(#2917) Update to align formatting by pre-commit. --------- Co-authored-by: Caroline Pascal <caroline8.pascal@gmail.com>	2026-02-17 20:09:42 +01:00
Pepijn	6600b60e7f	always use degrees (#2968 )	2026-02-13 13:49:01 +01:00
Caroline Pascal	adebbcf090	fix(dataset tools draccus): fixing draccus parsing for dataset edit operation type specification (#2949 ) * fix(edit dataset operation): fixing dataset tools CLI operation type specification * test(edit dataset operation): adding tests for dataset tools operation type specification * chore(format): running pre-commit * chore(backward compatibility): adding a type property in OperationConfig for backward compatibility Signed-off-by: Caroline Pascal <caroline8.pascal@gmail.com>	2026-02-12 18:56:04 +01:00
taken-yjyoon	3615160d89	fix(typo): Fixing wrong argparse examples in the comments (using 'True' not 'true') (#1040 ) Co-authored-by: juni <>	2026-02-12 18:13:51 +01:00
Steven Palma	fc8a388a25	feat(cameras): make backend configurable to the CLI (#2945 ) * feat(cameras): make backend configurable to the CLI * chore(cameras): address feedback * feat(Enum error messages): adding better instanciation error messages for Enum classes * chore(Enum error messages): propagating Enum error messages to all camera classes * chore(comments): removing superfluous comments * chore(format): applying ruff checks --------- Co-authored-by: CarolinePascal <caroline8.pascal@gmail.com>	2026-02-11 13:57:25 +01:00
Steven Palma	3c84d271d5	fix(motors): use decorator to fix precommit (#2951 )	2026-02-10 18:40:50 +01:00
Steven Palma	1ba3975020	chore: use is_connected decorators (#2948 ) * chore: use is_connected decorators * chore(robots): add is_connected to bi setups too	2026-02-10 17:49:30 +01:00
Steven Palma	35363c5798	chore(linter): ensure motors module passes MyPy type checks (#2939 ) * fix: ensure motors module passes MyPy type checks This commit fixes 62 mypy type errors in the motors module by: - Updating Protocol classes (PortHandler, PacketHandler, GroupSyncRead, GroupSyncWrite) to use class-level attribute declarations instead of __init__ body declarations - Adding missing `broadcastPing` method to PacketHandler Protocol - Fixing return type annotations (e.g., `_get_motor_model` returns str, not int) - Fixing parameter types to use `Sequence` for covariant list parameters - Fixing `Mapping` for covariant dict value types in `_normalize` - Updating method signatures to be consistent across parent and child classes (disable_torque, enable_torque, _get_half_turn_homings) - Adding explicit `int()` casts for MotorCalibration arguments - Adding explicit `return None` for functions returning Optional types - Adding type annotations for variables like `data_list: dict[int, int]` - Using `# type: ignore[method-assign]` for intentional monkeypatch - Fixing variable references (using `self.groups` instead of `groups`) Fixes #1723 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * chore(style): pre-commit after main merge * chore(linter): solve comments * chore(linter): apply pre-commit fixes to damiao * chore(linter): more fixes to damiao --------- Co-authored-by: yurekami <yurekami@users.noreply.github.com> Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>	2026-02-10 17:35:39 +01:00
whats2000	778db19a17	[Bug Fix] fix(ci): prevent runner group error on fork pushes (#2911 ) * fix(ci): prevent runner group error on fork pushes Add repository check to unbound_deps_tests workflow to ensure aws-general-8-plus runner group is only used on main repository, preventing 'Required runner group not found' errors on forks. * fix(ci): use gating job to prevent runner allocation on forks The previous approach failed because GitHub evaluates runs-on before if conditions. Now using a check-repo job that runs on ubuntu-latest first, and all jobs with special runners depend on it and check its output before being scheduled. * fix(ci): add gating job to full_tests to prevent runner allocation on forks Apply the same gating pattern used in