fix(language): address review — tools accessor, motion docs, conditional collate

* **`meta.tools` actually reads `info.json["tools"]`.** `DatasetInfo`
  had no `tools` field, so `from_dict` silently dropped the key (it
  warned about unknown fields then discarded them) and the property
  always returned `DEFAULT_TOOLS`. Added `tools: list[dict] | None`
  to the dataclass; `to_dict()` drops it when unset so existing
  datasets keep a clean `info.json`. Fixed the accessor to read
  `self.info.tools` (the previous `.get(...)` would have raised
  AttributeError on the dataclass anyway). Added regression tests:
  fallback when absent, round-trip from disk, and round-trip
  through `DatasetInfo.from_dict` / `to_dict`.

* **`motion` is not view-dependent — fix the docs.** The mdx claimed
  rows of style `motion` must carry `camera`, but `VIEW_DEPENDENT_STYLES
  = {"vqa", "trace"}` and the validator agrees: motion primitives are
  joint/Cartesian-frame, not pixel-space. Updated both call-out
  paragraphs in `language_and_recipes.mdx`.

* **Conditional `collate_fn` swap.** Added `meta.has_language_columns`
  and gate the `lerobot_collate_fn` swap in `lerobot_train.py` on it,
  so non-language datasets keep PyTorch's `default_collate`. Also
  added a pass-through test in `test_collate.py` that asserts on a
  plain tensor batch the custom collate matches `default_collate`
  key-for-key, plus a test for the `None`-sample drop path.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.github/workflows/documentation-upload-pr.yml

@@ -33,7 +33,7 @@ jobs:
       github.event.workflow_run.event == 'pull_request' &&
       github.event.workflow_run.conclusion == 'success' &&
       github.repository == 'huggingface/lerobot'
-    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@9ad2de8582b56c017cb530c1165116d40433f1c6  # main
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@2430c1ec91d04667414e2fa31ecfc36c153ea391  # main
     with:
       package_name: lerobot
     secrets:

.github/workflows/documentation.yml

@@ -55,7 +55,7 @@ jobs:
       github.repository == 'huggingface/lerobot'
     permissions:
       contents: read
-    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@90b4ee2c10b81b5c1a6367c4e6fc9e2fb510a7e3  # main
+    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@2430c1ec91d04667414e2fa31ecfc36c153ea391  # main
     with:
       commit_sha: ${{ github.sha }}
       package: lerobot
@@ -78,7 +78,7 @@ jobs:
     permissions:
       contents: read
       pull-requests: write
-    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@90b4ee2c10b81b5c1a6367c4e6fc9e2fb510a7e3  # main
+    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@2430c1ec91d04667414e2fa31ecfc36c153ea391  # main
     with:
       commit_sha: ${{ github.event.pull_request.head.sha }}
       pr_number: ${{ github.event.number }}

.github/workflows/stale.yml

@@ -24,14 +24,14 @@ on:
 
 env:
   CLOSE_ISSUE_MESSAGE: >
-    This issue was closed because it has been stalled for 14 days with no activity.
+    This issue was closed because it has been stalled for 30 days with no activity.
     Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
   CLOSE_PR_MESSAGE: >
-    This PR was closed because it has been stalled for 21 days with no activity.
+    This PR was closed because it has been stalled for 30 days with no activity.
     Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
   WARN_ISSUE_MESSAGE: >
     This issue has been automatically marked as stale because it has not had
-    recent activity (6 months). It will be closed if no further activity occurs.
+    recent activity (1 year). It will be closed if no further activity occurs.
     Any change, comment or update to this issue will reset this count.
     Thank you for your contributions.
   WARN_PR_MESSAGE: >
@@ -59,10 +59,10 @@ jobs:
           stale-pr-label: stale
           exempt-issue-labels: never-stale
           exempt-pr-labels: never-stale
-          days-before-issue-stale: 180
-          days-before-issue-close: 14
+          days-before-issue-stale: 365
+          days-before-issue-close: 30
           days-before-pr-stale: 365
-          days-before-pr-close: 21
+          days-before-pr-close: 30
           delete-branch: true
           close-issue-message: ${{ env.CLOSE_ISSUE_MESSAGE }}
           close-pr-message: ${{ env.CLOSE_PR_MESSAGE }}

AGENTS.md

@@ -1,5 +1,7 @@
 This file provides guidance to AI agents when working with code in this repository.
 
+> **User-facing help → [`AGENT_GUIDE.md`](./AGENT_GUIDE.md)** (SO-101 setup, recording, picking a policy, training duration, eval — with copy-pasteable commands).
+
 ## Project Overview
 
 LeRobot is a PyTorch-based library for real-world robotics, providing datasets, pretrained policies, and tools for training, evaluation, data collection, and robot control. It integrates with Hugging Face Hub for model/dataset sharing.

AGENT_GUIDE.md

@@ -0,0 +1,410 @@
+# AGENT_GUIDE.md — LeRobot Helper for AI Agents & Users
+
+This file is a practical, copy-paste-friendly companion for any AI agent (Cursor, Claude, ChatGPT, Codex, etc.) helping a user work with LeRobot. It complements [`AGENTS.md`](./AGENTS.md) (dev/contributor context) with **user-facing guidance**: how to start, what to train, how long, how to record, and how to calibrate an SO-101.
+
+---
+
+## 1. Start here — ask the user first (MANDATORY)
+
+Before suggesting any command, an agent MUST ask the user at least these questions and wait for answers:
+
+1. **What's your goal?** (e.g. "teach my SO-101 to fold a cloth", "train a policy on an existing HF dataset", "contribute a PR", "understand the codebase")
+2. **What hardware do you have?**
+   - Robot: none / SO-100 / SO-101 / Koch / LeKiwi / Reachy / other
+   - Teleop: leader arm / phone / keyboard / gamepad / none
+   - Cameras: how many, resolution, fixed or moving?
+3. **What machine will you train on?**
+   - GPU model + VRAM (e.g. "laptop 3060 6 GB", "RTX 4090 24 GB", "A100 80 GB", "CPU only")
+   - OS: macOS / Linux / Windows
+4. **Skill level & time budget?** First time, some ML, experienced? Hours, days, a weekend?
+5. **Do you already have a dataset?** Yes (HF repo id?) / no / want to record one
+6. **How can I help right now?** (pick one concrete next step)
+
+Only after you have answers, propose a concrete path. If something is ambiguous, ask again rather than guessing. Bias toward **the simplest thing that works** for the user's hardware and goal.
+
+---
+
+## 2. LeRobot in 60 seconds
+
+LeRobot = **datasets + policies + envs + robot control**, unified by a small set of strong abstractions.
+
+- **`LeRobotDataset`** — episode-aware dataset (video or images + actions + state), loadable from the Hub or disk.
+- **Policies** (`ACT`, `Diffusion`, `SmolVLA`, `π0`, `π0.5`, `Wall-X`, `X-VLA`, `VQ-BeT`, `TD-MPC`, …) — all inherit `PreTrainedPolicy` and can be pushed/pulled from the Hub.
+- **Processors** — small composable transforms between dataset → policy → robot.
+- **Envs** (sim) and **Robots** (real) — same action/observation contract so code swaps cleanly.
+- **CLI** — `lerobot-record`, `lerobot-train`, `lerobot-eval`, `lerobot-teleoperate`, `lerobot-calibrate`, `lerobot-find-port`, `lerobot-setup-motors`, `lerobot-replay`.
+
+See [`AGENTS.md`](./AGENTS.md) for repo architecture.
+
+---
+
+## 3. Quickstart paths (pick one)
+
+### Path A — "I have an SO-101 and want my first trained policy"
+
+Go to §4 (SO-101 end-to-end), then §5 (data tips), then §6 (pick a policy — likely **ACT**), then §7 (how long), then §8 (eval).
+
+### Path B — "No hardware, I want to train on an existing dataset"
+
+Skip §4. Pick a policy in §6, pick a duration in §7, then run `lerobot-train` per §4.9 with a Hub `--dataset.repo_id` and an `--env.type` for eval. Finish with §8.
+
+### Path C — "I just want to understand the codebase"
+
+Read §2 above, then `AGENTS.md` "Architecture", then open `src/lerobot/policies/act/` and `src/lerobot/datasets/lerobot_dataset.py` as canonical examples.
+
+---
+
+## 4. SO-101 end-to-end cheat-sheet
+
+Full details in [`docs/source/so101.mdx`](./docs/source/so101.mdx) and [`docs/source/il_robots.mdx`](./docs/source/il_robots.mdx). Minimum commands in order. Confirm arms are assembled + powered before issuing.
+
+**4.1 Install**
+
+```bash
+pip install 'lerobot[feetech]'              # SO-100/SO-101 motor stack
+# pip install 'lerobot[all]'                # everything
+# pip install 'lerobot[aloha,pusht]'        # specific features
+# pip install 'lerobot[smolvla]'            # add SmolVLA deps
+git lfs install && git lfs pull
+hf auth login                               # required to push datasets/policies
+```
+
+Contributors can alternatively use `uv sync --locked --extra feetech` (see `AGENTS.md`).
+
+**4.2 Find USB ports** — run once per arm, unplug when prompted.
+
+```bash
+lerobot-find-port
+```
+
+macOS: `/dev/tty.usbmodem...`; Linux: `/dev/ttyACM0` (may need `sudo chmod 666 /dev/ttyACM0`).
+
+**4.3 Setup motor IDs & baudrate** (one-time, per arm)
+
+```bash
+lerobot-setup-motors --robot.type=so101_follower --robot.port=<FOLLOWER_PORT>
+lerobot-setup-motors --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>
+```
+
+**4.4 Calibrate** — center all joints, press Enter, sweep each joint through its full range. The `id` is the calibration key — reuse it everywhere.
+
+```bash
+lerobot-calibrate --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower
+lerobot-calibrate --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>   --teleop.id=my_leader
+```
+
+**4.5 Teleoperate** (sanity check, no recording)
+
+```bash
+lerobot-teleoperate \
+  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
+  --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>  --teleop.id=my_leader \
+  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+  --display_data=true
+```
+
+> **Feetech timeout / comms error on SO-100 / SO-101?** Before touching software, check the **red motor LEDs** on the daisy chain.
+>
+> - **All steady red, gripper → base chain** → wiring OK.
+> - **One or more motors dark / chain stops mid-way** → wiring issue: reseat the 3-pin cables, check the controller-board power supply, and make sure each motor is fully clicked in.
+> - **LEDs blinking** → the motor is in an **error state**: usually overload (forcing a joint past its limit) **or wrong power supply voltage**. SO-100 / SO-101 ship in two variants — a **5 V / 7.4 V** build and a **12 V** build — they are NOT interchangeable. Using a 12 V PSU on a 5 V / 7.4 V arm (or vice-versa) will trip this error; confirm your motor variant before powering up.
+>
+> Most "timeout" errors are physical, not code.
+
+**4.6 Record a dataset** — keys: **→** next, **←** redo, **ESC** finish & upload.
+
+```bash
+HF_USER=$(NO_COLOR=1 hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+
+lerobot-record \
+  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
+  --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>  --teleop.id=my_leader \
+  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+  --dataset.repo_id=${HF_USER}/my_task \
+  --dataset.single_task="<describe the task in one sentence>" \
+  --dataset.num_episodes=50 \
+  --dataset.episode_time_s=30 \
+  --dataset.reset_time_s=10 \
+  --display_data=true
+```
+
+**4.7 Visualize** — **always** do this before training. Look for missing frames, camera blur, unreachable targets, inconsistent object positions.
+After upload: https://huggingface.co/spaces/lerobot/visualize_dataset → paste `${HF_USER}/my_task`. Works for **any LeRobot-formatted Hub dataset** — use it to scout other datasets, inspect episode quality, or debug your own data before retraining.
+
+**4.8 Replay an episode** (sanity check)
+
+```bash
+lerobot-replay --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
+  --dataset.repo_id=${HF_USER}/my_task --dataset.episode=0
+```
+
+**4.9 Train** (default: ACT — fastest, lowest memory). Apple silicon: `--policy.device=mps`. See §6/§7 for policy and duration.
+
+```bash
+lerobot-train \
+  --dataset.repo_id=${HF_USER}/my_task \
+  --policy.type=act \
+  --policy.device=cuda \
+  --output_dir=outputs/train/act_my_task \
+  --job_name=act_my_task \
+  --batch_size=8 \
+  --wandb.enable=true \
+  --policy.repo_id=${HF_USER}/act_my_task
+```
+
+**4.10 Evaluate on the real robot** — compare success rate to a teleoperated baseline.
+
+```bash
+lerobot-record \
+  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
+  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+  --dataset.repo_id=${HF_USER}/eval_my_task \
+  --dataset.single_task="<same task description as training>" \
+  --dataset.num_episodes=10 \
+  --policy.path=${HF_USER}/act_my_task
+```
+
+---
+
+## 5. Data collection tips (beginner → reliable policy)
+
+Good data beats clever models. Adopt these defaults and deviate only with evidence.
+
+### 5.1 Setup & ergonomics
+
+- **Fix the rig and cameras** before touching the software. If the rig vibrates or the operator gets frustrated, fix that first — more bad data won't help.
+- **Lighting matters more than resolution.** Diffuse, consistent light. Avoid moving shadows.
+- **"Can you do the task from the camera view alone?"** If no, your cameras are wrong. Fix before recording.
+- Enable **action interpolation** for rollouts when available for smoother trajectories.
+
+### 5.2 Practice before you record
+
+- Do 5–10 demos without recording. Build a deliberate, repeatable strategy.
+- Hesitant or inconsistent demos teach the model hesitation.
+
+### 5.3 Quality over speed
+
+Deliberate, high-quality execution beats fast sloppy runs. Optimize for speed only **after** strategy is dialed in — never trade quality for it.
+
+### 5.4 Consistency within and across episodes
+
+Same grasp, approach vector, and timing. Coherent strategies are much easier to learn than wildly varying movements.
+
+### 5.5 Start small, then extend (the golden rule)
+
+- **First 50 episodes = constrained version** of the task: one object, fixed position, fixed camera setup, one operator.
+- Train a quick ACT model. See what fails.
+- **Then add diversity** along one axis at a time: more positions → more lighting → more objects → more operators.
+- Don't try to collect the "perfect dataset" on day one. Iterate.
+
+### 5.6 Policy choice for beginners
+
+- **Laptop / first time / want results fast → ACT.** Works surprisingly well, trains fast even on a laptop GPU.
+- **Bigger GPU / language-conditioned / multi-task → SmolVLA.** Unfreezing the vision encoder (see §7) is a big win here.
+- Defer π0 / π0.5 / Wall-X / X-VLA until you have a proven ACT baseline and a 20+ GB GPU.
+
+### 5.7 Recommended defaults for your first task
+
+| Setting          | Value                                                                                                                                                 |
+| ---------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------- |
+| Episodes         | **50** to start, scale to 100–300 after first training                                                                                                |
+| Episode length   | 20–45 s (shorter is fine for grasp/place)                                                                                                             |
+| Reset time       | 10 s                                                                                                                                                  |
+| FPS              | 30                                                                                                                                                    |
+| Cameras          | **2 cameras recommended**: 1 fixed front + 1 wrist. Multi-view often outperforms single-view. A single fixed camera also works to keep things simple. |
+| Task description | Short, specific, action-phrased sentence                                                                                                              |
+
+### 5.8 Troubleshooting signal
+
+- Policy fails at one specific stage → record 10–20 more episodes **targeting that stage**.
+- Policy flaps / oscillates → likely inconsistent demos, or need more training; re-record worst episodes (use **←** to redo).
+- Policy ignores the object → camera framing or lighting issue, not a model issue.
+
+See also: [What makes a good dataset](https://huggingface.co/blog/lerobot-datasets#what-makes-a-good-dataset).
+
+---
+
+## 6. Which policy should I train?
+
+Match the policy to the user's **GPU memory** and **time budget**. Numbers below come from an internal profiling run (one training update per policy). They are **indicative only** — see caveats.
+
+### 6.1 Profiling snapshot (indicative)
+
+All policies typically train for **5–10 epochs** (see §7).
+
+| Policy      | Batch | Update (ms) | Peak GPU mem (GB) | Best for                                                                                         |
+| ----------- | ----: | ----------: | ----------------: | ------------------------------------------------------------------------------------------------ |
+| `act`       |     4 |    **83.9** |          **0.94** | First-time users, laptops, single-task. Fast and reliable.                                       |
+| `diffusion` |     4 |       168.6 |              4.94 | Multi-modal action distributions; needs mid-range GPU.                                           |
+| `smolvla`   |     1 |       357.8 |              3.93 | Language-conditioned, multi-task, small VLA. **Unfreeze vision encoder for big gains** (see §7). |
+| `xvla`      |     1 |       731.6 |             15.52 | Large VLA, multi-task.                                                                           |
+| `wall_x`    |     1 |       716.5 |             15.95 | Large VLA with world-model objective.                                                            |
+| `pi0`       |     1 |       940.3 |             15.50 | Strong large VLA baseline (Physical Intelligence).                                               |
+| `pi05`      |     1 |      1055.8 |             16.35 | Newer π policy; similar footprint to `pi0`.                                                      |
+
+**Critical caveats:**
+
+- **Optimizer:** measured with **SGD**. LeRobot's default is **AdamW**, which keeps extra optimizer state → **peak memory will be noticeably higher** with the default, especially for `pi0`, `pi05`, `wall_x`, `xvla`.
+- **Batch size:** the large policies were profiled at batch 1. In practice use a **larger batch** for stable training (see §7.4). Memory scales roughly linearly with batch.
+
+### 6.2 Decision rules
+
+- **< 8 GB VRAM (laptop, 3060, M-series Mac):** → `act`. Maybe `diffusion` if you have ~6–8 GB free.
+- **12–16 GB VRAM (4070/4080, A4000):** → `smolvla` with defaults, or `act`/`diffusion` with larger batch. `pi0`/`pi05`/`wall_x`/`xvla` feasible only with small batch + gradient accumulation.
+- **24+ GB VRAM (3090/4090/A5000):** → any policy. Prefer `smolvla` (unfrozen) for multi-task; `act` for single-task grasp-and-place (still often the best ROI). Could experiment with `pi0` or `pi05` or `xvla`
+- **80 GB (A100/H100):** → any, with healthy batch. `pi05`, `xvla`, `wall_x` become comfortable.
+- **CPU only:** → don't train here. Use Google Colab (see [`docs/source/notebooks.mdx`](./docs/source/notebooks.mdx)) or a rented GPU.
+
+---
+
+## 7. How long should I train?
+
+Robotics imitation learning usually converges in a **few epochs over the dataset**, not hundreds of thousands of raw steps. Think **epochs first**, then translate to steps.
+
+### 7.1 Rule of thumb
+
+- **Typical total: 5–10 epochs.** Start at 5, eval, then decide if more helps.
+- Very small datasets (< 30 episodes) may want slightly more epochs — but first, **collect more data**.
+- VLAs with a pretrained vision backbone typically need **fewer** epochs than training from scratch.
+
+### 7.2 Steps ↔ epochs conversion
+
+```
+total_frames     = sum of frames over all episodes      # e.g. 50 eps × 30 fps × 30 s ≈ 45,000
+steps_per_epoch  = ceil(total_frames / batch_size)
+total_steps      = epochs × steps_per_epoch
+```
+
+Examples for `--batch_size=8`:
+
+| Dataset size            |  Frames | Steps / epoch | 5 epochs | 10 epochs |
+| ----------------------- | ------: | ------------: | -------: | --------: |
+| 50 eps × 30 s @ 30 fps  |  45,000 |        ~5,625 |      28k |       56k |
+| 100 eps × 30 s @ 30 fps |  90,000 |       ~11,250 |      56k |      113k |
+| 300 eps × 30 s @ 30 fps | 270,000 |       ~33,750 |     169k |      338k |
+
+Pass the resulting total with `--steps=<N>`; eval at intermediate checkpoints (`outputs/train/.../checkpoints/`).
+
+### 7.3 Per-policy starting points (single-task, ~50 episodes)
+
+| Policy         | Batch | Steps (first run) | Notes                                                             |
+| -------------- | ----: | ----------------: | ----------------------------------------------------------------- |
+| `act`          |  8–16 |           30k–80k | Usually converges under 50k for single-task.                      |
+| `diffusion`    |  8–16 |          80k–150k | Benefits from longer training than ACT.                           |
+| `smolvla`      |   4–8 |           30k–80k | Pretrained VLM → converges fast.                                  |
+| `pi0` / `pi05` |   1–4 |           30k–80k | Memory-bound; use gradient accumulation for effective batch ≥ 16! |
+
+### 7.4 Batch size guidance
+
+- **Bigger batch is preferable** for stable gradients on teleop data.
+- If GPU memory is the bottleneck, use **gradient accumulation** to raise _effective_ batch without raising peak memory.
+- Scale **learning rate** gently with batch; most LeRobot defaults work fine for a 2–4× batch change.
+
+### 7.5 Scale LR schedule & checkpoints with `--steps`
+
+LeRobot's default schedulers (e.g. SmolVLA's cosine decay) use `scheduler_decay_steps=30_000`, which is sized for long training runs. When you shorten training (e.g. 5k–10k steps on a small dataset), **scale the scheduler down to match** — otherwise the LR stays near the peak and never decays. Same for checkpoint frequency.
+
+```bash
+lerobot-train ... \
+  --steps=5000 \
+  --policy.scheduler_decay_steps=5000 \
+  --save_freq=5000
+```
+
+Rule of thumb: set `scheduler_decay_steps ≈ steps`, and `save_freq` to whatever granularity you want for eval (e.g. every 1k–5k steps). Match `scheduler_warmup_steps` proportionally if your run is very short.
+
+### 7.6 SmolVLA: unfreeze the vision encoder for real gains
+
+SmolVLA ships with `freeze_vision_encoder=True`. Unfreezing usually **improves performance substantially** on specialized tasks, at the cost of more VRAM and slower steps. Enable with:
+
+```bash
+lerobot-train ... --policy.type=smolvla \
+  --policy.freeze_vision_encoder=false \
+  --policy.train_expert_only=false
+```
+
+### 7.7 Signals to stop / keep going
+
+- Train loss plateaus → stop, save a Hub checkpoint.
+- Train loss still dropping and you're under 10 epochs → keep going.
+
+---
+
+## 8. Evaluation & benchmarks
+
+Two flavors of evaluation:
+
+### 8.1 Real-robot eval (SO-101, etc.)
+
+Reuse `lerobot-record` with `--policy.path` to run the trained policy on-robot and save the run as an eval dataset. Convention: prefix the dataset with `eval_`.
+
+```bash
+lerobot-record \
+  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
+  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+  --dataset.repo_id=${HF_USER}/eval_my_task \
+  --dataset.single_task="<same task description used during training>" \
+  --dataset.num_episodes=10 \
+  --policy.path=${HF_USER}/act_my_task
+```
+
+Report success rate across episodes. Compare to a teleoperated baseline and to an earlier checkpoint to catch regressions.
+
+### 8.2 Sim-benchmark eval
+
+For policies trained on sim datasets (PushT, Aloha, LIBERO, MetaWorld, RoboCasa, …) use `lerobot-eval` against the matching `env.type`:
+
+```bash
+lerobot-eval \
+  --policy.path=${HF_USER}/diffusion_pusht \
+  --env.type=pusht \
+  --eval.n_episodes=50 \
+  --eval.batch_size=10 \
+  --policy.device=cuda
+```
+
+- Use `--policy.path=outputs/train/.../checkpoints/<step>/pretrained_model` for local checkpoints.
+- `--eval.n_episodes` should be ≥ 50 for a stable success-rate estimate.
+- Available envs live in `src/lerobot/envs/`. See [`docs/source/libero.mdx`](./docs/source/libero.mdx), [`metaworld.mdx`](./docs/source/metaworld.mdx), [`robocasa.mdx`](./docs/source/robocasa.mdx), [`vlabench.mdx`](./docs/source/vlabench.mdx) for specific benchmarks.
+- To add a new benchmark, see [`docs/source/adding_benchmarks.mdx`](./docs/source/adding_benchmarks.mdx) and [`envhub.mdx`](./docs/source/envhub.mdx).
+
+### 8.2b Dockerfiles for benchmark eval
+
+Benchmark envs have native dependencies that are painful to install locally. The repo ships **pre-baked Dockerfiles** for each supported benchmark — use these to run `lerobot-eval` in a reproducible environment:
+
+| Benchmark   | Dockerfile                                                                             |
+| ----------- | -------------------------------------------------------------------------------------- |
+| LIBERO      | [`docker/Dockerfile.benchmark.libero`](./docker/Dockerfile.benchmark.libero)           |
+| LIBERO+     | [`docker/Dockerfile.benchmark.libero_plus`](./docker/Dockerfile.benchmark.libero_plus) |
+| MetaWorld   | [`docker/Dockerfile.benchmark.metaworld`](./docker/Dockerfile.benchmark.metaworld)     |
+| RoboCasa    | [`docker/Dockerfile.benchmark.robocasa`](./docker/Dockerfile.benchmark.robocasa)       |
+| RoboCerebra | [`docker/Dockerfile.benchmark.robocerebra`](./docker/Dockerfile.benchmark.robocerebra) |
+| RoboMME     | [`docker/Dockerfile.benchmark.robomme`](./docker/Dockerfile.benchmark.robomme)         |
+| RoboTwin    | [`docker/Dockerfile.benchmark.robotwin`](./docker/Dockerfile.benchmark.robotwin)       |
+| VLABench    | [`docker/Dockerfile.benchmark.vlabench`](./docker/Dockerfile.benchmark.vlabench)       |
+
+Build and run (adapt to your benchmark):
+
+```bash
+docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-bench-robomme .
+docker run --gpus all --rm -it \
+  -v $HOME/.cache/huggingface:/root/.cache/huggingface \
+  lerobot-bench-robomme \
+  lerobot-eval --policy.path=<your_policy> --env.type=<env> --eval.n_episodes=50
+```
+
+See [`docker/README.md`](./docker/README.md) for base-image details.
+
+### 8.3 Target success rates
+
+Single-task grasp-and-place with 50 clean episodes: ACT should reach **> 70% success** on the training configuration. Less → data problem (see §5), not model problem. Expect a drop when generalizing to new positions — scale episodes or diversity to recover.
+
+---
+
+## 9. Further reading & resources
+
+- **Getting started:** [`installation.mdx`](./docs/source/installation.mdx) · [`il_robots.mdx`](./docs/source/il_robots.mdx) · [What makes a good dataset](https://huggingface.co/blog/lerobot-datasets)
+- **Per-policy docs:** browse [`docs/source/*.mdx`](./docs/source/) (policies, hardware, benchmarks, advanced training).
+- **Community:** [Discord](https://discord.com/invite/s3KuuzsPFb) · [Hub `LeRobot` tag](https://huggingface.co/datasets?other=LeRobot) · [Dataset visualizer](https://huggingface.co/spaces/lerobot/visualize_dataset)
+
+> Keep this file current. If you learn a rule that would prevent a class of user mistakes, add it here and in [`AGENTS.md`](./AGENTS.md).

MANIFEST.in

@@ -1,3 +1,4 @@
 include src/lerobot/templates/lerobot_modelcard_template.md
+include src/lerobot/templates/lerobot_rewardmodel_modelcard_template.md
 include src/lerobot/datasets/card_template.md
 include src/lerobot/envs/metaworld_config.json

docker/Dockerfile.internal

@@ -18,9 +18,8 @@
 # docker build -f docker/Dockerfile.internal -t lerobot-internal .
 
 # Configure the base image for CI with GPU access
-# TODO(Steven): Bump these versions
-ARG CUDA_VERSION=12.4.1
-ARG OS_VERSION=22.04
+ARG CUDA_VERSION=12.6.3
+ARG OS_VERSION=24.04
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}
 
 # Define Python version argument
@@ -36,16 +35,13 @@ ENV DEBIAN_FRONTEND=noninteractive \
 
 # Install Python, system dependencies, and uv (as root)
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    software-properties-common build-essential git curl \
-    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    build-essential git curl \
+    libglib2.0-0 libgl1 libegl1 ffmpeg \
     libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
     cmake pkg-config ninja-build \
-    && add-apt-repository -y ppa:deadsnakes/ppa \
-    && apt-get update \
-    && apt-get install -y --no-install-recommends \
-       python${PYTHON_VERSION} \
-       python${PYTHON_VERSION}-venv \
-       python${PYTHON_VERSION}-dev \
+    python${PYTHON_VERSION} \
+    python${PYTHON_VERSION}-venv \
+    python${PYTHON_VERSION}-dev \
     && curl -LsSf https://astral.sh/uv/install.sh | sh \
     && mv /root/.local/bin/uv /usr/local/bin/uv \
     && useradd --create-home --shell /bin/bash user_lerobot \

docs/source/_toctree.yml

@@ -63,6 +63,8 @@
     title: SARM
   title: "Reward Models"
 - sections:
+  - local: inference
+    title: Policy Deployment (lerobot-rollout)
   - local: async
     title: Use Async Inference
   - local: rtc

docs/source/hil_data_collection.mdx

@@ -50,30 +50,30 @@ This process can be repeated iteratively: deploy, collect, fine-tune, repeat. Ea
 
 ### Teleoperator Requirements
 
-The `examples/hil` HIL scripts require **teleoperators with active motors** that can:
+The `lerobot-rollout --strategy.type=dagger` mode requires **teleoperators with active motors** that can:
 
 - Enable/disable torque programmatically
 - Move to target positions (to mirror the robot state when pausing)
 
-**Compatible teleoperators in the current `examples/hil` scripts:**
+**Compatible teleoperators:**
 
 - `openarm_mini` - OpenArm Mini
 - `so_leader` - SO100 / SO101 leader arm
 
 > [!IMPORTANT]
-> The provided `examples/hil` commands default to `bi_openarm_follower` + `openarm_mini`.
+> The provided commands default to `bi_openarm_follower` + `openarm_mini`.
 > `so_follower` + `so_leader` configs are also registered and can be used via CLI flags.
 
 ---
 
 ## Script
 
-A single script handles both synchronous and RTC-based inference. Toggle RTC with `--rtc.enabled=true`:
+Use `lerobot-rollout` with `--strategy.type=dagger` for HIL data collection. Select the inference backend with `--inference.type=sync|rtc`:
 
-| Mode                     | Flag                 | Models                |
-| ------------------------ | -------------------- | --------------------- |
-| Standard (default)       | _(no flag needed)_   | ACT, Diffusion Policy |
-| Real-Time Chunking (RTC) | `--rtc.enabled=true` | Pi0, Pi0.5, SmolVLA   |
+| Mode                     | Flag                   | Models                |
+| ------------------------ | ---------------------- | --------------------- |
+| Standard (default)       | _(no flag needed)_     | ACT, Diffusion Policy |
+| Real-Time Chunking (RTC) | `--inference.type=rtc` | Pi0, Pi0.5, SmolVLA   |
 
 ---
 
@@ -97,7 +97,7 @@ python src/lerobot/scripts/lerobot_train.py \
 **Standard inference (ACT, Diffusion Policy):**
 
 ```bash
-python examples/hil/hil_data_collection.py \
+lerobot-rollout --strategy.type=dagger \
     --robot.type=bi_openarm_follower \
     --robot.left_arm_config.port=can1 \
     --robot.left_arm_config.side=left \
@@ -108,11 +108,10 @@ python examples/hil/hil_data_collection.py \
     --teleop.port_left=/dev/ttyACM0 \
     --teleop.port_right=/dev/ttyACM1 \
     --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
-    --dataset.repo_id=your-username/hil-dataset \
+    --dataset.repo_id=your-username/rollout_hil_dataset \
     --dataset.single_task="Fold the T-shirt properly" \
     --dataset.fps=30 \
-    --dataset.episode_time_s=1000 \
-    --dataset.num_episodes=50 \
+    --strategy.num_episodes=50 \
     --interpolation_multiplier=2
 ```
 
@@ -121,11 +120,11 @@ python examples/hil/hil_data_collection.py \
 For models with high inference latency, enable RTC for smooth execution:
 
 ```bash
-python examples/hil/hil_data_collection.py \
-    --rtc.enabled=true \
-    --rtc.execution_horizon=20 \
-    --rtc.max_guidance_weight=5.0 \
-    --rtc.prefix_attention_schedule=LINEAR \
+lerobot-rollout --strategy.type=dagger \
+    --inference.type=rtc \
+    --inference.rtc.execution_horizon=20 \
+    --inference.rtc.max_guidance_weight=5.0 \
+    --inference.rtc.prefix_attention_schedule=LINEAR \
     --robot.type=bi_openarm_follower \
     --robot.left_arm_config.port=can1 \
     --robot.left_arm_config.side=left \
@@ -136,11 +135,10 @@ python examples/hil/hil_data_collection.py \
     --teleop.port_left=/dev/ttyACM0 \
     --teleop.port_right=/dev/ttyACM1 \
     --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
-    --dataset.repo_id=your-username/hil-rtc-dataset \
+    --dataset.repo_id=your-username/rollout_hil_rtc_dataset \
     --dataset.single_task="Fold the T-shirt properly" \
     --dataset.fps=30 \
-    --dataset.episode_time_s=1000 \
-    --dataset.num_episodes=50 \
+    --strategy.num_episodes=50 \
     --interpolation_multiplier=3
 ```
 
@@ -235,7 +233,7 @@ This HIL data collection approach builds on ideas from interactive imitation lea
 
 - **HG-DAgger** (Kelly et al., 2019) made this practical for robotics: a human expert monitors the robot and only intervenes when needed, rather than labeling every state. The gating between autonomous and human control is exactly the pause → takeover → return-to-policy loop used in the scripts here.
 
-- **RaC** (Hu et al., 2025) scales this loop to long-horizon tasks by explicitly decomposing interventions into **recovery** (teleoperating back to a good state) and **correction** (demonstrating the right behavior from there). This decomposition is the protocol followed by the HIL scripts in `examples/hil`.
+- **RaC** (Hu et al., 2025) scales this loop to long-horizon tasks by explicitly decomposing interventions into **recovery** (teleoperating back to a good state) and **correction** (demonstrating the right behavior from there). This decomposition is the protocol followed by the DAgger strategy in `lerobot-rollout`.
 
 - **π0.6/RECAP** (Physical Intelligence, 2025) applies the same iterative collect-and-finetune loop at scale with VLA models, showing that even large pretrained policies benefit substantially from targeted human corrections on their own failure modes. π0.6 is trained using RECAP.
 

docs/source/il_robots.mdx

@@ -509,121 +509,42 @@ hf upload ${HF_USER}/act_so101_test${CKPT} \
 
 ## Run inference and evaluate your policy
 
-You can use the `record` script from [`lerobot-record`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_record.py) with a policy checkpoint as input, to run inference and evaluate your policy. For instance, run this command or API example to run inference and record 10 evaluation episodes:
+Use `lerobot-rollout` to deploy a trained policy on your robot. You can choose different strategies depending on your needs:
 