unbound_deps_tests to full_tests.yml to prevent GitHub from trying to allocate custom runners when workflows run on forks. The check-repo job runs first on ubuntu-latest and all jobs with custom runners depend on it and check its output. * fix(ci): add repository check to unbound_deps_tests workflow Add 'if: github.repository == huggingface/lerobot' check to build-and-push-docker job to prevent runner group access errors on forks, matching the pattern used in nightly.yml * fix(ci): add repository check to full_tests workflow Add 'if: github.repository == huggingface/lerobot' check to build-and-push-docker and gpu-tests jobs to prevent runner group access errors on forks * refactor(ci): remove redundant check from gpu-tests job gpu-tests depends on build-and-push-docker via needs, so it will automatically skip when the parent job is skipped * refactor(ci): remove unnecessary fork check from full-tests job full-tests runs on ubuntu-latest which is available to all forks, no need to restrict it --------- Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>	2026-02-10 15:21:40 +01:00
Jai Kumaar Ratadia	d2d01399d6	docs: clarify installation steps are sequential, not optional (#2925 ) * docs: clarify installation steps are sequential, not optional Add intro paragraph noting conda is one path (not the only one) and number the three sections as steps so readers understand miniforge and environment setup are prerequisites, not independent choices. * Update installation guide link for LeRobot Signed-off-by: Jai Kumaar Ratadia <jaikumaarratadia@gmail.com> * Fix link formatting in installation guide again Signed-off-by: Jai Kumaar Ratadia <jaikumaarratadia@gmail.com> --------- Signed-off-by: Jai Kumaar Ratadia <jaikumaarratadia@gmail.com> Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>	2026-02-10 15:18:32 +01:00
Aoqun Jin	5eba4ce6f4	Change LIBERO init_state_id when reset. (#2899 ) * Change LIBERO init_state_id when reset. Signed-off-by: Aoqun Jin <aojiaojiao@foxmail.com> * Change LIBERO init_state_id when reset. Signed-off-by: Aoqun Jin <aojiaojiao@foxmail.com> * pre-commit run --------- Signed-off-by: Aoqun Jin <aojiaojiao@foxmail.com> Co-authored-by: Jade Choghari <chogharijade@gmail.com> Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>	2026-02-10 16:39:17 +03:00
Stepan Feduniak	cca0296cd6	fix(pipeline): use FeatureType for STATE features in Libero processor (#2888 ) * fix the types * pre-commit --------- Co-authored-by: Jade Choghari <chogharijade@gmail.com> Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>	2026-02-10 15:55:11 +03:00
Steven Palma	489cb7b6b9	fix(scripts): correct can import check (#2937 )	2026-02-09 16:58:32 +01:00
Reece O'Mahoney	e14bdf57d0	Convert tensors to scalars (#2903 ) Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>	2026-02-09 14:46:12 +01:00
Reece O'Mahoney	97e7e0f9ed	feat(datasets): improve image transform support (#2885 ) * improve image transform support * add tests * Add stricter transform check and extra test * improve subclass check	2026-02-05 15:39:58 +01:00
jwang078	0f39248445	Small docstring fix in diffusion configuration (#2847 )	2026-02-03 19:19:00 +01:00
Iori Yanokura	a6370dd783	fix(wandb): truncate init tags to 64-character limit (#995 )	2026-02-03 14:17:04 +01:00
Michel Aractingi	14a15f90e7	Add missing RL config options: add_ee_pose_to_observation and gripper_penalty_in_reward (#2873 ) * fix(RL) add missing config arguments * respond to copilot review * fix(revert penalty in reward): reverting gripper penalty addition in reward. This is already done in compute_loss_discrete_critic. --------- Co-authored-by: CarolinePascal <caroline8.pascal@gmail.com>	2026-02-02 22:14:03 +01:00
Hirokazu Ishida	9c24a09665	docs: update document in response to Simplify configs PR (#1596 ) * docs: update document input/output_shapes -> input/output_features * fix inconsistent quote (suggested by copilot reviewer) * docs: shapes => PolicyFeature * docs: relfect normalization_mapping and remove outdated	2026-02-02 20:05:58 +01:00