 <hfoptions id="eval">
-<hfoption id="Command">
+<hfoption id="Base mode (no recording)">
 ```bash
-lerobot-record  \
+lerobot-rollout \
+  --strategy.type=base \
+  --policy.path=${HF_USER}/my_policy \
   --robot.type=so100_follower \
   --robot.port=/dev/ttyACM1 \
   --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
-  --robot.id=my_awesome_follower_arm \
-  --display_data=false \
-  --dataset.repo_id=${HF_USER}/eval_so100 \
-  --dataset.single_task="Put lego brick into the transparent box" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
-  # <- Teleop optional if you want to teleoperate in between episodes \
-  # --teleop.type=so100_leader \
-  # --teleop.port=/dev/ttyACM0 \
-  # --teleop.id=my_awesome_leader_arm \
-  --policy.path=${HF_USER}/my_policy
+  --task="Put lego brick into the transparent box" \
+  --duration=60
 ```
 </hfoption>
-<hfoption id="API example">
-
-<!-- prettier-ignore-start -->
-```python
-from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.datasets import LeRobotDataset
-from lerobot.utils.feature_utils import hw_to_dataset_features
-from lerobot.policies.act import ACTPolicy
-from lerobot.policies import make_pre_post_processors
-from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
-from lerobot.scripts.lerobot_record import record_loop
-from lerobot.common.control_utils import init_keyboard_listener
-from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
-
-
-NUM_EPISODES = 5
-FPS = 30
-EPISODE_TIME_SEC = 60
-TASK_DESCRIPTION = "My task description"
-HF_MODEL_ID = "<hf_username>/<model_repo_id>"
-HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
-
-# Create the robot configuration
-camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
-robot_config = SO100FollowerConfig(
-    port="/dev/tty.usbmodem58760434471", id="my_awesome_follower_arm", cameras=camera_config
-)
-
-# Initialize the robot
-robot = SO100Follower(robot_config)
-
-# Initialize the policy
-policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
-
-# Configure the dataset features
-action_features = hw_to_dataset_features(robot.action_features, "action")
-obs_features = hw_to_dataset_features(robot.observation_features, "observation")
-dataset_features = {**action_features, **obs_features}
-
-# Create the dataset
-dataset = LeRobotDataset.create(
-    repo_id=HF_DATASET_ID,
-    fps=FPS,
-    features=dataset_features,
-    robot_type=robot.name,
-    use_videos=True,
-    image_writer_threads=4,
-)
-
-# Initialize the keyboard listener and rerun visualization
-_, events = init_keyboard_listener()
-init_rerun(session_name="recording")
-
-# Connect the robot
-robot.connect()
-
-preprocessor, postprocessor = make_pre_post_processors(
-    policy_cfg=policy,
-    pretrained_path=HF_MODEL_ID,
-    dataset_stats=dataset.meta.stats,
-)
-
-for episode_idx in range(NUM_EPISODES):
-    log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
-
-    # Run the policy inference loop
-    record_loop(
-        robot=robot,
-        events=events,
-        fps=FPS,
-        policy=policy,
-        preprocessor=preprocessor,
-        postprocessor=postprocessor,
-        dataset=dataset,
-        control_time_s=EPISODE_TIME_SEC,
-        single_task=TASK_DESCRIPTION,
-        display_data=True,
-    )
-
-    dataset.save_episode()
-
-# Clean up
-robot.disconnect()
-dataset.push_to_hub()
+<hfoption id="Sentry mode (with recording)">
+```bash
+lerobot-rollout \
+  --strategy.type=sentry \
+  --strategy.upload_every_n_episodes=5 \
+  --policy.path=${HF_USER}/my_policy \
+  --robot.type=so100_follower \
+  --robot.port=/dev/ttyACM1 \
+  --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
+  --dataset.repo_id=${HF_USER}/eval_so100 \
+  --dataset.single_task="Put lego brick into the transparent box" \
+  --duration=600
 ```
-<!-- prettier-ignore-end -->
-
 </hfoption>
 </hfoptions>
 
-As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+The `--strategy.type` flag selects the execution mode:
 
-1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with (e.g. `outputs/train/eval_act_so101_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_so101_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_so101_test`).
+- `base`: Autonomous rollout with no data recording (useful for quick evaluation)
+- `sentry`: Continuous recording with auto-upload (useful for large-scale evaluation)
+- `highlight`: Ring buffer recording with keystroke save (useful for capturing interesting events)
+- `dagger`: Human-in-the-loop data collection (see [HIL Data Collection](./hil_data_collection))
+
+All strategies support `--inference.type=rtc` for smooth execution with slow VLA models (Pi0, Pi0.5, SmolVLA).

docs/source/inference.mdx

@@ -0,0 +1,261 @@
+# Policy Deployment (lerobot-rollout)
+
+`lerobot-rollout` is the single CLI for deploying trained policies on real robots. It supports multiple execution strategies and inference backends, from quick evaluation to continuous recording and human-in-the-loop data collection.
+
+## Quick Start
+
+No extra dependencies are needed beyond your robot and policy extras.
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --policy.path=lerobot/act_koch_real \
+    --robot.type=koch_follower \
+    --robot.port=/dev/ttyACM0 \
+    --task="pick up cube" \
+    --duration=30
+```
+
+This runs the policy for 30 seconds with no recording.
+
+---
+
+## Strategies
+
+Select a strategy with `--strategy.type=<name>`. Each strategy defines a different control loop with its own recording and interaction semantics.
+
+### Base (`--strategy.type=base`)
+
+Autonomous policy execution with no data recording. Use this for quick evaluation, demos, or when you only need to observe the robot.
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --task="Put lego brick into the box" \
+    --duration=60
+```
+
+| Flag             | Description                                            |
+| ---------------- | ------------------------------------------------------ |
+| `--duration`     | Run time in seconds (0 = infinite)                     |
+| `--task`         | Task description passed to the policy                  |
+| `--display_data` | Stream observations/actions to Rerun for visualization |
+
+### Sentry (`--strategy.type=sentry`)
+
+Continuous autonomous recording with periodic upload to the Hugging Face Hub. Episode boundaries are auto-computed from camera resolution and FPS so each saved episode produces a complete video file, keeping uploads efficient.
+
+Policy state (hidden state, RTC queue) persists across episode boundaries: the robot does not reset between episodes.
+
+```bash
+lerobot-rollout \
+    --strategy.type=sentry \
+    --strategy.upload_every_n_episodes=5 \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --dataset.repo_id=${HF_USER}/rollout_eval_data \
+    --dataset.single_task="Put lego brick into the box" \
+    --duration=3600
+```
+
+| Flag                                   | Description                                                 |
+| -------------------------------------- | ----------------------------------------------------------- |
+| `--strategy.upload_every_n_episodes`   | Push to Hub every N episodes (default: 5)                   |
+| `--strategy.target_video_file_size_mb` | Target video file size for episode rotation (default: auto) |
+| `--dataset.repo_id`                    | **Required.** Hub repository for the recorded dataset       |
+| `--dataset.push_to_hub`                | Whether to push to Hub on teardown (default: true)          |
+
+### Highlight (`--strategy.type=highlight`)
+
+Autonomous rollout with on-demand recording via a memory-bounded ring buffer. The robot runs continuously while the buffer captures the last N seconds of telemetry. Press the save key to flush the buffer and start live recording; press it again to save the episode.
+
+```bash
+lerobot-rollout \
+    --strategy.type=highlight \
+    --strategy.ring_buffer_seconds=30 \
+    --strategy.save_key=s \
+    --strategy.push_key=h \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=koch_follower \
+    --robot.port=/dev/ttyACM0 \
+    --dataset.repo_id=${HF_USER}/rollout_highlight_data \
+    --dataset.single_task="Pick up the red cube"
+```
+
+**Keyboard controls:**
+
+| Key                | Action                                                   |
+| ------------------ | -------------------------------------------------------- |
+| `s` (configurable) | Start recording (flushes buffer) / stop and save episode |
+| `h` (configurable) | Push dataset to Hub                                      |
+| `ESC`              | Stop the session                                         |
+
+| Flag                                   | Description                                    |
+| -------------------------------------- | ---------------------------------------------- |
+| `--strategy.ring_buffer_seconds`       | Duration of buffered telemetry (default: 30)   |
+| `--strategy.ring_buffer_max_memory_mb` | Memory cap for the ring buffer (default: 2048) |
+| `--strategy.save_key`                  | Key to toggle recording (default: `s`)         |
+| `--strategy.push_key`                  | Key to push to Hub (default: `h`)              |
+
+### DAgger (`--strategy.type=dagger`)
+
+Human-in-the-loop data collection. Alternates between autonomous policy execution and human intervention via a teleoperator. Intervention frames are tagged with `intervention=True`. Requires a teleoperator (`--teleop.type`).
+
+See the [Human-In-the-Loop Data Collection](./hil_data_collection) guide for a detailed walkthrough.
+
+**Corrections-only mode** (default): Only human correction windows are recorded. Each correction becomes one episode.
+
+```bash
+lerobot-rollout \
+    --strategy.type=dagger \
+    --strategy.num_episodes=20 \
+    --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
+    --robot.type=bi_openarm_follower \
+    --teleop.type=openarm_mini \
+    --dataset.repo_id=${HF_USER}/rollout_hil_data \
+    --dataset.single_task="Fold the T-shirt"
+```
+
+**Continuous recording mode** (`--strategy.record_autonomous=true`): Both autonomous and correction frames are recorded with time-based episode rotation (same as Sentry).
+
+```bash
+lerobot-rollout \
+    --strategy.type=dagger \
+    --strategy.record_autonomous=true \
+    --strategy.num_episodes=50 \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/ttyACM1 \
+    --dataset.repo_id=${HF_USER}/rollout_dagger_data \
+    --dataset.single_task="Grasp the block"
+```
+
+**Keyboard controls** (default input device):
+
+| Key     | Action                                      |
+| ------- | ------------------------------------------- |
+| `Space` | Pause / resume policy execution             |
+| `Tab`   | Start / stop human correction               |
+| `Enter` | Push dataset to Hub (corrections-only mode) |
+| `ESC`   | Stop the session                            |
+
+Foot pedal input is also supported via `--strategy.input_device=pedal`. Configure pedal codes with `--strategy.pedal.*` flags.
+
+| Flag                                 | Description                                             |
+| ------------------------------------ | ------------------------------------------------------- |
+| `--strategy.num_episodes`            | Number of correction episodes to record (default: 10)   |
+| `--strategy.record_autonomous`       | Record autonomous frames too (default: false)           |
+| `--strategy.upload_every_n_episodes` | Push to Hub every N episodes (default: 5)               |
+| `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard) |
+| `--teleop.type`                      | **Required.** Teleoperator type                         |
+
+---
+
+## Inference Backends
+
+Select a backend with `--inference.type=<name>`. All strategies work with both backends.
+
+### Sync (default)
+
+One policy call per control tick. The main loop blocks until the action is computed.
+
+Works with all policies. No extra flags needed.
+
+### Real-Time Chunking (`--inference.type=rtc`)
+
+A background thread produces action chunks asynchronously. The main control loop polls for the next ready action while the policy computes the next chunk in parallel.
+
+Use RTC with large, slow VLA models (Pi0, Pi0.5, SmolVLA) for smooth, continuous motion despite high inference latency.
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --inference.type=rtc \
+    --inference.rtc.execution_horizon=10 \
+    --inference.rtc.max_guidance_weight=10.0 \
+    --policy.path=${HF_USER}/pi0_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --task="Pick up the cube" \
+    --duration=60 \
+    --device=cuda
+```
+
+| Flag                                        | Description                                                    |
+| ------------------------------------------- | -------------------------------------------------------------- |
+| `--inference.rtc.execution_horizon`         | Steps to blend with previous chunk (default: varies by policy) |
+| `--inference.rtc.max_guidance_weight`       | Consistency enforcement strength (default: varies by policy)   |
+| `--inference.rtc.prefix_attention_schedule` | Blend schedule: `LINEAR`, `EXP`, `ONES`, `ZEROS`               |
+| `--inference.queue_threshold`               | Max queue size before backpressure (default: 30)               |
+
+See the [Real-Time Chunking](./rtc) guide for details on tuning RTC parameters.
+
+---
+
+## Common Flags
+
+| Flag                              | Description                                                       | Default |
+| --------------------------------- | ----------------------------------------------------------------- | ------- |
+| `--policy.path`                   | **Required.** HF Hub model ID or local checkpoint path            | --      |
+| `--robot.type`                    | **Required.** Robot type (e.g. `so100_follower`, `koch_follower`) | --      |
+| `--robot.port`                    | Serial port for the robot                                         | --      |
+| `--robot.cameras`                 | Camera configuration (JSON dict)                                  | --      |
+| `--fps`                           | Control loop frequency                                            | 30      |
+| `--duration`                      | Run time in seconds (0 = infinite)                                | 0       |
+| `--device`                        | Torch device (`cpu`, `cuda`, `mps`)                               | auto    |
+| `--task`                          | Task description (used when no dataset is provided)               | --      |
+| `--display_data`                  | Stream telemetry to Rerun visualization                           | false   |
+| `--display_ip` / `--display_port` | Remote Rerun server address                                       | --      |
+| `--interpolation_multiplier`      | Action interpolation factor                                       | 1       |
+| `--use_torch_compile`             | Enable `torch.compile` for inference                              | false   |
+| `--resume`                        | Resume a previous recording session                               | false   |
+| `--play_sounds`                   | Vocal synthesis for events                                        | true    |
+
+---
+
+## Programmatic Usage
+
+For custom deployments (e.g. with kinematics processors), use the rollout module API directly:
+
+```python
+from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
+from lerobot.rollout.inference import SyncInferenceConfig
+from lerobot.rollout.strategies import BaseStrategy
+from lerobot.utils.process import ProcessSignalHandler
+
+cfg = RolloutConfig(
+    robot=my_robot_config,
+    policy=my_policy_config,
+    strategy=BaseStrategyConfig(),
+    inference=SyncInferenceConfig(),
+    fps=30,
+    duration=60,
+    task="my task",
+)
+
+signal_handler = ProcessSignalHandler(use_threads=True)
+ctx = build_rollout_context(
+    cfg,
+    signal_handler.shutdown_event,
+    robot_action_processor=my_custom_action_processor,       # optional
+    robot_observation_processor=my_custom_obs_processor,     # optional
+)
+
+strategy = BaseStrategy(cfg.strategy)
+try:
+    strategy.setup(ctx)
+    strategy.run(ctx)
+finally:
+    strategy.teardown(ctx)
+```
+
+See `examples/so100_to_so100_EE/rollout.py` and `examples/phone_to_so100/rollout.py` for full examples with kinematics processors.

docs/source/language_and_recipes.mdx

@@ -1,7 +1,34 @@
 # Language columns and recipes
 
-LeRobot stores reusable language annotations directly next to frame data in `data/chunk-*/file-*.parquet`.
-The two optional columns are:
+Most LeRobot datasets ship with a single `task` string per episode — fine for
+short, single-instruction skills, but not enough for the longer-horizon,
+multi-modal robot policies the field is moving toward (high-level planning,
+memory, interjections, VQA, tool use). To support those policies without
+forking the dataset format, LeRobot extends `LeRobotDataset` with two optional
+language columns and a small recipe layer that turns those rows into
+chat-style training samples on the fly.
+
+The design splits cleanly into three layers:
+
+1. **Data in the dataset** — language annotations stored next to frames in
+   `data/chunk-*/file-*.parquet` as two optional columns (`language_persistent`
+   and `language_events`). Datasets without these columns keep their existing
+   behavior.
+2. **Recipe** — a YAML file that declares which annotation rows to bind and
+   how to lay them out as chat turns (`role`, `content`, optional images,
+   optional tool calls). Recipes are pure config; no Python required to add a
+   new one.
+3. **Training format** — at sample time, `RenderMessagesStep` resolves the
+   recipe against the per-frame annotations and emits HF-style `messages` plus
+   LeRobot-specific sidecars (`message_streams`, `target_message_indices`)
+   that policy processors consume.
+
+This page describes each layer in turn.
+
+## Layer 1 — language columns in the dataset
+
+The two optional columns live next to frame data in
+`data/chunk-*/file-*.parquet`:
 
 - `language_persistent`: a list of rows broadcast across every frame in an episode for state that remains active, such as `subtask`, `plan`, and `memory`.
 - `language_events`: a list of rows only on the exact frame where an event was emitted, such as `interjection`, `vqa`, and speech tool calls.
@@ -19,16 +46,17 @@ tool_calls: list[Json] | null
 ```
 
 The `camera` field tags rows whose `content` is grounded in a specific camera
-view. Rows of view-dependent styles (`vqa`, and the reserved `motion` /
-`trace`) MUST set `camera` to the matching `observation.images.*` feature key.
-Rows of every other style MUST leave `camera` as `null`. Pipeline writers and
-the validator enforce this via `validate_camera_field(style, camera)`.
+view. Rows of view-dependent styles (`vqa` and `trace`) MUST set `camera` to
+the matching `observation.images.*` feature key. Rows of every other style —
+including `motion`, which describes robot-frame primitives in joint / Cartesian
+terms — MUST leave `camera` as `null`. Pipeline writers and the validator
+enforce this via `validate_camera_field(style, camera)`.
 
 `meta/tasks.parquet` remains the canonical source for the task. The special `${task}` recipe binding always reads that task string and does not depend on language annotations.
 
-## Architecture
+### Architecture
 
-The language stack has three layers:
+The language stack itself has three internal modules backing layer 1:
 
 1. `lerobot.datasets.language` defines the schema, style registry, and `column_for_style`.
 2. `lerobot.datasets.language_render` resolves rows and renders messages.
@@ -36,7 +64,7 @@ The language stack has three layers:
 
 `LeRobotDataset` stays recipe-agnostic. It passes `language_persistent` and `language_events` through when present, and unannotated datasets keep their existing behavior.
 
-## Temporal semantics
+### Temporal semantics
 
 Persistent styles are active after emission until replaced:
 
@@ -52,9 +80,9 @@ Event styles only exist on their exact timestamp:
 
 Exact event matching has no tolerance window, so writers must stamp event rows with frame timestamps from the parquet data.
 
-## View-dependent resolution
+### View-dependent resolution
 
-For view-dependent styles (`vqa`, `motion`, `trace`), the resolver gains a
+For view-dependent styles (`vqa` and `trace`), the resolver gains a
 `camera=` filter parallel to `role=` and `tool_name=`. Datasets with multiple
 cameras typically emit one (`vqa`, `user`) + (`vqa`, `assistant`) pair per
 camera at the same timestamp; without `camera=`, those resolvers see two
@@ -73,14 +101,22 @@ ask_vqa_top:
       content:
         - { type: image, feature: observation.images.top }
         - { type: text, text: "${vqa_query}" }
-    - { role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa }
+    - {
+        role: assistant,
+        content: "${vqa}",
+        stream: high_level,
+        target: true,
+        if_present: vqa,
+      }
 ```
 
 Add one such sub-recipe per camera the dataset records.
 
-## Recipe anatomy
+## Layer 2 — recipe anatomy
 
-Recipes are YAML files backed by `TrainingRecipe` and `MessageTurn`.
+Recipes are YAML files backed by `TrainingRecipe` and `MessageTurn`. They
+declare which annotation rows to pull (via `bindings`) and how to compose them
+into chat turns (`messages`).
 
 ```yaml
 messages:
@@ -88,6 +124,13 @@ messages:
   - { role: assistant, content: "${subtask}", stream: low_level, target: true }
 ```
 
+A recipe can also branch into a weighted **blend** of sub-recipes. At sample
+time, exactly one branch is selected deterministically from the sample index,
+so different frames train different objectives (e.g. memory updates vs.
+low-level execution vs. VQA) without any Python wiring.
+
+## Layer 3 — training format
+
 Rendered samples use HF-style chat messages plus LeRobot sidecars:
 
 ```python
@@ -96,12 +139,7 @@ sample["message_streams"]
 sample["target_message_indices"]
 ```
 
-The renderer does not apply a tokenizer chat template. Policy processors decide how to serialize the messages for their backbone.
-
-## Blends
-
-Blend recipes select one weighted sub-recipe deterministically from the sample index.
-The canonical `recipes/pi05_hirobot.yaml` combines memory updates, interjection responses, high-level subtask prediction, low-level execution, and VQA.
+The renderer does not apply a tokenizer chat template. Policy processors decide how to serialize the messages for their backbone, which keeps the same dataset usable across SmolVLA, Pi0.5, and any future VLM that expects OpenAI-style chat messages.
 
 ## Graceful absence
 

docs/source/rename_map.mdx

@@ -61,17 +61,6 @@ lerobot-eval \
   --rename_map='{"observation.images.image": "observation.images.base_0_rgb", "observation.images.image2": "observation.images.left_wrist_0_rgb"}'
 ```
 
-### Recording
-
-`lerobot-record` also supports rename maps, nested under the dataset config:
-
-```bash
-lerobot-record \ # When running inference
-  --policy.path="<user>/smolVLA_finetuned" \
-  ... \
-  --dataset.rename_map='{"observation.images.glove2": "observation.images.image"}'
-```
-
 ## Alternative: edit the policy config directly
 
 If you always use the same dataset or environment, you can **edit the policy's `config.json`** so its observation keys match your data source. Then no rename map is needed.
@@ -105,10 +94,10 @@ XVLA-base has three visual inputs and `empty_cameras=0` by default. Your dataset
 
 ## Quick reference
 
-| Goal                                      | What to do                                                                  |
-| ----------------------------------------- | --------------------------------------------------------------------------- |
-| Dataset keys ≠ policy keys                | `--rename_map='{"dataset_key": "policy_key", ...}'`                         |
-| Env keys ≠ policy keys (eval)             | `--rename_map='{"env_key": "policy_key", ...}'`                             |
-| Recording with different keys (inference) | `--dataset.rename_map='{"source_key": "policy_key", ...}'`.                 |
-| Fewer cameras than policy expects         | `--policy.empty_cameras=N` (supported by PI0, PI05, PI0Fast, SmolVLA, XVLA) |
-| Avoid passing a rename map                | Edit the policy's `config.json` so its keys match your data source          |
+| Goal                                    | What to do                                                                  |
+| --------------------------------------- | --------------------------------------------------------------------------- |
+| Dataset keys ≠ policy keys              | `--rename_map='{"dataset_key": "policy_key", ...}'`                         |
+| Env keys ≠ policy keys (eval)           | `--rename_map='{"env_key": "policy_key", ...}'`                             |
+| Rollout with different keys (inference) | `--rename_map='{"source_key": "policy_key", ...}'`.                         |
+| Fewer cameras than policy expects       | `--policy.empty_cameras=N` (supported by PI0, PI05, PI0Fast, SmolVLA, XVLA) |
+| Avoid passing a rename map              | Edit the policy's `config.json` so its keys match your data source          |

docs/source/rtc.mdx

@@ -34,7 +34,7 @@ pip install -e ".[smolvla]"
 
 ### Using RTC with Pi0
 
-You can find a complete reference implementation in [eval_with_real_robot.py](examples/rtc/eval_with_real_robot.py).
+You can use `lerobot-rollout --strategy.type=base --inference.type=rtc` for RTC deployment on real robots.
 The snippet below provides a simplified pseudo-example of how RTC operates with Pi0 in your pipeline:
 
 ```python
@@ -137,8 +137,12 @@ The script generates a visualization of the denoising process, comparing standar
 ## Testing RTC with a Real Robot
 
 ```bash
-python examples/rtc/eval_with_real_robot.py \
+lerobot-rollout \
+    --strategy.type=base \
     --policy.path=${HF_USERNAME}/policy_repo_id \
+    --inference.type=rtc \
+    --inference.rtc.execution_horizon=10 \
+    --inference.rtc.max_guidance_weight=10.0 \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58FA0834591 \
     --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
@@ -178,7 +182,7 @@ visualizer = RTCDebugVisualizer()
 # ... create plots
 ```
 
-See `examples/rtc/eval_dataset.py` for a complete example of visualization.
+See `examples/rtc/eval_dataset.py` for a complete example of offline RTC visualization.
 
 ## References
 

docs/source/sarm.mdx

@@ -46,7 +46,7 @@ This ensures identical task states map to consistent progress values, even acros
 
 ## Inputs and Targets (What the new code expects)
 
-SARM is trained through its processor (`src/lerobot/policies/sarm/processor_sarm.py`), which:
+SARM is trained through its processor (`src/lerobot/rewards/sarm/processor_sarm.py`), which:
 
 - **Encodes** images and task text with CLIP (ViT-B/32) into `video_features` and `text_features`
 - **Pads/truncates** robot state into `state_features` (up to `max_state_dim`)
@@ -347,7 +347,7 @@ Use `compute_rabc_weights.py` with `--visualize-only` to visualize model predict
 <hfoption id="single_stage">
 
 ```bash
-python src/lerobot/policies/sarm/compute_rabc_weights.py \
+python -m lerobot.rewards.sarm.compute_rabc_weights \
   --dataset-repo-id your-username/your-dataset \
   --reward-model-path your-username/sarm-model \
   --visualize-only \
@@ -360,7 +360,7 @@ python src/lerobot/policies/sarm/compute_rabc_weights.py \
 <hfoption id="dense_only">
 
 ```bash
-python src/lerobot/policies/sarm/compute_rabc_weights.py \
+python -m lerobot.rewards.sarm.compute_rabc_weights \
   --dataset-repo-id your-username/your-dataset \
   --reward-model-path your-username/sarm-model \
   --visualize-only \
@@ -373,7 +373,7 @@ python src/lerobot/policies/sarm/compute_rabc_weights.py \
 <hfoption id="dual">
 
 ```bash
-python src/lerobot/policies/sarm/compute_rabc_weights.py \
+python -m lerobot.rewards.sarm.compute_rabc_weights \
   --dataset-repo-id your-username/your-dataset \
   --reward-model-path your-username/sarm-model \
   --visualize-only \
@@ -429,7 +429,7 @@ The weighting follows **Equations 8-9** from the paper:
 First, run the SARM model on all frames in your dataset to compute progress values:
 
 ```bash
-python src/lerobot/policies/sarm/compute_rabc_weights.py \
+python -m lerobot.rewards.sarm.compute_rabc_weights \
   --dataset-repo-id your-username/your-dataset \
   --reward-model-path your-username/sarm-model \
   --head-mode sparse \
@@ -465,15 +465,15 @@ This script:
 
 ### Step 5b: Train Policy with RA-BC
 
-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:
+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`) if not explicitly provided. Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
 
 ```bash
 lerobot-train \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
-  --use_rabc=true \
-  --rabc_head_mode=sparse \
-  --rabc_kappa=0.01 \
+  --sample_weighting.type=rabc \
+  --sample_weighting.head_mode=sparse \
+  --sample_weighting.kappa=0.01 \
   --output_dir=outputs/train/policy_rabc \
   --batch_size=32 \
   --steps=40000
@@ -488,12 +488,13 @@ The training script automatically:
 
 **RA-BC Arguments:**
 
-| Argument               | Description                                                | Default                            |
-| ---------------------- | ---------------------------------------------------------- | ---------------------------------- |
-| `--use_rabc`           | Enable RA-BC sample weighting                              | `false`                            |
-| `--rabc_progress_path` | Path to progress parquet file (auto-detected from dataset) | `sarm_progress.parquet` in dataset |
-| `--rabc_head_mode`     | Which SARM head's progress to use: `sparse` or `dense`     | `sparse`                           |
-| `--rabc_kappa`         | Threshold κ for high-quality samples                       | `0.01`                             |
+| Argument                           | Description                                            | Default                 |
+| ---------------------------------- | ------------------------------------------------------ | ----------------------- |
+| `--sample_weighting.type`          | Weighting strategy type (`rabc` or `uniform`)          | `rabc`                  |
+| `--sample_weighting.progress_path` | Path to progress parquet file                          | `sarm_progress.parquet` |
+| `--sample_weighting.head_mode`     | Which SARM head's progress to use: `sparse` or `dense` | `sparse`                |
+| `--sample_weighting.kappa`         | Threshold κ for high-quality samples                   | `0.01`                  |
+| `--sample_weighting.epsilon`       | Small constant for numerical stability                 | `1e-6`                  |
 
 ### Tuning RA-BC Kappa
 
@@ -511,30 +512,30 @@ The `kappa` parameter is the threshold that determines which samples get full we
 
 Monitor these WandB metrics during training:
 
-| Metric             | Healthy Range | Problem Indicator         |
-| ------------------ | ------------- | ------------------------- |
-| `rabc_mean_weight` | 0.3 - 0.8     | ≈ 1.0 means kappa too low |
-| `rabc_delta_mean`  | > 0           | Should be positive        |
-| `rabc_delta_std`   | > 0           | Variance in data quality  |
+| Metric                        | Healthy Range | Problem Indicator         |
+| ----------------------------- | ------------- | ------------------------- |
+| `sample_weight_mean_weight`   | 0.3 - 0.8     | ≈ 1.0 means kappa too low |
+| `sample_weighting/delta_mean` | > 0           | Should be positive        |
+| `sample_weighting/delta_std`  | > 0           | Variance in data quality  |
 
-**If `rabc_mean_weight ≈ 1.0`:** Your kappa is too low. Most samples have `delta > kappa` and bypass the soft-weighting entirely. RA-BC becomes equivalent to vanilla BC.
+**If `sample_weight_mean_weight ≈ 1.0`:** Your kappa is too low. Most samples have `delta > kappa` and bypass the soft-weighting entirely. RA-BC becomes equivalent to vanilla BC.
 
 **Setting kappa based on your data:**
 
-The default `kappa=0.01` was tuned for the paper's T-shirt folding task (~90s episodes at 30fps). For your dataset, check the logged `rabc_delta_mean` and `rabc_delta_std`:
+The default `kappa=0.01` was tuned for the paper's T-shirt folding task (~90s episodes at 30fps). For your dataset, check the logged `sample_weighting/delta_mean` and `sample_weighting/delta_std`:
 
 ```
 # If delta_mean ≈ 0.03 and delta_std ≈ 0.02:
 # Most deltas fall in range [0.01, 0.05]
 
 # Option 1: Set kappa = delta_mean (medium selectivity)
---rabc_kappa=0.03
+--sample_weighting.kappa=0.03
 
 # Option 2: Set kappa = delta_mean + delta_std (high selectivity)
---rabc_kappa=0.05
+--sample_weighting.kappa=0.05
 
 # Option 3: Set kappa = delta_mean + 2*delta_std (very selective)
---rabc_kappa=0.07
+--sample_weighting.kappa=0.07
 ```
 
 **When RA-BC may not help:**
@@ -550,8 +551,8 @@ accelerate launch \
   src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
-  --use_rabc=true \
-  --rabc_kappa=0.01 \
+  --sample_weighting.type=rabc \
+  --sample_weighting.kappa=0.01 \
   --output_dir=outputs/train/policy_rabc \
   --batch_size=32 \
   --steps=40000
@@ -576,7 +577,7 @@ accelerate launch \
 ### RA-BC
 
 1. **Train SARM first**: RA-BC quality depends entirely on SARM quality
-2. **Monitor `rabc_mean_weight`**: If it's ≈ 1.0, increase kappa (see [Tuning RA-BC Kappa](#tuning-ra-bc-kappa))
+2. **Monitor `sample_weight_mean_weight`**: If it's ≈ 1.0, increase kappa (see [Tuning RA-BC Kappa](#tuning-ra-bc-kappa))
 
 ---
 

docs/source/tools.mdx

@@ -6,9 +6,9 @@ runtime dispatches to a real implementation (TTS, controller, logger, …).
 
 This page covers:
 
-1. Where the tool catalog lives (PR 1).
-2. How the annotation pipeline produces tool-call atoms (PR 2).
-3. How to add your own tool (PR 3).
+1. Where the tool catalog lives.
+2. How the annotation pipeline produces tool-call atoms.
+3. How to add your own tool.
 
 ## Where tools are declared
 
@@ -29,7 +29,10 @@ Two layers.
         "parameters": {
           "type": "object",
           "properties": {
-            "text": { "type": "string", "description": "The verbatim text to speak." }
+            "text": {
+              "type": "string",
+              "description": "The verbatim text to speak."
+            }
           },
           "required": ["text"]
         }
@@ -64,12 +67,12 @@ prompt_str = tokenizer.apply_chat_template(
 ```
 
 **The implementations** — runnable Python — live under
-`src/lerobot/tools/`, one file per tool. The `say` implementation
-arrives in PR 3 and wraps Kyutai's pocket-tts model.
+`src/lerobot/tools/`, one file per tool. The canonical `say`
+implementation wraps Kyutai's pocket-tts model.
 
-## Per-row tool *invocations*
+## Per-row tool _invocations_
 
-The catalog above describes *what can be called*. The actual *call* — the
+The catalog above describes _what can be called_. The actual _call_ — the
 function name plus the argument values — is stored per-row, on the
 assistant atoms in `language_events`:
 
@@ -94,13 +97,18 @@ user_interjection_response:
   bindings:
     speech: "emitted_at(t, role=assistant, tool_name=say)"
   messages:
-    - { role: user,      content: "${task}",         stream: high_level }
-    - { role: assistant, content: "${current_plan}", stream: high_level,
-        target: true, tool_calls_from: speech }
+    - { role: user, content: "${task}", stream: high_level }
+    - {
+        role: assistant,
+        content: "${current_plan}",
+        stream: high_level,
+        target: true,
+        tool_calls_from: speech,
+      }
 ```
 
 The model's training target is one assistant turn that carries both the
-plan text *and* the `say` tool call. At inference, the runtime parses
+plan text _and_ the `say` tool call. At inference, the runtime parses
 the generated text back into structured `tool_calls` and dispatches to
 the matching implementation.
 
@@ -113,9 +121,8 @@ loop.
 ### Step 1 — declare the schema
 
 Add an entry under `meta/info.json["tools"]`. Either edit the file
-directly on disk *before* running the annotation pipeline (it'll be
-preserved) or hand it to `lerobot-annotate` via a config flag (PR 2 —
-exact CLI lands with the pipeline change).
+directly on disk _before_ running the annotation pipeline (it'll be
+preserved) or hand it to `lerobot-annotate` via a config flag.
 
 ```json
 {
@@ -129,7 +136,10 @@ exact CLI lands with the pipeline change).
         "parameters": {
           "type": "object",
           "properties": {
-            "label": { "type": "string", "description": "Short label for the saved image." }
+            "label": {
+              "type": "string",
+              "description": "Short label for the saved image."
+            }
           },
           "required": ["label"]
         }
@@ -167,12 +177,12 @@ class RecordObservationTool:
 ```
 
 One file per tool keeps dependencies isolated — `record_observation`
-might pull `pillow`, while `say` (PR 3) pulls `pocket-tts`. Users
-installing only the tools they need avoid heavy transitive deps.
+might pull `pillow`, while `say` pulls `pocket-tts`. Users installing
+only the tools they need avoid heavy transitive deps.
 
 ### Step 3 — register it
 
-Add to `src/lerobot/tools/registry.py` (PR 3):
+Add to `src/lerobot/tools/registry.py`:
 
 ```python
 from .record_observation import RecordObservationTool
@@ -184,15 +194,7 @@ That's it. At runtime `get_tools(meta)` looks up each schema in
 `meta.tools`, instantiates the matching registered class, and returns
 a name → instance dict the dispatcher can route into.
 
-## Where this fits in the three-PR stack
-
-| Layer | PR | What lands |
-|---|---|---|
-| Catalog storage in `meta/info.json` + `meta.tools` accessor | PR 1 | This page; `SAY_TOOL_SCHEMA`, `DEFAULT_TOOLS` constants in `lerobot.datasets.language`; `LeRobotDatasetMetadata.tools` property |
-| Annotation pipeline writes `tools` to meta after a run; honors anything users pre-populated | PR 2 | `lerobot-annotate` ensures `meta/info.json["tools"]` includes the canonical `say` and merges any user-declared tools |
-| Runnable implementations under `src/lerobot/tools/`; runtime dispatcher; `say.py` wired to Kyutai's pocket-tts | PR 3 | One file per tool; `Tool` protocol; `TOOL_REGISTRY`; optional `[tools]` extra in `pyproject.toml` |
-
-If you want to use a tool *without* writing an implementation (e.g. for
+If you want to use a tool _without_ writing an implementation (e.g. for
 training-time chat-template formatting only), step 1 alone is enough —
-the model still learns to *generate* the call. Steps 2 and 3 are only
-needed to actually *execute* it at inference.
+the model still learns to _generate_ the call. Steps 2 and 3 are only
+needed to actually _execute_ it at inference.

docs/source/unitree_g1.mdx

@@ -274,7 +274,8 @@ python src/lerobot/scripts/lerobot_train.py \
 Once trained, we recommend deploying policies using inference-time RTC:
 
 ```bash
-python examples/rtc/eval_with_real_robot.py \
+lerobot-rollout \
+  --strategy.type=base \
   --policy.path=your-username/your-repo-id \
   --policy.device=cuda \
   --robot.type=unitree_g1 \
@@ -284,7 +285,7 @@ python examples/rtc/eval_with_real_robot.py \
   --task="task_description" \
   --duration=1000 \
   --fps=30 \
-  --rtc.enabled=true
+  --inference.type=rtc
 ```
 
 ---

docs/source/xvla.mdx

@@ -220,7 +220,7 @@ REAL_DIM = 12
 # Postprocessing: Trim 20D predictions to 12D for deployment
 ```
 
-See the [action_hub.py](/home/jade_choghari/robot/lerobot/src/lerobot/policies/xvla/action_hub.py) implementation for details.
+See the [action_hub.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/action_hub.py) implementation for details.
 
 #### Auto Action Mode (Recommended)
 
@@ -519,9 +519,9 @@ If you use X-VLA in your research, please cite:
 
 - [X-VLA Paper](https://arxiv.org/pdf/2510.10274)
 - [LeRobot Documentation](https://github.com/huggingface/lerobot)
-- [Action Registry Implementation](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/action_hub.py)
-- [Processor Implementation](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/processor_xvla.py)
-- [Model Configuration](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/configuration_xvla.py)
+- [Action Registry Implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/action_hub.py)
+- [Processor Implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/processor_xvla.py)
+- [Model Configuration](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/configuration_xvla.py)
 
 ## Contributing
 

examples/dataset/slurm_compute_rabc.py

@@ -69,7 +69,7 @@ class ComputeProgressShards(PipelineStep):
         import torch
         from tqdm import tqdm
 
-        from lerobot.policies.sarm.compute_rabc_weights import (
+        from lerobot.rewards.sarm.compute_rabc_weights import (
             generate_all_frame_indices,
             interpolate_progress,
             load_sarm_resources,

examples/hil/hil_utils.py

@@ -1,226 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""Shared utilities for Human-in-the-Loop data collection scripts."""
-
-import logging
-import time
-from dataclasses import dataclass, field
-from pathlib import Path
-
-from lerobot.common.control_utils import is_headless
-from lerobot.processor import (
-    IdentityProcessorStep,
-    RobotAction,
-    RobotObservation,
-    RobotProcessorPipeline,
-    observation_to_transition,
-    robot_action_observation_to_transition,
-    transition_to_observation,
-    transition_to_robot_action,
-)
-from lerobot.robots import Robot
-from lerobot.teleoperators import Teleoperator
-from lerobot.utils.robot_utils import precise_sleep
-
-logger = logging.getLogger(__name__)
-
-
-@dataclass
-class HILDatasetConfig:
-    repo_id: str
-    single_task: str
-    root: str | Path | None = None
-    fps: int = 30
-    episode_time_s: float = 120
-    num_episodes: int = 50
-    video: bool = True
-    push_to_hub: bool = True
-    private: bool = False
-    tags: list[str] | None = None
-    num_image_writer_processes: int = 0
-    num_image_writer_threads_per_camera: int = 4
-    video_encoding_batch_size: int = 1
-    vcodec: str = "auto"
-    streaming_encoding: bool = True
-    encoder_queue_maxsize: int = 30
-    encoder_threads: int | None = None
-    rename_map: dict[str, str] = field(default_factory=dict)
-
-
-def teleop_has_motor_control(teleop: Teleoperator) -> bool:
-    """Check if teleoperator has motor control capabilities."""
-    return all(hasattr(teleop, attr) for attr in ("enable_torque", "disable_torque", "write_goal_positions"))
-
-
-def teleop_disable_torque(teleop: Teleoperator) -> None:
-    """Disable teleop torque if supported."""
-    if hasattr(teleop, "disable_torque"):
-        teleop.disable_torque()
-
-
-def teleop_enable_torque(teleop: Teleoperator) -> None:
-    """Enable teleop torque if supported."""
-    if hasattr(teleop, "enable_torque"):
-        teleop.enable_torque()
-
-
-def teleop_smooth_move_to(teleop: Teleoperator, target_pos: dict, duration_s: float = 2.0, fps: int = 50):
-    """Smoothly move teleop to target position if motor control is available."""
-    if not teleop_has_motor_control(teleop):
-        logger.warning("Teleop does not support motor control - cannot mirror robot position")
-        return
-
-    teleop_enable_torque(teleop)
-    current = teleop.get_action()
-    steps = max(int(duration_s * fps), 1)
-
-    for step in range(steps + 1):
-        t = step / steps
-        interp = {}
-        for k in current:
-            if k in target_pos:
-                interp[k] = current[k] * (1 - t) + target_pos[k] * t
-            else:
-                interp[k] = current[k]
-        teleop.write_goal_positions(interp)
-        time.sleep(1 / fps)
-
-
-def init_keyboard_listener():
-    """Initialize keyboard listener with HIL controls."""
-    events = {
-        "exit_early": False,
-        "rerecord_episode": False,
-        "stop_recording": False,
-        "policy_paused": False,
-        "correction_active": False,
-        "resume_policy": False,
-        "in_reset": False,
-        "start_next_episode": False,
-    }
-
-    if is_headless():
-        logger.warning("Headless environment - keyboard controls unavailable")
-        return None, events
-
-    from pynput import keyboard
-
-    def on_press(key):
-        try:
-            if events["in_reset"]:
-                if key in [keyboard.Key.space, keyboard.Key.right]:
-                    logger.info("[HIL] Starting next episode...")
-                    events["start_next_episode"] = True
-                elif hasattr(key, "char") and key.char == "c":
-                    events["start_next_episode"] = True
-                elif key == keyboard.Key.esc:
-                    logger.info("[HIL] ESC - Stop recording, pushing to hub...")
-                    events["stop_recording"] = True
-                    events["start_next_episode"] = True
-            else:
-                if key == keyboard.Key.space:
-                    if not events["policy_paused"] and not events["correction_active"]:
-                        logger.info("[HIL] PAUSED - Press 'c' to take control or 'p' to resume policy")
-                        events["policy_paused"] = True
-                elif hasattr(key, "char") and key.char == "c":
-                    if events["policy_paused"] and not events["correction_active"]:
-                        logger.info("[HIL] Taking control...")
-                        events["start_next_episode"] = True
-                elif hasattr(key, "char") and key.char == "p":
-                    if events["policy_paused"] or events["correction_active"]:
-                        logger.info("[HIL] Resuming policy...")
-                        events["resume_policy"] = True
-                elif key == keyboard.Key.right:
-                    logger.info("[HIL] End episode")
-                    events["exit_early"] = True
-                elif key == keyboard.Key.left:
-                    logger.info("[HIL] Re-record episode")
-                    events["rerecord_episode"] = True
-                    events["exit_early"] = True
-                elif key == keyboard.Key.esc:
-                    logger.info("[HIL] ESC - Stop recording...")
-                    events["stop_recording"] = True
-                    events["exit_early"] = True
-        except Exception as e:
-            logger.info(f"Key error: {e}")
-
-    listener = keyboard.Listener(on_press=on_press)
-    listener.start()
-    return listener, events
-
-
-def make_identity_processors():
-    """Create identity processors for recording."""
-    teleop_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-        steps=[IdentityProcessorStep()],
-        to_transition=robot_action_observation_to_transition,
-        to_output=transition_to_robot_action,
-    )
-    obs_proc = RobotProcessorPipeline[RobotObservation, RobotObservation](
-        steps=[IdentityProcessorStep()],
-        to_transition=observation_to_transition,
-        to_output=transition_to_observation,
-    )
-    return teleop_proc, obs_proc
-
-
-def reset_loop(robot: Robot, teleop: Teleoperator, events: dict, fps: int):
-    """Reset period where human repositions environment."""
-    logger.info("[HIL] RESET")
-
-    events["in_reset"] = True
-    events["start_next_episode"] = False
-
-    obs = robot.get_observation()
-    robot_pos = {k: v for k, v in obs.items() if k.endswith(".pos") and k in robot.observation_features}
-    teleop_smooth_move_to(teleop, robot_pos, duration_s=2.0, fps=50)
-
-    logger.info("Press any key to enable teleoperation")
-    while not events["start_next_episode"] and not events["stop_recording"]:
-        precise_sleep(0.05)
-
-    if events["stop_recording"]:
-        return
-
-    events["start_next_episode"] = False
-    teleop_disable_torque(teleop)
-    logger.info("Teleop enabled - press any key to start episode")
-
-    while not events["start_next_episode"] and not events["stop_recording"]:
-        loop_start = time.perf_counter()
-        action = teleop.get_action()
-        robot.send_action(action)
-        precise_sleep(1 / fps - (time.perf_counter() - loop_start))
-
-    events["in_reset"] = False
-    events["start_next_episode"] = False
-    events["exit_early"] = False
-    events["policy_paused"] = False
-    events["correction_active"] = False
-    events["resume_policy"] = False
-
-
-def print_controls(rtc: bool = False):
-    """Print control instructions."""
-    mode = "Human-in-the-Loop Data Collection" + (" (RTC)" if rtc else "")
-    logger.info(
-        "%s\n  Controls:\n"
-        "    SPACE  - Pause policy\n"
-        "    c      - Take control\n"
-        "    p      - Resume policy after pause/correction\n"
-        "    →      - End episode\n"
-        "    ESC    - Stop and push to hub",
-        mode,
-    )

examples/lekiwi/evaluate.py

@@ -14,17 +14,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.common.control_utils import init_keyboard_listener
+import logging
+import time
+
+from lerobot.common.control_utils import init_keyboard_listener, predict_action
 from lerobot.datasets import LeRobotDataset
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
+from lerobot.policies.utils import make_robot_action
 from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.constants import ACTION, OBS_STR
-from lerobot.utils.feature_utils import hw_to_dataset_features
+from lerobot.utils.feature_utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 NUM_EPISODES = 2
 FPS = 30
@@ -35,6 +39,9 @@ HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
 
 
 def main():
+    # NOTE: For production policy deployment, use `lerobot-rollout` CLI instead.
+    # This script provides a self-contained example for educational purposes.
+
     # Create the robot configuration & robot
     robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
 
@@ -83,43 +90,67 @@ def main():
             raise ValueError("Robot is not connected!")
 
         print("Starting evaluate loop...")
+        control_interval = 1 / FPS
         recorded_episodes = 0
         while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
             log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
 
-            # Main record loop
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                policy=policy,
-                preprocessor=preprocessor,  # Pass the pre and post policy processors
-                postprocessor=postprocessor,
-                dataset=dataset,
-                control_time_s=EPISODE_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=teleop_action_processor,
-                robot_action_processor=robot_action_processor,
-                robot_observation_processor=robot_observation_processor,
-            )
+            # Inline evaluation loop: predict actions and send to robot
+            timestamp = 0
+            start_episode_t = time.perf_counter()
+            while timestamp < EPISODE_TIME_SEC:
+                start_loop_t = time.perf_counter()
+
+                if events["exit_early"]:
+                    events["exit_early"] = False
+                    break
+
+                # Get robot observation
+                obs = robot.get_observation()
+                obs_processed = robot_observation_processor(obs)
+                observation_frame = build_dataset_frame(dataset.features, obs_processed, prefix=OBS_STR)
+
+                # Predict action using the policy
+                action_tensor = predict_action(
+                    observation=observation_frame,
+                    policy=policy,
+                    device=policy.config.device,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    use_amp=policy.config.device.type == "cuda",
+                    task=TASK_DESCRIPTION,
+                    robot_type=robot.name,
+                )
+
+                # Convert policy output to robot action dict
+                action_values = make_robot_action(action_tensor, dataset.features)
+
+                # Process and send action to robot
+                robot_action_to_send = robot_action_processor((action_values, obs))
+                robot.send_action(robot_action_to_send)
+
+                # Write to dataset
+                action_frame = build_dataset_frame(dataset.features, action_values, prefix=ACTION)
+                frame = {**observation_frame, **action_frame, "task": TASK_DESCRIPTION}
+                dataset.add_frame(frame)
+
+                log_rerun_data(observation=obs_processed, action=action_values)
+
+                dt_s = time.perf_counter() - start_loop_t
+                sleep_time_s = control_interval - dt_s
+                if sleep_time_s < 0:
+                    logging.warning(
+                        f"Evaluate loop is running slower ({1 / dt_s:.1f} Hz) than the target FPS ({FPS} Hz)."
+                    )
+                precise_sleep(max(sleep_time_s, 0.0))
+                timestamp = time.perf_counter() - start_episode_t
 
             # Reset the environment if not stopping or re-recording
             if not events["stop_recording"] and (
                 (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
             ):
                 log_say("Reset the environment")
-                record_loop(
-                    robot=robot,
-                    events=events,
-                    fps=FPS,
-                    control_time_s=EPISODE_TIME_SEC,
-                    single_task=TASK_DESCRIPTION,
-                    display_data=True,
-                    teleop_action_processor=teleop_action_processor,
-                    robot_action_processor=robot_action_processor,
-                    robot_observation_processor=robot_observation_processor,
-                )
+                log_say("Waiting for environment reset, press right arrow key when ready...")
 
             if events["rerecord_episode"]:
                 log_say("Re-record episode")

examples/lekiwi/record.py

@@ -45,9 +45,6 @@ def main():
     leader_arm = SO100Leader(leader_arm_config)
     keyboard = KeyboardTeleop(keyboard_config)
 
-    # TODO(Steven): Update this example to use pipelines
-    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
-
     # Configure the dataset features
     action_features = hw_to_dataset_features(robot.action_features, ACTION)
     obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
@@ -77,6 +74,10 @@ def main():
         if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
             raise ValueError("Robot or teleop is not connected!")
 
+        teleop_action_processor, robot_action_processor, robot_observation_processor = (
+            make_default_processors()
+        )
+
         print("Starting record loop...")
         recorded_episodes = 0
         while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
@@ -87,14 +88,14 @@ def main():
                 robot=robot,
                 events=events,
                 fps=FPS,
+                teleop_action_processor=teleop_action_processor,
+                robot_action_processor=robot_action_processor,
+                robot_observation_processor=robot_observation_processor,
                 dataset=dataset,
                 teleop=[leader_arm, keyboard],
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
-                teleop_action_processor=teleop_action_processor,
-                robot_action_processor=robot_action_processor,
-                robot_observation_processor=robot_observation_processor,
             )
 
             # Reset the environment if not stopping or re-recording
@@ -106,13 +107,13 @@ def main():
                     robot=robot,
                     events=events,
                     fps=FPS,
+                    teleop_action_processor=teleop_action_processor,
+                    robot_action_processor=robot_action_processor,
+                    robot_observation_processor=robot_observation_processor,
                     teleop=[leader_arm, keyboard],
                     control_time_s=RESET_TIME_SEC,
                     single_task=TASK_DESCRIPTION,
                     display_data=True,
-                    teleop_action_processor=teleop_action_processor,
-                    robot_action_processor=robot_action_processor,
-                    robot_observation_processor=robot_observation_processor,
                 )
 
             if events["rerecord_episode"]:

examples/lekiwi/rollout.py

@@ -0,0 +1,77 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Run a trained policy on LeKiwi without recording (base rollout).
+
+Uses the rollout engine's :class:`BaseStrategy` (autonomous execution,
+no dataset) with :class:`SyncInferenceConfig` (inline policy call per
+control tick).  For a CLI entry point with the same capabilities plus
+recording, upload, and human-in-the-loop variants, see ``lerobot-rollout``.
+"""
+
+from lerobot.configs import PreTrainedConfig
+from lerobot.robots.lekiwi import LeKiwiClientConfig
+from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
+from lerobot.rollout.inference import SyncInferenceConfig
+from lerobot.rollout.strategies import BaseStrategy
+from lerobot.utils.process import ProcessSignalHandler
+from lerobot.utils.utils import init_logging
+
+FPS = 30
+DURATION_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+
+
+def main():
+    init_logging()
+
+    # Robot: LeKiwi client — make sure lekiwi_host is already running on the robot.
+    robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
+
+    # Policy: load the pretrained config.  ``pretrained_path`` is read downstream
+    # by ``build_rollout_context`` to reload the full model.
+    policy_config = PreTrainedConfig.from_pretrained(HF_MODEL_ID)
+    policy_config.pretrained_path = HF_MODEL_ID
+
+    # Assemble the rollout config: base strategy (no recording) + sync inference.
+    cfg = RolloutConfig(
+        robot=robot_config,
+        policy=policy_config,
+        strategy=BaseStrategyConfig(),
+        inference=SyncInferenceConfig(),
+        fps=FPS,
+        duration=DURATION_SEC,
+        task=TASK_DESCRIPTION,
+    )
+
+    # Graceful Ctrl-C: the strategy loop exits when shutdown_event is set.
+    signal_handler = ProcessSignalHandler(use_threads=True)
+
+    # Build the context (connects robot, loads policy, wires the inference strategy).
+    # No custom processors here — LeKiwi runs on raw joint features.
+    ctx = build_rollout_context(cfg, signal_handler.shutdown_event)
+
+    strategy = BaseStrategy(cfg.strategy)
+    try:
+        strategy.setup(ctx)
+        strategy.run(ctx)
+    finally:
+        strategy.teardown(ctx)
+
+
+if __name__ == "__main__":
+    main()

examples/phone_to_so100/evaluate.py

@@ -14,13 +14,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import logging
+import time
+
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.common.control_utils import init_keyboard_listener, predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
+from lerobot.policies.utils import make_robot_action
 from lerobot.processor import (
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
@@ -34,11 +38,12 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.feature_utils import combine_feature_dicts
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 NUM_EPISODES = 5
 FPS = 30
@@ -49,6 +54,9 @@ HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
 
 
 def main():
+    # NOTE: For production policy deployment, use `lerobot-rollout` CLI instead.
+    # This script provides a self-contained example for educational purposes.
+
     # Create the robot configuration & robot
     camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
     robot_config = SO100FollowerConfig(
@@ -143,43 +151,67 @@ def main():
             raise ValueError("Robot is not connected!")
 
         print("Starting evaluate loop...")
+        control_interval = 1 / FPS
         episode_idx = 0
         for episode_idx in range(NUM_EPISODES):
             log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
-            # Main record loop
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                policy=policy,
-                preprocessor=preprocessor,  # Pass the pre and post policy processors
-                postprocessor=postprocessor,
-                dataset=dataset,
-                control_time_s=EPISODE_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=make_default_teleop_action_processor(),
-                robot_action_processor=robot_ee_to_joints_processor,
-                robot_observation_processor=robot_joints_to_ee_pose_processor,
-            )
+            # Inline evaluation loop: predict actions and send to robot
+            timestamp = 0
+            start_episode_t = time.perf_counter()
+            while timestamp < EPISODE_TIME_SEC:
+                start_loop_t = time.perf_counter()
+
+                if events["exit_early"]:
+                    events["exit_early"] = False
+                    break
+
+                # Get robot observation
+                obs = robot.get_observation()
+                obs_processed = robot_joints_to_ee_pose_processor(obs)
+                observation_frame = build_dataset_frame(dataset.features, obs_processed, prefix=OBS_STR)
+
+                # Predict action using the policy
+                action_tensor = predict_action(
+                    observation=observation_frame,
+                    policy=policy,
+                    device=policy.config.device,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    use_amp=policy.config.device.type == "cuda",
+                    task=TASK_DESCRIPTION,
+                    robot_type=robot.name,
+                )
+
+                # Convert policy output to robot action dict
+                action_values = make_robot_action(action_tensor, dataset.features)
+
+                # Process and send action to robot (EE -> joints via IK)
+                robot_action_to_send = robot_ee_to_joints_processor((action_values, obs))
+                robot.send_action(robot_action_to_send)
+
+                # Write to dataset
+                action_frame = build_dataset_frame(dataset.features, action_values, prefix=ACTION)
+                frame = {**observation_frame, **action_frame, "task": TASK_DESCRIPTION}
+                dataset.add_frame(frame)
+
+                log_rerun_data(observation=obs_processed, action=action_values)
+
+                dt_s = time.perf_counter() - start_loop_t
+                sleep_time_s = control_interval - dt_s
+                if sleep_time_s < 0:
+                    logging.warning(
+                        f"Evaluate loop is running slower ({1 / dt_s:.1f} Hz) than the target FPS ({FPS} Hz)."
+                    )
+                precise_sleep(max(sleep_time_s, 0.0))
+                timestamp = time.perf_counter() - start_episode_t
 
             # Reset the environment if not stopping or re-recording
             if not events["stop_recording"] and (
                 (episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]
             ):
                 log_say("Reset the environment")
-                record_loop(
-                    robot=robot,
-                    events=events,
-                    fps=FPS,
-                    control_time_s=EPISODE_TIME_SEC,
-                    single_task=TASK_DESCRIPTION,
-                    display_data=True,
-                    teleop_action_processor=make_default_teleop_action_processor(),
-                    robot_action_processor=robot_ee_to_joints_processor,
-                    robot_observation_processor=robot_joints_to_ee_pose_processor,
-                )
+                log_say("Waiting for environment reset, press right arrow key when ready...")
 
             if events["rerecord_episode"]:
                 log_say("Re-record episode")
@@ -190,7 +222,6 @@ def main():
 
             # Save episode
             dataset.save_episode()
-            episode_idx += 1
     finally:
         # Clean up
         log_say("Stop recording")

examples/phone_to_so100/record.py

@@ -65,14 +65,15 @@ def main():
     robot = SO100Follower(robot_config)
     phone = Phone(teleop_config)
 
-    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo:
+    #   https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
     kinematics_solver = RobotKinematics(
         urdf_path="./SO101/so101_new_calib.urdf",
         target_frame_name="gripper_frame_link",
         joint_names=list(robot.bus.motors.keys()),
     )
 
-    # Build pipeline to convert phone action to EE action
+    # Build pipeline to convert phone action to EE action (with gripper velocity mapped to joint).
     phone_to_robot_ee_pose_processor = RobotProcessorPipeline[
         tuple[RobotAction, RobotObservation], RobotAction
     ](
@@ -94,7 +95,7 @@ def main():
         to_output=transition_to_robot_action,
     )
 
-    # Build pipeline to convert EE action to joints action
+    # Build pipeline to convert EE action to joints action (IK).
     robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
         steps=[
             InverseKinematicsEEToJoints(
@@ -107,7 +108,7 @@ def main():
         to_output=transition_to_robot_action,
     )
 
-    # Build pipeline to convert joint observation to EE observation
+    # Build pipeline to convert joint observation to EE observation (FK).
     robot_joints_to_ee_pose = RobotProcessorPipeline[RobotObservation, RobotObservation](
         steps=[
             ForwardKinematicsJointsToEE(
@@ -118,13 +119,12 @@ def main():
         to_output=transition_to_observation,
     )
 
-    # Create the dataset
+    # Create the dataset, deriving features from the pipelines so the on-disk schema
+    # matches exactly what the pipelines produce at runtime.
     dataset = LeRobotDataset.create(
         repo_id=HF_REPO_ID,
         fps=FPS,
         features=combine_feature_dicts(
-            # Run the feature contract of the pipelines
-            # This tells you how the features would look like after the pipeline steps
             aggregate_pipeline_dataset_features(
                 pipeline=phone_to_robot_ee_pose_processor,
                 initial_features=create_initial_features(action=phone.action_features),
@@ -163,14 +163,14 @@ def main():
                 robot=robot,
                 events=events,
                 fps=FPS,
+                teleop_action_processor=phone_to_robot_ee_pose_processor,
+                robot_action_processor=robot_ee_to_joints_processor,
+                robot_observation_processor=robot_joints_to_ee_pose,
                 teleop=phone,
                 dataset=dataset,
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
-                teleop_action_processor=phone_to_robot_ee_pose_processor,
-                robot_action_processor=robot_ee_to_joints_processor,
-                robot_observation_processor=robot_joints_to_ee_pose,
             )
 
             # Reset the environment if not stopping or re-recording
@@ -182,13 +182,13 @@ def main():
                     robot=robot,
                     events=events,
                     fps=FPS,
+                    teleop_action_processor=phone_to_robot_ee_pose_processor,
+                    robot_action_processor=robot_ee_to_joints_processor,
+                    robot_observation_processor=robot_joints_to_ee_pose,
                     teleop=phone,
                     control_time_s=RESET_TIME_SEC,
                     single_task=TASK_DESCRIPTION,
                     display_data=True,
-                    teleop_action_processor=phone_to_robot_ee_pose_processor,
-                    robot_action_processor=robot_ee_to_joints_processor,
-                    robot_observation_processor=robot_joints_to_ee_pose,
                 )
 
             if events["rerecord_episode"]:

examples/phone_to_so100/rollout.py

@@ -0,0 +1,126 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Run a trained EE-space policy on SO100 (phone-trained) without recording.
+
+Mirrors ``examples/so100_to_so100_EE/rollout.py`` — the model was trained
+with phone teleoperation in EE space, so at deployment we only need the
+joint↔EE conversion on the robot side; the phone is not used.
+
+Uses :class:`BaseStrategy` (no recording) + :class:`SyncInferenceConfig`
+(inline policy call).  For recording during rollout, switch to Sentry,
+Highlight, or DAgger via ``lerobot-rollout --strategy.type=...``.
+"""
+
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.configs import PreTrainedConfig
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import (
+    RobotProcessorPipeline,
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
+    ForwardKinematicsJointsToEE,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
+from lerobot.rollout.inference import SyncInferenceConfig
+from lerobot.rollout.strategies import BaseStrategy
+from lerobot.types import RobotAction, RobotObservation
+from lerobot.utils.process import ProcessSignalHandler
+from lerobot.utils.utils import init_logging
+
+FPS = 30
+DURATION_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+
+
+def main():
+    init_logging()
+
+    camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem58760434471",
+        id="my_awesome_follower_arm",
+        cameras=camera_config,
+        use_degrees=True,
+    )
+
+    # Peek at motor names once to build the kinematic solver.
+    temp_robot = SO100Follower(robot_config)
+    motor_names = list(temp_robot.bus.motors.keys())
+
+    kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=motor_names,
+    )
+
+    robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=motor_names)],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+
+    robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[
+            InverseKinematicsEEToJoints(
+                kinematics=kinematics_solver,
+                motor_names=motor_names,
+                initial_guess_current_joints=True,
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+
+    policy_config = PreTrainedConfig.from_pretrained(HF_MODEL_ID)
+    policy_config.pretrained_path = HF_MODEL_ID
+
+    cfg = RolloutConfig(
+        robot=robot_config,
+        policy=policy_config,
+        strategy=BaseStrategyConfig(),
+        inference=SyncInferenceConfig(),
+        fps=FPS,
+        duration=DURATION_SEC,
+        task=TASK_DESCRIPTION,
+    )
+
+    signal_handler = ProcessSignalHandler(use_threads=True)
+
+    ctx = build_rollout_context(
+        cfg,
+        signal_handler.shutdown_event,
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose_processor,
+    )
+
+    strategy = BaseStrategy(cfg.strategy)
+    try:
+        strategy.setup(ctx)
+        strategy.run(ctx)
+    finally:
+        strategy.teardown(ctx)
+
+
+if __name__ == "__main__":
+    main()

examples/rtc/eval_with_real_robot.py

@@ -1,673 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-Demo script showing how to use Real-Time Chunking (RTC) with action chunking policies on real robots.
-
-This script demonstrates:
-1. Creating a robot and policy (SmolVLA, Pi0, etc.) with RTC
-2. Consuming actions from the policy while the robot executes
-3. Periodically requesting new action chunks in the background using threads
-4. Managing action buffers and timing for real-time operation
-
-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=<USER>/smolvla_check_rtc_last3 \
-        --policy.device=mps \
-        --rtc.enabled=true \
-        --rtc.execution_horizon=20 \
-        --robot.type=so100_follower \
-        --robot.port=/dev/tty.usbmodem58FA0834591 \
-        --robot.id=so100_follower \
-        --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-        --task="Move green small object into the purple platform" \
-        --duration=120
-
-    # Run RTC with Real robot without RTC
-    uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --policy.device=mps \
-        --rtc.enabled=false \
-        --robot.type=so100_follower \
-        --robot.port=/dev/tty.usbmodem58FA0834591 \
-        --robot.id=so100_follower \
-        --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-        --task="Move green small object into the purple platform" \
-        --duration=120
-
-    # Run RTC with Real robot with pi0.5 policy
-    uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/pi05_check_rtc \
-        --policy.device=mps \
-        --rtc.enabled=true \
-        --rtc.execution_horizon=20 \
-        --robot.type=so100_follower \
-        --robot.port=/dev/tty.usbmodem58FA0834591 \
-        --robot.id=so100_follower \
-        --robot.cameras="{ gripper: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}}" \
-        --task="Move green small object into the purple platform" \
-        --duration=120
-
-    # Run RTC with bi_openarm_follower (dual-arm OpenArms) and pi0.5 policy
-    python examples/rtc/eval_with_real_robot.py \
-        --policy.path=lerobot-data-collection/folding_final \
-        --robot.type=bi_openarm_follower \
-        --robot.cameras='{left_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}}' \
-        --robot.left_arm_config.port=can0 \
-        --robot.left_arm_config.side=left \
-        --robot.left_arm_config.can_interface=socketcan \
-        --robot.left_arm_config.disable_torque_on_disconnect=true \
-        --robot.left_arm_config.max_relative_target=8.0 \
-        --robot.right_arm_config.port=can1 \
-        --robot.right_arm_config.side=right \
-        --robot.right_arm_config.can_interface=socketcan \
-        --robot.right_arm_config.disable_torque_on_disconnect=true \
-        --robot.right_arm_config.max_relative_target=8.0 \
-        --task="Fold the T-shirt properly" \
-        --fps=30 \
-        --duration=2000 \
-        --interpolation_multiplier=3 \
-        --rtc.enabled=true \
-        --rtc.execution_horizon=20 \
-        --rtc.max_guidance_weight=5.0 \
-        --rtc.prefix_attention_schedule=LINEAR \
-        --device=cuda
-"""
-
-import logging
-import math
-import sys
-import time
-import traceback
-from dataclasses import dataclass, field
-from threading import Event, Lock, Thread
-
-import torch
-from torch import Tensor
-
-from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
-from lerobot.cameras.realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.cameras.zmq import ZMQCameraConfig  # noqa: F401
-from lerobot.configs import PreTrainedConfig, RTCAttentionSchedule, parser
-from lerobot.policies import get_policy_class, make_pre_post_processors
-from lerobot.policies.rtc import ActionInterpolator, ActionQueue, LatencyTracker, RTCConfig
-from lerobot.processor import (
-    NormalizerProcessorStep,
-    RelativeActionsProcessorStep,
-    TransitionKey,
-    create_transition,
-    make_default_robot_action_processor,
-    make_default_robot_observation_processor,
-    to_relative_actions,
-)
-from lerobot.rl.process import ProcessSignalHandler
-from lerobot.robots import (  # noqa: F401
-    Robot,
-    RobotConfig,
-    bi_openarm_follower,
-    bi_so_follower,
-    koch_follower,
-    so_follower,
-    unitree_g1,
-)
-from lerobot.robots.utils import make_robot_from_config
-from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
-from lerobot.utils.feature_utils import build_dataset_frame, hw_to_dataset_features
-from lerobot.utils.hub import HubMixin
-from lerobot.utils.utils import init_logging
-
-logging.basicConfig(level=logging.INFO)
-logger = logging.getLogger(__name__)
-
-
-class RobotWrapper:
-    def __init__(self, robot: Robot):
-        self.robot = robot
-        self.lock = Lock()
-
-    def get_observation(self) -> dict[str, Tensor]:
-        with self.lock:
-            return self.robot.get_observation()
-
-    def send_action(self, action: Tensor):
-        with self.lock:
-            self.robot.send_action(action)
-
-    def observation_features(self) -> list[str]:
-        with self.lock:
-            return self.robot.observation_features
-
-    def action_features(self) -> list[str]:
-        with self.lock:
-            return self.robot.action_features
-
-
-@dataclass
-class RTCDemoConfig(HubMixin):
-    """Configuration for RTC demo with action chunking policies and real robots."""
-
-    # Policy configuration
-    policy: PreTrainedConfig | None = None
-
-    # Robot configuration
-    robot: RobotConfig | None = None
-
-    # RTC configuration
-    rtc: RTCConfig = field(
-        default_factory=lambda: RTCConfig(
-            execution_horizon=10,
-            max_guidance_weight=1.0,
-            prefix_attention_schedule=RTCAttentionSchedule.EXP,
-        )
-    )
-
-    # Demo parameters
-    duration: float = 30.0  # Duration to run the demo (seconds)
-    fps: float = 10.0  # Action execution frequency (Hz)
-    interpolation_multiplier: int = 1  # Control rate multiplier (1=off, 2=2x, 3=3x)
-
-    # Compute device
-    device: str | None = None  # Device to run on (cuda, cpu, auto)
-
-    # Get new actions horizon. The amount of executed steps after which will be requested new actions.
-    # It should be higher than inference delay + execution horizon.
-    action_queue_size_to_get_new_actions: int = 30
-
-    # Task to execute
-    task: str = field(default="", metadata={"help": "Task to execute"})
-
-    # Torch compile configuration
-    use_torch_compile: bool = field(
-        default=False,
-        metadata={"help": "Use torch.compile for faster inference (PyTorch 2.0+)"},
-    )
-
-    torch_compile_backend: str = field(
-        default="inductor",
-        metadata={"help": "Backend for torch.compile (inductor, aot_eager, cudagraphs)"},
-    )
-
-    torch_compile_mode: str = field(
-        default="default",
-        metadata={"help": "Compilation mode (default, reduce-overhead, max-autotune)"},
-    )
-
-    torch_compile_disable_cudagraphs: bool = field(
-        default=True,
-        metadata={
-            "help": "Disable CUDA graphs in torch.compile. Required due to in-place tensor "
-            "operations in denoising loop (x_t += dt * v_t) which cause tensor aliasing issues."
-        },
-    )
-
-    def __post_init__(self):
-        # HACK: We parse again the cli args here to get the pretrained path if there was one.
-        policy_path = parser.get_path_arg("policy")
-        if policy_path:
-            cli_overrides = parser.get_cli_overrides("policy")
-            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
-            self.policy.pretrained_path = policy_path
-        else:
-            raise ValueError("Policy path is required")
-
-        # Validate that robot configuration is provided
-        if self.robot is None:
-            raise ValueError("Robot configuration must be provided")
-
-    @classmethod
-    def __get_path_fields__(cls) -> list[str]:
-        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
-        return ["policy"]
-
-
-def is_image_key(k: str) -> bool:
-    return k.startswith(OBS_IMAGES)
-
-
-def _reanchor_relative_rtc_prefix(
-    prev_actions_absolute: Tensor,
-    current_state: Tensor,
-    relative_step: RelativeActionsProcessorStep,
-    normalizer_step: NormalizerProcessorStep | None,
-    policy_device: torch.device | str,
-) -> Tensor:
-    """Convert absolute leftovers into model-space for relative-action RTC policies.
-
-    When a policy uses relative actions, the RTC prefix (leftover actions from
-    the previous chunk) is stored in absolute space. Before feeding it back to
-    the policy we need to re-express it relative to the *current* robot state
-    and then re-normalize.
-    """
-    state = current_state.detach().cpu()
-    if state.dim() == 1:
-        state = state.unsqueeze(0)
-
-    action_cpu = prev_actions_absolute.detach().cpu()
-    mask = relative_step._build_mask(action_cpu.shape[-1])
-    relative_actions = to_relative_actions(action_cpu, state, mask)
-
-    transition = create_transition(action=relative_actions)
-    if normalizer_step is not None:
-        transition = normalizer_step(transition)
-
-    return transition[TransitionKey.ACTION].to(policy_device)
-
-
-def get_actions(
-    policy,
-    robot: RobotWrapper,
-    robot_observation_processor,
-    action_queue: ActionQueue,
-    shutdown_event: Event,
-    cfg: RTCDemoConfig,
-):
-    """Thread function to request action chunks from the policy.
-
-    Args:
-        policy: The policy instance (SmolVLA, Pi0, etc.)
-        robot: The robot instance for getting observations
-        robot_observation_processor: Processor for raw robot observations
-        action_queue: Queue to put new action chunks
-        shutdown_event: Event to signal shutdown
-        cfg: Demo configuration
-    """
-    try:
-        logger.info("[GET_ACTIONS] Starting get actions thread")
-
-        latency_tracker = LatencyTracker()  # Track latency of action chunks
-        fps = cfg.fps
-        time_per_chunk = 1.0 / fps
-
-        # Only keep .pos joints + camera streams if the policy was trained on positions,
-        # not the full pos/vel/torque state the robot exposes.
-        observation_features_hw = {
-            key: value
-            for key, value in robot.observation_features().items()
-            if key.endswith(".pos") or isinstance(value, tuple)
-        }
-
-        dataset_features = hw_to_dataset_features(observation_features_hw, "observation")
-        policy_device = policy.config.device
-
-        # Load preprocessor and postprocessor from pretrained files
-        # The stats are embedded in the processor .safetensors files
-        logger.info(f"[GET_ACTIONS] Loading preprocessor/postprocessor from {cfg.policy.pretrained_path}")
-
-        preprocessor, postprocessor = make_pre_post_processors(
-            policy_cfg=cfg.policy,
-            pretrained_path=cfg.policy.pretrained_path,
-            dataset_stats=None,  # Will load from pretrained processor files
-            preprocessor_overrides={
-                "device_processor": {"device": cfg.policy.device},
-            },
-        )
-
-        logger.info("[GET_ACTIONS] Preprocessor/postprocessor loaded successfully with embedded stats")
-
-        relative_step = next(
-            (s for s in preprocessor.steps if isinstance(s, RelativeActionsProcessorStep) and s.enabled),
-            None,
-        )
-        normalizer_step = next(
-            (s for s in preprocessor.steps if isinstance(s, NormalizerProcessorStep)),
-            None,
-        )
-        if relative_step is not None:
-            if relative_step.action_names is None:
-                cfg_names = getattr(cfg.policy, "action_feature_names", None)
-                if cfg_names:
-                    relative_step.action_names = list(cfg_names)
-                else:
-                    relative_step.action_names = [
-                        k for k in robot.robot.action_features if k.endswith(".pos")
-                    ]
-            logger.info("[GET_ACTIONS] Relative actions enabled: will re-anchor RTC prefix")
-
-        get_actions_threshold = cfg.action_queue_size_to_get_new_actions
-
-        if not cfg.rtc.enabled:
-            get_actions_threshold = 0
-
-        while not shutdown_event.is_set():
-            if action_queue.qsize() <= get_actions_threshold:
-                current_time = time.perf_counter()
-                action_index_before_inference = action_queue.get_action_index()
-                prev_actions = action_queue.get_left_over()
-
-                inference_latency = latency_tracker.max()
-                inference_delay = math.ceil(inference_latency / time_per_chunk)
-
-                obs = robot.get_observation()
-
-                # Apply robot observation processor
-                obs_processed = robot_observation_processor(obs)
-
-                obs_with_policy_features = build_dataset_frame(
-                    dataset_features, obs_processed, prefix="observation"
-                )
-
-                for name in obs_with_policy_features:
-                    obs_with_policy_features[name] = torch.from_numpy(obs_with_policy_features[name])
-                    if "image" in name:
-                        obs_with_policy_features[name] = (
-                            obs_with_policy_features[name].type(torch.float32) / 255
-                        )
-                        obs_with_policy_features[name] = (
-                            obs_with_policy_features[name].permute(2, 0, 1).contiguous()
-                        )
-                    obs_with_policy_features[name] = obs_with_policy_features[name].unsqueeze(0)
-                    obs_with_policy_features[name] = obs_with_policy_features[name].to(policy_device)
-
-                obs_with_policy_features["task"] = [cfg.task]  # Task should be a list, not a string!
-                obs_with_policy_features["robot_type"] = (
-                    robot.robot.name if hasattr(robot.robot, "name") else ""
-                )
-
-                preproceseded_obs = preprocessor(obs_with_policy_features)
-
-                # Re-anchor leftover actions for relative-action policies.
-                # We need the *postprocessed* (absolute) leftover, not the original
-                # (normalized/relative) one that get_left_over() returns.
-                if (
-                    prev_actions is not None
-                    and relative_step is not None
-                    and OBS_STATE in obs_with_policy_features
-                ):
-                    with action_queue.lock:
-                        if action_queue.queue is not None:
-                            prev_actions_abs = action_queue.queue[action_queue.last_index :].clone()
-                        else:
-                            prev_actions_abs = None
-                    if prev_actions_abs is not None and prev_actions_abs.numel() > 0:
-                        prev_actions = _reanchor_relative_rtc_prefix(
-                            prev_actions_absolute=prev_actions_abs,
-                            current_state=obs_with_policy_features[OBS_STATE],
-                            relative_step=relative_step,
-                            normalizer_step=normalizer_step,
-                            policy_device=policy_device,
-                        )
-
-                # Generate actions WITH RTC
-                actions = policy.predict_action_chunk(
-                    preproceseded_obs,
-                    inference_delay=inference_delay,
-                    prev_chunk_left_over=prev_actions,
-                )
-
-                # Store original actions (before postprocessing) for RTC
-                original_actions = actions.squeeze(0).clone()
-
-                postprocessed_actions = postprocessor(actions)
-
-                postprocessed_actions = postprocessed_actions.squeeze(0)
-
-                new_latency = time.perf_counter() - current_time
-                new_delay = math.ceil(new_latency / time_per_chunk)
-                latency_tracker.add(new_latency)
-
-                if cfg.action_queue_size_to_get_new_actions < cfg.rtc.execution_horizon + new_delay:
-                    logger.warning(
-                        "[GET_ACTIONS] cfg.action_queue_size_to_get_new_actions Too small, It should be higher than inference delay + execution horizon."
-                    )
-
-                action_queue.merge(
-                    original_actions, postprocessed_actions, new_delay, action_index_before_inference
-                )
-            else:
-                # Small sleep to prevent busy waiting
-                time.sleep(0.1)
-
-        logger.info("[GET_ACTIONS] get actions thread shutting down")
-    except Exception as e:
-        logger.error(f"[GET_ACTIONS] Fatal exception in get_actions thread: {e}")
-        logger.error(traceback.format_exc())
-        sys.exit(1)
-
-
-def actor_control(
-    robot: RobotWrapper,
-    robot_action_processor,
-    action_queue: ActionQueue,
-    shutdown_event: Event,
-    cfg: RTCDemoConfig,
-):
-    """Thread function to execute actions on the robot.
-
-    Args:
-        robot: The robot instance
-        action_queue: Queue to get actions from
-        shutdown_event: Event to signal shutdown
-        cfg: Demo configuration
-    """
-    try:
-        logger.info("[ACTOR] Starting actor thread")
-
-        action_keys = [k for k in robot.action_features() if k.endswith(".pos")]
-
-        action_count = 0
-        interpolator = ActionInterpolator(multiplier=cfg.interpolation_multiplier)
-        action_interval = interpolator.get_control_interval(cfg.fps)
-
-        while not shutdown_event.is_set():
-            start_time = time.perf_counter()
-
-            if interpolator.needs_new_action():
-                new_action = action_queue.get()
-                if new_action is not None:
-                    interpolator.add(new_action.cpu())
-
-            action = interpolator.get()
-            if action is not None:
-                action = action.cpu()
-                action_dict = {key: action[i].item() for i, key in enumerate(action_keys)}
-                action_processed = robot_action_processor((action_dict, None))
-                robot.send_action(action_processed)
-                action_count += 1
-
-            dt_s = time.perf_counter() - start_time
-            time.sleep(max(0, (action_interval - dt_s) - 0.001))
-
-        logger.info(f"[ACTOR] Actor thread shutting down. Total actions executed: {action_count}")
-    except Exception as e:
-        logger.error(f"[ACTOR] Fatal exception in actor_control thread: {e}")
-        logger.error(traceback.format_exc())
-        sys.exit(1)
-
-
-def _apply_torch_compile(policy, cfg: RTCDemoConfig):
-    """Apply torch.compile to the policy's predict_action_chunk method.
-
-    Args:
-        policy: Policy instance to compile
-        cfg: Configuration containing torch compile settings
-
-    Returns:
-        Policy with compiled predict_action_chunk method
-    """
-
-    # PI models handle their own compilation
-    if policy.type == "pi05" or policy.type == "pi0":
-        return policy
-
-    try:
-        # Check if torch.compile is available (PyTorch 2.0+)
-        if not hasattr(torch, "compile"):
-            logger.warning(
-                f"torch.compile is not available. Requires PyTorch 2.0+. "
-                f"Current version: {torch.__version__}. Skipping compilation."
-            )
-            return policy
-
-        logger.info("Applying torch.compile to predict_action_chunk...")
-        logger.info(f"  Backend: {cfg.torch_compile_backend}")
-        logger.info(f"  Mode: {cfg.torch_compile_mode}")
-        logger.info(f"  Disable CUDA graphs: {cfg.torch_compile_disable_cudagraphs}")
-
-        # Compile the predict_action_chunk method
-        # - CUDA graphs disabled to prevent tensor aliasing from in-place ops (x_t += dt * v_t)
-        compile_kwargs = {
-            "backend": cfg.torch_compile_backend,
-            "mode": cfg.torch_compile_mode,
-        }
-
-        # Disable CUDA graphs if requested (prevents tensor aliasing issues)
-        if cfg.torch_compile_disable_cudagraphs:
-            compile_kwargs["options"] = {"triton.cudagraphs": False}
-
-        original_method = policy.predict_action_chunk
-        compiled_method = torch.compile(original_method, **compile_kwargs)
-        policy.predict_action_chunk = compiled_method
-        logger.info("✓ Successfully compiled predict_action_chunk")
-
-    except Exception as e:
-        logger.error(f"Failed to apply torch.compile: {e}")
-        logger.warning("Continuing without torch.compile")
-
-    return policy
-
-
-@parser.wrap()
-def demo_cli(cfg: RTCDemoConfig):
-    """Main entry point for RTC demo with draccus configuration."""
-
-    # Initialize logging
-    init_logging()
-
-    logger.info(f"Using device: {cfg.device}")
-
-    # Setup signal handler for graceful shutdown
-    signal_handler = ProcessSignalHandler(use_threads=True, display_pid=False)
-    shutdown_event = signal_handler.shutdown_event
-
-    policy = None
-    robot = None
-    get_actions_thread = None
-    actor_thread = None
-
-    policy_class = get_policy_class(cfg.policy.type)
-
-    # Load config and set compile_model for pi0/pi05 models
-    config = PreTrainedConfig.from_pretrained(cfg.policy.pretrained_path)
-
-    if cfg.policy.type == "pi05" or cfg.policy.type == "pi0":
-        config.compile_model = cfg.use_torch_compile
-
-    if config.use_peft:
-        from peft import PeftConfig, PeftModel
-
-        peft_pretrained_path = cfg.policy.pretrained_path
-        peft_config = PeftConfig.from_pretrained(peft_pretrained_path)
-
-        policy = policy_class.from_pretrained(
-            pretrained_name_or_path=peft_config.base_model_name_or_path, config=config
-        )
-        policy = PeftModel.from_pretrained(policy, peft_pretrained_path, config=peft_config)
-    else:
-        policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
-
-    # Turn on RTC
-    policy.config.rtc_config = cfg.rtc
-
-    # Init RTC processort, as by default if RTC disabled in the config
-    # The processor won't be created
-    policy.init_rtc_processor()
-
-    assert policy.name in ["smolvla", "pi05", "pi0"], "Only smolvla, pi05, and pi0 are supported for RTC"
-
-    policy = policy.to(cfg.device)
-    policy.eval()
-
-    # Apply torch.compile to predict_action_chunk method if enabled
-    if cfg.use_torch_compile:
-        policy = _apply_torch_compile(policy, cfg)
-
-    # Create robot
-    logger.info(f"Initializing robot: {cfg.robot.type}")
-    robot = make_robot_from_config(cfg.robot)
-    robot.connect()
-    robot_wrapper = RobotWrapper(robot)
-
-    # Create robot observation processor
-    robot_observation_processor = make_default_robot_observation_processor()
-    robot_action_processor = make_default_robot_action_processor()
-
-    # Create action queue for communication between threads
-    action_queue = ActionQueue(cfg.rtc)
-
-    # Start chunk requester thread
-    get_actions_thread = Thread(
-        target=get_actions,
-        args=(policy, robot_wrapper, robot_observation_processor, action_queue, shutdown_event, cfg),
-        daemon=True,
-        name="GetActions",
-    )
-    get_actions_thread.start()
-    logger.info("Started get actions thread")
-
-    # Start action executor thread
-    actor_thread = Thread(
-        target=actor_control,
-        args=(robot_wrapper, robot_action_processor, action_queue, shutdown_event, cfg),
-        daemon=True,
-        name="Actor",
-    )
-    actor_thread.start()
-    logger.info("Started actor thread")
-
-    logger.info("Started stop by duration thread")
-
-    # Main thread monitors for duration or shutdown
-    logger.info(f"Running demo for {cfg.duration} seconds...")
-    start_time = time.time()
-
-    while not shutdown_event.is_set() and (time.time() - start_time) < cfg.duration:
-        time.sleep(10)
-
-        # Log queue status periodically
-        if int(time.time() - start_time) % 5 == 0:
-            logger.info(f"[MAIN] Action queue size: {action_queue.qsize()}")
-
-        if time.time() - start_time > cfg.duration:
-            break
-
-    logger.info("Demo duration reached or shutdown requested")
-
-    # Signal shutdown
-    shutdown_event.set()
-
-    # Wait for threads to finish
-    if get_actions_thread and get_actions_thread.is_alive():
-        logger.info("Waiting for chunk requester thread to finish...")
-        get_actions_thread.join()
-
-    if actor_thread and actor_thread.is_alive():
-        logger.info("Waiting for action executor thread to finish...")
-        actor_thread.join()
-
-    # Cleanup robot
-    if robot:
-        robot.disconnect()
-        logger.info("Robot disconnected")
-
-    logger.info("Cleanup completed")
-
-
-if __name__ == "__main__":
-    demo_cli()
-    logging.info("RTC demo finished")

examples/so100_to_so100_EE/evaluate.py

@@ -14,13 +14,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import logging
+import time
+
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.common.control_utils import init_keyboard_listener, predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
+from lerobot.policies.utils import make_robot_action
 from lerobot.processor import (
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
@@ -34,11 +38,12 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.feature_utils import combine_feature_dicts
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 NUM_EPISODES = 5
 FPS = 30
@@ -49,6 +54,9 @@ HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
 
 
 def main():
+    # NOTE: For production policy deployment, use `lerobot-rollout` CLI instead.
+    # This script provides a self-contained example for educational purposes.
+
     # Create the robot configuration & robot
     camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
     robot_config = SO100FollowerConfig(
@@ -143,43 +151,67 @@ def main():
             raise ValueError("Robot is not connected!")
 
         print("Starting evaluate loop...")
+        control_interval = 1 / FPS
         episode_idx = 0
         for episode_idx in range(NUM_EPISODES):
             log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
-            # Main record loop
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                policy=policy,
-                preprocessor=preprocessor,  # Pass the pre and post policy processors
-                postprocessor=postprocessor,
-                dataset=dataset,
-                control_time_s=EPISODE_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=make_default_teleop_action_processor(),
-                robot_action_processor=robot_ee_to_joints_processor,
-                robot_observation_processor=robot_joints_to_ee_pose_processor,
-            )
+            # Inline evaluation loop: predict actions and send to robot
+            timestamp = 0
+            start_episode_t = time.perf_counter()
+            while timestamp < EPISODE_TIME_SEC:
+                start_loop_t = time.perf_counter()
+
+                if events["exit_early"]:
+                    events["exit_early"] = False
+                    break
+
+                # Get robot observation
+                obs = robot.get_observation()
+                obs_processed = robot_joints_to_ee_pose_processor(obs)
+                observation_frame = build_dataset_frame(dataset.features, obs_processed, prefix=OBS_STR)
+
+                # Predict action using the policy
+                action_tensor = predict_action(
+                    observation=observation_frame,
+                    policy=policy,
+                    device=policy.config.device,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    use_amp=policy.config.device.type == "cuda",
+                    task=TASK_DESCRIPTION,
+                    robot_type=robot.name,
+                )
+
+                # Convert policy output to robot action dict
+                action_values = make_robot_action(action_tensor, dataset.features)
+
+                # Process and send action to robot (EE -> joints via IK)
+                robot_action_to_send = robot_ee_to_joints_processor((action_values, obs))
+                robot.send_action(robot_action_to_send)
+
+                # Write to dataset
+                action_frame = build_dataset_frame(dataset.features, action_values, prefix=ACTION)
+                frame = {**observation_frame, **action_frame, "task": TASK_DESCRIPTION}
+                dataset.add_frame(frame)
+
+                log_rerun_data(observation=obs_processed, action=action_values)
+
+                dt_s = time.perf_counter() - start_loop_t
+                sleep_time_s = control_interval - dt_s
+                if sleep_time_s < 0:
+                    logging.warning(
+                        f"Evaluate loop is running slower ({1 / dt_s:.1f} Hz) than the target FPS ({FPS} Hz)."
+                    )
+                precise_sleep(max(sleep_time_s, 0.0))
+                timestamp = time.perf_counter() - start_episode_t
 
             # Reset the environment if not stopping or re-recording
             if not events["stop_recording"] and (
                 (episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]
             ):
                 log_say("Reset the environment")
-                record_loop(
-                    robot=robot,
-                    events=events,
-                    fps=FPS,
-                    control_time_s=EPISODE_TIME_SEC,
-                    single_task=TASK_DESCRIPTION,
-                    display_data=True,
-                    teleop_action_processor=make_default_teleop_action_processor(),
-                    robot_action_processor=robot_ee_to_joints_processor,
-                    robot_observation_processor=robot_joints_to_ee_pose_processor,
-                )
+                log_say("Waiting for environment reset, press right arrow key when ready...")
 
             if events["rerecord_episode"]:
                 log_say("Re-record episode")
@@ -190,7 +222,6 @@ def main():
 
             # Save episode
             dataset.save_episode()
-            episode_idx += 1
     finally:
         # Clean up
         log_say("Stop recording")

examples/so100_to_so100_EE/record.py

@@ -62,21 +62,20 @@ def main():
     follower = SO100Follower(follower_config)
     leader = SO100Leader(leader_config)
 
-    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo:
+    #   https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
     follower_kinematics_solver = RobotKinematics(
         urdf_path="./SO101/so101_new_calib.urdf",
         target_frame_name="gripper_frame_link",
         joint_names=list(follower.bus.motors.keys()),
     )
-
-    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
     leader_kinematics_solver = RobotKinematics(
         urdf_path="./SO101/so101_new_calib.urdf",
         target_frame_name="gripper_frame_link",
         joint_names=list(leader.bus.motors.keys()),
     )
 
-    # Build pipeline to convert follower joints to EE observation
+    # Build pipeline to convert follower joints to EE observation.
     follower_joints_to_ee = RobotProcessorPipeline[RobotObservation, RobotObservation](
         steps=[
             ForwardKinematicsJointsToEE(
@@ -87,7 +86,7 @@ def main():
         to_output=transition_to_observation,
     )
 
-    # Build pipeline to convert leader joints to EE action
+    # Build pipeline to convert leader joints to EE action.
     leader_joints_to_ee = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
         steps=[
             ForwardKinematicsJointsToEE(
@@ -98,9 +97,9 @@ def main():
         to_output=transition_to_robot_action,
     )
 
-    # Build pipeline to convert EE action to follower joints
+    # Build pipeline to convert EE action to follower joints (with safety bounds).
     ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-        [
+        steps=[
             EEBoundsAndSafety(
                 end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
                 max_ee_step_m=0.10,
@@ -115,13 +114,12 @@ def main():
         to_output=transition_to_robot_action,
     )
 
-    # Create the dataset
+    # Create the dataset, deriving features from the pipelines so the on-disk schema
+    # matches exactly what the pipelines produce at runtime.
     dataset = LeRobotDataset.create(
         repo_id=HF_REPO_ID,
         fps=FPS,
         features=combine_feature_dicts(
-            # Run the feature contract of the pipelines
-            # This tells you how the features would look like after the pipeline steps
             aggregate_pipeline_dataset_features(
                 pipeline=leader_joints_to_ee,
                 initial_features=create_initial_features(action=leader.action_features),
@@ -144,7 +142,7 @@ def main():
 
     # Initialize the keyboard listener and rerun visualization
     listener, events = init_keyboard_listener()
-    init_rerun(session_name="recording_phone")
+    init_rerun(session_name="recording_so100_ee")
 
     try:
         if not leader.is_connected or not follower.is_connected:
@@ -160,14 +158,14 @@ def main():
                 robot=follower,
                 events=events,
                 fps=FPS,
+                teleop_action_processor=leader_joints_to_ee,
+                robot_action_processor=ee_to_follower_joints,
+                robot_observation_processor=follower_joints_to_ee,
                 teleop=leader,
                 dataset=dataset,
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
-                teleop_action_processor=leader_joints_to_ee,
-                robot_action_processor=ee_to_follower_joints,
-                robot_observation_processor=follower_joints_to_ee,
             )
 
             # Reset the environment if not stopping or re-recording
@@ -179,13 +177,13 @@ def main():
                     robot=follower,
                     events=events,
                     fps=FPS,
+                    teleop_action_processor=leader_joints_to_ee,
+                    robot_action_processor=ee_to_follower_joints,
+                    robot_observation_processor=follower_joints_to_ee,
                     teleop=leader,
                     control_time_s=RESET_TIME_SEC,
                     single_task=TASK_DESCRIPTION,
                     display_data=True,
-                    teleop_action_processor=leader_joints_to_ee,
-                    robot_action_processor=ee_to_follower_joints,
-                    robot_observation_processor=follower_joints_to_ee,
                 )
 
             if events["rerecord_episode"]:

examples/so100_to_so100_EE/rollout.py

@@ -0,0 +1,134 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Run a trained EE-space policy on SO100 without recording (base rollout).
+
+Uses the rollout engine's :class:`BaseStrategy` (autonomous execution,
+no dataset) with :class:`SyncInferenceConfig` (inline policy call per
+control tick).  The custom observation/action processors convert between
+joint space (robot hardware) and end-effector space (policy I/O) via
+forward/inverse kinematics.
+"""
+
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.configs import PreTrainedConfig
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import (
+    RobotProcessorPipeline,
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
+    ForwardKinematicsJointsToEE,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
+from lerobot.rollout.inference import SyncInferenceConfig
+from lerobot.rollout.strategies import BaseStrategy
+from lerobot.types import RobotAction, RobotObservation
+from lerobot.utils.process import ProcessSignalHandler
+from lerobot.utils.utils import init_logging
+
+FPS = 30
+DURATION_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+
+
+def main():
+    init_logging()
+
+    # Robot configuration — the rollout engine will connect it inside build_rollout_context.
+    camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411",
+        id="my_awesome_follower_arm",
+        cameras=camera_config,
+        use_degrees=True,
+    )
+
+    # Kinematic solver: we need the motor-name list, so peek at the robot once.
+    # (The rollout engine owns the connected instance; we only use this for introspection.)
+    temp_robot = SO100Follower(robot_config)
+    motor_names = list(temp_robot.bus.motors.keys())
+
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo:
+    #   https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=motor_names,
+    )
+
+    # Joint-space observation → EE-space observation (consumed by the policy).
+    robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=motor_names)],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+
+    # EE-space action (produced by the policy) → joint-space action (sent to robot).
+    robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[
+            InverseKinematicsEEToJoints(
+                kinematics=kinematics_solver,
+                motor_names=motor_names,
+                initial_guess_current_joints=True,
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+
+    # Policy config (full model is loaded inside build_rollout_context).
+    policy_config = PreTrainedConfig.from_pretrained(HF_MODEL_ID)
+    policy_config.pretrained_path = HF_MODEL_ID
+
+    cfg = RolloutConfig(
+        robot=robot_config,
+        policy=policy_config,
+        strategy=BaseStrategyConfig(),
+        inference=SyncInferenceConfig(),
+        fps=FPS,
+        duration=DURATION_SEC,
+        task=TASK_DESCRIPTION,
+    )
+
+    signal_handler = ProcessSignalHandler(use_threads=True)
+
+    # Pass the EE kinematic processors via kwargs; the defaults (identity) would
+    # otherwise skip the joint↔EE conversion and the policy would receive the
+    # wrong observation/action space.
+    ctx = build_rollout_context(
+        cfg,
+        signal_handler.shutdown_event,
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose_processor,
+    )
+
+    strategy = BaseStrategy(cfg.strategy)
+    try:
+        strategy.setup(ctx)
+        strategy.run(ctx)
+    finally:
+        strategy.teardown(ctx)
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/rl/hilserl_example.py

@@ -10,7 +10,7 @@ from lerobot.datasets import LeRobotDataset
 from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
 from lerobot.policies import SACConfig
 from lerobot.policies.sac.modeling_sac import SACPolicy
-from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
+from lerobot.rewards.classifier.modeling_classifier import Classifier
 from lerobot.rl.buffer import ReplayBuffer
 from lerobot.rl.gym_manipulator import make_robot_env
 from lerobot.robots.so_follower import SO100FollowerConfig

examples/tutorial/rl/reward_classifier_example.py

@@ -1,7 +1,7 @@
 import torch
 
 from lerobot.datasets import LeRobotDataset
-from lerobot.policies import RewardClassifierConfig, make_policy, make_pre_post_processors
+from lerobot.rewards import RewardClassifierConfig, make_reward_model, make_reward_pre_post_processors
 
 
 def main():
@@ -22,10 +22,10 @@ def main():
         model_name="microsoft/resnet-18",
     )
 
-    # Make policy, preprocessor, and optimizer
-    policy = make_policy(config, ds_meta=dataset.meta)
-    optimizer = config.get_optimizer_preset().build(policy.parameters())
-    preprocessor, _ = make_pre_post_processors(policy_cfg=config, dataset_stats=dataset.meta.stats)
+    # Make reward model, preprocessor, and optimizer
+    reward_model = make_reward_model(config, dataset_stats=dataset.meta.stats)
+    optimizer = config.get_optimizer_preset().build(reward_model.parameters())
+    preprocessor, _ = make_reward_pre_post_processors(config, dataset_stats=dataset.meta.stats)
 
     classifier_id = "<user>/reward_classifier_hil_serl_example"
 
@@ -42,7 +42,7 @@ def main():
             batch = preprocessor(batch)
 
             # Forward pass
-            loss, output_dict = policy.forward(batch)
+            loss, output_dict = reward_model.forward(batch)
 
             # Backward pass and optimization
             optimizer.zero_grad()
@@ -58,8 +58,8 @@ def main():
 
     print("Training finished!")
 
-    # You can now save the trained policy.
-    policy.push_to_hub(classifier_id)
+    # You can now save the trained reward model.
+    reward_model.push_to_hub(classifier_id)
 
 
 if __name__ == "__main__":

pyproject.toml

@@ -128,7 +128,7 @@ dataset_viz = ["lerobot[dataset]", "lerobot[viz]"]
 av-dep = ["av>=15.0.0,<16.0.0"]
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.9.17"]
-transformers-dep = ["transformers==5.3.0"] # TODO(Steven): https://github.com/huggingface/lerobot/pull/3249
+transformers-dep = ["transformers>=5.4.0,<5.6.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
 can-dep = ["python-can>=4.2.0,<5.0.0"]
 peft-dep = ["peft>=0.18.0,<1.0.0"]
@@ -289,6 +289,7 @@ lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
 lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
 lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
 lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
+lerobot-rollout="lerobot.scripts.lerobot_rollout:main"
 
 # ---------------- Tool Configurations ----------------
 [tool.setuptools.package-data]

src/lerobot/cameras/realsense/camera_realsense.py

@@ -17,6 +17,7 @@ Provides the RealSenseCamera class for capturing frames from Intel RealSense cam
 """
 
 import logging
+import sys
 import time
 from threading import Event, Lock, Thread
 from typing import TYPE_CHECKING, Any
@@ -41,6 +42,7 @@ from ..utils import get_cv2_rotation
 from .configuration_realsense import RealSenseCameraConfig
 
 logger = logging.getLogger(__name__)
+pkg_name = "pyrealsense2-macosx" if sys.platform == "darwin" else "pyrealsense2"
 
 
 class RealSenseCamera(Camera):
@@ -114,7 +116,7 @@ class RealSenseCamera(Camera):
         Args:
             config: The configuration settings for the camera.
         """
-        require_package("pyrealsense2", extra="intelrealsense")
+        require_package(pkg_name, extra="intelrealsense", import_name="pyrealsense2")
         super().__init__(config)
 
         self.config = config

src/lerobot/common/wandb_utils.py

@@ -41,8 +41,12 @@ def cfg_to_group(
             return tag
         return tag[:max_tag_length]
 
+    if cfg.is_reward_model_training:
+        trainable_tag = f"reward_model:{cfg.reward_model.type}"
+    else:
+        trainable_tag = f"policy:{cfg.policy.type}"
     lst = [
-        f"policy:{cfg.policy.type}",
+        trainable_tag,
         f"seed:{cfg.seed}",
     ]
     if cfg.dataset is not None:

src/lerobot/configs/__init__.py

@@ -21,6 +21,7 @@ are intentionally NOT re-exported here to avoid circular dependencies
 Import them directly: ``from lerobot.configs.train import TrainPipelineConfig``
 """
 
+from .dataset import DatasetRecordConfig
 from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
 from .recipe import MessageTurn, TrainingRecipe, load_recipe
@@ -40,6 +41,7 @@ __all__ = [
     "PolicyFeature",
     "RTCAttentionSchedule",
     # Config classes
+    "DatasetRecordConfig",
     "DatasetConfig",
     "EvalConfig",
     "MessageTurn",

src/lerobot/configs/dataset.py

@@ -0,0 +1,80 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Shared dataset recording configuration used by both ``lerobot-record`` and ``lerobot-rollout``."""
+
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+
+
+@dataclass
+class DatasetRecordConfig:
+    # Dataset identifier. By convention it should match '{hf_username}/{dataset_name}' (e.g. `lerobot/test`).
+    repo_id: str = ""
+    # A short but accurate description of the task performed during the recording (e.g. "Pick the Lego block and drop it in the box on the right.")
+    single_task: str = ""
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    root: str | Path | None = None
+    # Limit the frames per second.
+    fps: int = 30
+    # Number of seconds for data recording for each episode.
+    episode_time_s: int | float = 60
+    # Number of seconds for resetting the environment after each episode.
+    reset_time_s: int | float = 60
+    # Number of episodes to record.
+    num_episodes: int = 50
+    # Encode frames in the dataset into video
+    video: bool = True
+    # Upload dataset to Hugging Face hub.
+    push_to_hub: bool = True
+    # Upload on private repository on the Hugging Face hub.
+    private: bool = False
+    # Add tags to your dataset on the hub.
+    tags: list[str] | None = None
+    # Number of subprocesses handling the saving of frames as PNG. Set to 0 to use threads only;
+    # set to ≥1 to use subprocesses, each using threads to write images. The best number of processes
+    # and threads depends on your system. We recommend 4 threads per camera with 0 processes.
+    # If fps is unstable, adjust the thread count. If still unstable, try using 1 or more subprocesses.
+    num_image_writer_processes: int = 0
+    # Number of threads writing the frames as png images on disk, per camera.
+    # Too many threads might cause unstable teleoperation fps due to main thread being blocked.
+    # Not enough threads might cause low camera fps.
+    num_image_writer_threads_per_camera: int = 4
+    # Number of episodes to record before batch encoding videos
+    # Set to 1 for immediate encoding (default behavior), or higher for batched encoding
+    video_encoding_batch_size: int = 1
+    # Video codec for encoding videos. Options: 'h264', 'hevc', 'libsvtav1', 'auto',
+    # or hardware-specific: 'h264_videotoolbox', 'h264_nvenc', 'h264_vaapi', 'h264_qsv'.
+    # Use 'auto' to auto-detect the best available hardware encoder.
+    vcodec: str = "libsvtav1"
+    # Enable streaming video encoding: encode frames in real-time during capture instead
+    # of writing PNG images first. Makes save_episode() near-instant. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding
+    streaming_encoding: bool = False
+    # Maximum number of frames to buffer per camera when using streaming encoding.
+    # ~1s buffer at 30fps. Provides backpressure if the encoder can't keep up.
+    encoder_queue_maxsize: int = 30
+    # Number of threads per encoder instance. None = auto (codec default).
+    # Lower values reduce CPU usage, maps to 'lp' (via svtav1-params) for libsvtav1 and 'threads' for h264/hevc..
+    encoder_threads: int | None = None
+
+    def stamp_repo_id(self) -> None:
+        """Append a date-time tag to ``repo_id`` so each recording session gets a unique name.
+
+        Must be called explicitly at dataset *creation* time — not on resume,
+        where the existing ``repo_id`` (already stamped) must be preserved.
+        """
+        if self.repo_id:
+            timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+            self.repo_id = f"{self.repo_id}_{timestamp}"

src/lerobot/configs/recipes/pi05_hirobot.yaml

@@ -1,74 +0,0 @@
-blend:
-
-  memory_update:
-    weight: 0.10
-    bindings:
-      prior_memory: "nth_prev(style=memory, offset=1)"
-      current_memory: "emitted_at(t, style=memory)"
-      completed_subtask: "nth_prev(style=subtask, offset=1)"
-    messages:
-      - {role: user, content: "${task}", stream: high_level}
-      - {role: assistant, content: "Previous memory: ${prior_memory}", stream: high_level, if_present: prior_memory}
-      - {role: user, content: "Completed subtask: ${completed_subtask}", stream: high_level, if_present: completed_subtask}
-      - {role: assistant, content: "${current_memory}", stream: high_level, target: true, if_present: current_memory}
-
-  user_interjection_response:
-    weight: 0.16
-    bindings:
-      prior_plan: "nth_prev(style=plan, offset=1)"
-      current_plan: "emitted_at(t, style=plan)"
-      interjection: "emitted_at(t, style=interjection)"
-      speech: "emitted_at(t, role=assistant, tool_name=say)"
-    messages:
-      - {role: user, content: "${task}", stream: high_level}
-      - {role: assistant, content: "Previous plan:\n${prior_plan}", stream: high_level, if_present: prior_plan}
-      - {role: user, content: "${interjection}", stream: high_level, if_present: interjection}
-      - {role: assistant, content: "${current_plan}", stream: high_level, target: true, if_present: current_plan, tool_calls_from: speech}
-
-  high_level_subtask:
-    weight: 0.15
-    bindings:
-      next_subtask: "nth_next(style=subtask, offset=1)"
-    messages:
-      - {role: user, content: "${task}\nPlan: ${plan}\nMemory: ${memory}", stream: high_level}
-      - {role: user, content: "Current subtask: ${subtask}", stream: high_level, if_present: subtask}
-      - {role: assistant, content: "${next_subtask}", stream: high_level, target: true}
-
-  low_level_execution:
-    weight: 0.35
-    messages:
-      - {role: user, content: "${task}\nPlan: ${plan}\nMemory: ${memory}", stream: high_level}
-      - {role: assistant, content: "${subtask}", stream: low_level, target: true}
-
-  # VQA is view-dependent: bbox / keypoint / count answers only make sense for
-  # the camera they were grounded against. Each camera gets its own sub-recipe
-  # so the resolver can disambiguate via `camera=...` and the user-turn carries
-  # the matching image block. Adjust the camera keys (and add more sub-recipes)
-  # to match the cameras present on your dataset.
-  ask_vqa_top:
-    weight: 0.10
-    bindings:
-      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.top)"
-      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.top)"
-    messages:
-      - role: user
-        stream: high_level
-        if_present: vqa_query
-        content:
-          - {type: image, feature: observation.images.top}
-          - {type: text, text: "${vqa_query}"}
-      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}
-
-  ask_vqa_wrist:
-    weight: 0.10
-    bindings:
-      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.wrist)"
-      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.wrist)"
-    messages:
-      - role: user
-        stream: high_level
-        if_present: vqa_query
-        content:
-          - {type: image, feature: observation.images.wrist}
-          - {type: text, text: "${vqa_query}"}
-      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}

src/lerobot/configs/rewards.py

@@ -0,0 +1,163 @@
+# 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.
+
+import abc
+import builtins
+import json
+import logging
+import os
+import tempfile
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, TypeVar
+
+import draccus
+from huggingface_hub import hf_hub_download
+from huggingface_hub.constants import CONFIG_NAME
+from huggingface_hub.errors import HfHubHTTPError
+
+from lerobot.configs.types import PolicyFeature
+from lerobot.optim.optimizers import OptimizerConfig
+from lerobot.optim.schedulers import LRSchedulerConfig
+from lerobot.utils.device_utils import auto_select_torch_device, is_torch_device_available
+from lerobot.utils.hub import HubMixin
+
+T = TypeVar("T", bound="RewardModelConfig")
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class RewardModelConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
+    """Base configuration for reward models.
+
+    Args:
+    input_features: A dictionary defining the PolicyFeature of the input data for the reward. 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 reward. The key represents
+        the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+    """
+
+    # Reuses PolicyFeature
+    input_features: dict[str, PolicyFeature] = field(default_factory=dict)
+    output_features: dict[str, PolicyFeature] = field(default_factory=dict)
+
+    device: str | None = None
+
+    pretrained_path: str | None = None
+
+    push_to_hub: bool = False
+    repo_id: str | None = None
+
+    # Hub metadata
+    license: str | None = None
+    tags: list[str] | None = None
+    private: bool | None = None
+
+    def __post_init__(self) -> None:
+        if not self.device or not is_torch_device_available(self.device):
+            auto_device = auto_select_torch_device()
+            logger.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")
+            self.device = auto_device.type
+
+    @property
+    def type(self) -> str:
+        choice_name = self.get_choice_name(self.__class__)
+        if not isinstance(choice_name, str):
+            raise TypeError(f"Expected string from get_choice_name, got {type(choice_name)}")
+        return choice_name
+
+    @property
+    def observation_delta_indices(self) -> list | None:  # type: ignore[type-arg]
+        return None
+
+    @property
+    def action_delta_indices(self) -> list | None:  # type: ignore[type-arg]
+        return None
+
+    @property
+    def reward_delta_indices(self) -> list | None:  # type: ignore[type-arg]
+        return None
+
+    @abc.abstractmethod
+    def get_optimizer_preset(self) -> OptimizerConfig:
+        raise NotImplementedError
+
+    def get_scheduler_preset(self) -> LRSchedulerConfig | None:
+        return None
+
+    def validate_features(self) -> None:
+        pass
+
+    def _save_pretrained(self, save_directory: Path) -> None:
+        with open(save_directory / CONFIG_NAME, "w") as f, draccus.config_type("json"):
+            draccus.dump(self, f, indent=4)
+
+    @classmethod
+    def from_pretrained(
+        cls: builtins.type[T],
+        pretrained_name_or_path: str | Path,
+        *,
+        force_download: bool = False,
+        resume_download: bool | None = None,
+        proxies: dict[Any, Any] | None = None,
+        token: str | bool | None = None,
+        cache_dir: str | Path | None = None,
+        local_files_only: bool = False,
+        revision: str | None = None,
+        **reward_kwargs: Any,
+    ) -> T:
+        model_id = str(pretrained_name_or_path)
+        config_file: str | None = None
+        if Path(model_id).is_dir():
+            if CONFIG_NAME in os.listdir(model_id):
+                config_file = os.path.join(model_id, CONFIG_NAME)
+            else:
+                logger.error(f"{CONFIG_NAME} not found in {Path(model_id).resolve()}")
+        else:
+            try:
+                config_file = hf_hub_download(
+                    repo_id=model_id,
+                    filename=CONFIG_NAME,
+                    revision=revision,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    resume_download=resume_download,
+                    token=token,
+                    local_files_only=local_files_only,
+                )
+            except HfHubHTTPError as e:
+                raise FileNotFoundError(
+                    f"{CONFIG_NAME} not found on the HuggingFace Hub in {model_id}"
+                ) from e
+
+        if config_file is None:
+            raise FileNotFoundError(f"{CONFIG_NAME} not found in {model_id}")
+
+        # HACK: Parse the original config to get the config subclass, so that we can
+        # apply cli overrides.
+        with draccus.config_type("json"):
+            orig_config = draccus.parse(cls, config_file, args=[])
+
+        with open(config_file) as f:
+            config = json.load(f)
+
+        config.pop("type", None)
+        with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as f:
+            json.dump(config, f)
+            config_file = f.name
+
+        cli_overrides = reward_kwargs.pop("cli_overrides", [])
+        with draccus.config_type("json"):
+            return draccus.parse(orig_config.__class__, config_file, args=cli_overrides)

src/lerobot/configs/train.py

@@ -13,7 +13,9 @@
 # limitations under the License.
 import builtins
 import datetime as dt
+import json
 import os
+import tempfile
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any
@@ -26,18 +28,57 @@ from lerobot import envs
 from lerobot.configs import parser
 from lerobot.optim import LRSchedulerConfig, OptimizerConfig
 from lerobot.utils.hub import HubMixin
+from lerobot.utils.sample_weighting import SampleWeightingConfig
 
 from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
+from .rewards import RewardModelConfig
 
 TRAIN_CONFIG_NAME = "train_config.json"
 
 
+def _migrate_legacy_rabc_fields(config: dict[str, Any]) -> dict[str, Any] | None:
+    """Return migrated payload for legacy RA-BC fields, or None when no migration is needed."""
+    legacy_fields = (
+        "use_rabc",
+        "rabc_progress_path",
+        "rabc_kappa",
+        "rabc_epsilon",
+        "rabc_head_mode",
+    )
+    if not any(key in config for key in legacy_fields):
+        return None
+
+    migrated_config = dict(config)
+    use_rabc = bool(migrated_config.pop("use_rabc", False))
+    rabc_progress_path = migrated_config.pop("rabc_progress_path", None)
+    rabc_kappa = migrated_config.pop("rabc_kappa", None)
+    rabc_epsilon = migrated_config.pop("rabc_epsilon", None)
+    rabc_head_mode = migrated_config.pop("rabc_head_mode", None)
+
+    # New configs may already define sample_weighting explicitly. In that case,
+    # legacy fields are ignored after being stripped from the payload.
+    if migrated_config.get("sample_weighting") is None and use_rabc:
+        sample_weighting: dict[str, Any] = {"type": "rabc"}
+        if rabc_progress_path is not None:
+            sample_weighting["progress_path"] = rabc_progress_path
+        if rabc_kappa is not None:
+            sample_weighting["kappa"] = rabc_kappa
+        if rabc_epsilon is not None:
+            sample_weighting["epsilon"] = rabc_epsilon
+        if rabc_head_mode is not None:
+            sample_weighting["head_mode"] = rabc_head_mode
+        migrated_config["sample_weighting"] = sample_weighting
+
+    return migrated_config
+
+
 @dataclass
 class TrainPipelineConfig(HubMixin):
     dataset: DatasetConfig
     env: envs.EnvConfig | None = None
     policy: PreTrainedConfig | None = None
+    reward_model: RewardModelConfig | None = None
     # Set `dir` to where you would like to save all of the run outputs. If you run another training session
     # with the same value for `dir` its contents will be overwritten unless you set `resume` to true.
     output_dir: Path | None = None
@@ -72,27 +113,41 @@ class TrainPipelineConfig(HubMixin):
     wandb: WandBConfig = field(default_factory=WandBConfig)
     peft: PeftConfig | None = None
 
-    # RA-BC (Reward-Aligned Behavior Cloning) parameters
-    use_rabc: bool = False  # Enable reward-weighted training
-    rabc_progress_path: str | None = None  # Path to precomputed SARM progress parquet file
-    rabc_kappa: float = 0.01  # Hard threshold for high-quality samples
-    rabc_epsilon: float = 1e-6  # Small constant for numerical stability
-    rabc_head_mode: str | None = "sparse"  # For dual-head models: "sparse" or "dense"
+    # Sample weighting configuration (e.g., for RA-BC training)
+    sample_weighting: SampleWeightingConfig | None = None
 
     # Rename map for the observation to override the image and state keys
     rename_map: dict[str, str] = field(default_factory=dict)
     checkpoint_path: Path | None = field(init=False, default=None)
 
+    @property
+    def is_reward_model_training(self) -> bool:
+        """True when the config targets a reward model rather than a policy."""
+        return self.reward_model is not None
+
+    @property
+    def trainable_config(self) -> PreTrainedConfig | RewardModelConfig:
+        """Return whichever config (policy or reward_model) is active."""
+        if self.is_reward_model_training:
+            return self.reward_model  # type: ignore[return-value]
+        return self.policy  # type: ignore[return-value]
+
     def validate(self) -> None:
         # HACK: We parse again the cli args here to get the pretrained paths if there was some.
         policy_path = parser.get_path_arg("policy")
-        if policy_path:
-            # Only load the policy config
+        reward_model_path = parser.get_path_arg("reward_model")
+
+        if reward_model_path:
+            cli_overrides = parser.get_cli_overrides("reward_model")
+            self.reward_model = RewardModelConfig.from_pretrained(
+                reward_model_path, cli_overrides=cli_overrides
+            )
+            self.reward_model.pretrained_path = str(Path(reward_model_path))
+        elif policy_path:
             cli_overrides = parser.get_cli_overrides("policy")
             self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
             self.policy.pretrained_path = Path(policy_path)
         elif self.resume:
-            # The entire train config is already loaded, we just need to get the checkpoint dir
             config_path = parser.parse_arg("config_path")
             if not config_path:
                 raise ValueError(
@@ -108,18 +163,22 @@ class TrainPipelineConfig(HubMixin):
             policy_dir = Path(config_path).parent
             if self.policy is not None:
                 self.policy.pretrained_path = policy_dir
+            if self.reward_model is not None:
+                self.reward_model.pretrained_path = str(policy_dir)
             self.checkpoint_path = policy_dir.parent
 
-        if self.policy is None:
+        if self.policy is None and self.reward_model is None:
             raise ValueError(
-                "Policy is not configured. Please specify a pretrained policy with `--policy.path`."
+                "Neither policy nor reward_model is configured. "
+                "Please specify one with `--policy.path` or `--reward_model.path`."
             )
 
+        active_cfg = self.trainable_config
         if not self.job_name:
             if self.env is None:
-                self.job_name = f"{self.policy.type}"
+                self.job_name = f"{active_cfg.type}"
             else:
-                self.job_name = f"{self.env.type}_{self.policy.type}"
+                self.job_name = f"{self.env.type}_{active_cfg.type}"
 
         if not self.resume and isinstance(self.output_dir, Path) and self.output_dir.is_dir():
             raise FileExistsError(
@@ -137,26 +196,16 @@ class TrainPipelineConfig(HubMixin):
         if not self.use_policy_training_preset and (self.optimizer is None or self.scheduler is None):
             raise ValueError("Optimizer and Scheduler must be set when the policy presets are not used.")
         elif self.use_policy_training_preset and not self.resume:
-            self.optimizer = self.policy.get_optimizer_preset()
-            self.scheduler = self.policy.get_scheduler_preset()
+            self.optimizer = active_cfg.get_optimizer_preset()
+            self.scheduler = active_cfg.get_scheduler_preset()
 
-        if self.policy.push_to_hub and not self.policy.repo_id:
-            raise ValueError(
-                "'policy.repo_id' argument missing. Please specify it to push the model to the hub."
-            )
-
-        if self.use_rabc and not self.rabc_progress_path:
-            # Auto-detect from dataset path
-            repo_id = self.dataset.repo_id
-            if self.dataset.root:
-                self.rabc_progress_path = str(Path(self.dataset.root) / "sarm_progress.parquet")
-            else:
-                self.rabc_progress_path = f"hf://datasets/{repo_id}/sarm_progress.parquet"
+        if hasattr(active_cfg, "push_to_hub") and active_cfg.push_to_hub and not active_cfg.repo_id:
+            raise ValueError("'repo_id' argument missing. Please specify it to push the model to the hub.")
 
     @classmethod
     def __get_path_fields__(cls) -> list[str]:
-        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
-        return ["policy"]
+        """Keys for draccus pretrained-path loading."""
+        return ["policy", "reward_model"]
 
     def to_dict(self) -> dict[str, Any]:
         return draccus.encode(self)  # type: ignore[no-any-return]  # because of the third-party library draccus uses Any as the return type
@@ -207,6 +256,15 @@ class TrainPipelineConfig(HubMixin):
                 ) from e
 
         cli_args = kwargs.pop("cli_args", [])
+        if config_file is not None:
+            with open(config_file) as f:
+                config = json.load(f)
+            migrated_config = _migrate_legacy_rabc_fields(config)
+            if migrated_config is not None:
+                with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as f:
+                    json.dump(migrated_config, f)
+                    config_file = f.name
+
         with draccus.config_type("json"):
             return draccus.parse(cls, config_file, args=cli_args)
 

src/lerobot/datasets/aggregate.py

@@ -97,8 +97,8 @@ def update_data_df(df, src_meta, dst_meta):
         pd.DataFrame: Updated DataFrame with adjusted indices.
     """
 
-    df["episode_index"] = df["episode_index"] + dst_meta.info["total_episodes"]
-    df["index"] = df["index"] + dst_meta.info["total_frames"]
+    df["episode_index"] = df["episode_index"] + dst_meta.info.total_episodes
+    df["index"] = df["index"] + dst_meta.info.total_frames
 
     src_task_names = src_meta.tasks.index.take(df["task_index"].to_numpy())
     df["task_index"] = dst_meta.tasks.loc[src_task_names, "task_index"].to_numpy()
@@ -225,9 +225,9 @@ def update_meta_data(
         # Clean up temporary columns
         df = df.drop(columns=["_orig_chunk", "_orig_file"])
 
-    df["dataset_from_index"] = df["dataset_from_index"] + dst_meta.info["total_frames"]
-    df["dataset_to_index"] = df["dataset_to_index"] + dst_meta.info["total_frames"]
-    df["episode_index"] = df["episode_index"] + dst_meta.info["total_episodes"]
+    df["dataset_from_index"] = df["dataset_from_index"] + dst_meta.info.total_frames
+    df["dataset_to_index"] = df["dataset_to_index"] + dst_meta.info.total_frames
+    df["episode_index"] = df["episode_index"] + dst_meta.info.total_episodes
 
     return df
 
@@ -237,8 +237,8 @@ def aggregate_datasets(
     aggr_repo_id: str,
     roots: list[Path] | None = None,
     aggr_root: Path | None = None,
-    data_files_size_in_mb: float | None = None,
-    video_files_size_in_mb: float | None = None,
+    data_files_size_in_mb: int | None = None,
+    video_files_size_in_mb: int | None = None,
     chunk_size: int | None = None,
 ):
     """Aggregates multiple LeRobot datasets into a single unified dataset.
@@ -313,8 +313,8 @@ def aggregate_datasets(
         # to avoid interference between different source datasets
         data_idx.pop("src_to_dst", None)
 
-        dst_meta.info["total_episodes"] += src_meta.total_episodes
-        dst_meta.info["total_frames"] += src_meta.total_frames
+        dst_meta.info.total_episodes += src_meta.total_episodes
+        dst_meta.info.total_frames += src_meta.total_frames
 
     finalize_aggregation(dst_meta, all_metadata)
     logging.info("Aggregation complete.")
@@ -640,14 +640,10 @@ def finalize_aggregation(aggr_meta, all_metadata):
     write_tasks(aggr_meta.tasks, aggr_meta.root)
 
     logging.info("write info")
-    aggr_meta.info.update(
-        {
-            "total_tasks": len(aggr_meta.tasks),
-            "total_episodes": sum(m.total_episodes for m in all_metadata),
-            "total_frames": sum(m.total_frames for m in all_metadata),
-            "splits": {"train": f"0:{sum(m.total_episodes for m in all_metadata)}"},
-        }
-    )
+    aggr_meta.info.total_tasks = len(aggr_meta.tasks)
+    aggr_meta.info.total_episodes = sum(m.total_episodes for m in all_metadata)
+    aggr_meta.info.total_frames = sum(m.total_frames for m in all_metadata)
+    aggr_meta.info.splits = {"train": f"0:{sum(m.total_episodes for m in all_metadata)}"}
     write_info(aggr_meta.info, aggr_meta.root)
 
     logging.info("write stats")

src/lerobot/datasets/dataset_metadata.py

@@ -36,13 +36,11 @@ from .io_utils import (
     load_stats,
     load_tasks,
     write_info,
-    write_json,
     write_stats,
     write_tasks,
 )
 from .utils import (
     DEFAULT_EPISODES_PATH,
-    INFO_PATH,
     check_version_compatibility,
     get_safe_version,
     has_legacy_hub_download_metadata,
@@ -226,7 +224,7 @@ class LeRobotDatasetMetadata:
     @property
     def _version(self) -> packaging.version.Version:
         """Codebase version used to create this dataset."""
-        return packaging.version.parse(self.info["codebase_version"])
+        return packaging.version.parse(self.info.codebase_version)
 
     def get_data_file_path(self, ep_index: int) -> Path:
         """Return the relative parquet file path for the given episode index.
@@ -281,27 +279,27 @@ class LeRobotDatasetMetadata:
     @property
     def data_path(self) -> str:
         """Formattable string for the parquet files."""
-        return self.info["data_path"]
+        return self.info.data_path
 
     @property
     def video_path(self) -> str | None:
         """Formattable string for the video files."""
-        return self.info["video_path"]
+        return self.info.video_path
 
     @property
     def robot_type(self) -> str | None:
         """Robot type used in recording this dataset."""
-        return self.info["robot_type"]
+        return self.info.robot_type
 
     @property
     def fps(self) -> int:
         """Frames per second used during data collection."""
-        return self.info["fps"]
+        return self.info.fps
 
     @property
     def features(self) -> dict[str, dict]:
         """All features contained in the dataset."""
-        return self.info["features"]
+        return self.info.features
 
     @property
     def image_keys(self) -> list[str]:
@@ -318,6 +316,17 @@ class LeRobotDatasetMetadata:
         """Keys to access visual modalities (regardless of their storage method)."""
         return [key for key, ft in self.features.items() if ft["dtype"] in ["video", "image"]]
 
+    @property
+    def has_language_columns(self) -> bool:
+        """Return ``True`` if the dataset declares any language column.
+
+        Used to gate language-aware code paths (collate, render step) so
+        unannotated datasets keep PyTorch's default collate behavior.
+        """
+        from .language import LANGUAGE_COLUMNS  # noqa: PLC0415  (avoid circular import)
+
+        return any(col in self.features for col in LANGUAGE_COLUMNS)
+
     @property
     def tools(self) -> list[dict]:
         """OpenAI-style tool schemas declared by this dataset.
@@ -335,8 +344,8 @@ class LeRobotDatasetMetadata:
         """
         from .language import DEFAULT_TOOLS  # noqa: PLC0415  (avoid circular import)
 
-        declared = self.info.get("tools")
-        if isinstance(declared, list) and declared:
+        declared = self.info.tools
+        if declared:
             return [dict(t) for t in declared]
         return [dict(t) for t in DEFAULT_TOOLS]
 
@@ -353,32 +362,32 @@ class LeRobotDatasetMetadata:
     @property
     def total_episodes(self) -> int:
         """Total number of episodes available."""
-        return self.info["total_episodes"]
+        return self.info.total_episodes
 
     @property
     def total_frames(self) -> int:
         """Total number of frames saved in this dataset."""
-        return self.info["total_frames"]
+        return self.info.total_frames
 
     @property
     def total_tasks(self) -> int:
         """Total number of different tasks performed in this dataset."""
-        return self.info["total_tasks"]
+        return self.info.total_tasks
 
     @property
     def chunks_size(self) -> int:
         """Max number of files per chunk."""
-        return self.info["chunks_size"]
+        return self.info.chunks_size
 
     @property
     def data_files_size_in_mb(self) -> int:
         """Max size of data file in mega bytes."""
-        return self.info["data_files_size_in_mb"]
+        return self.info.data_files_size_in_mb
 
     @property
     def video_files_size_in_mb(self) -> int:
         """Max size of video file in mega bytes."""
-        return self.info["video_files_size_in_mb"]
+        return self.info.video_files_size_in_mb
 
     def get_task_index(self, task: str) -> int | None:
         """
@@ -522,10 +531,10 @@ class LeRobotDatasetMetadata:
         self._save_episode_metadata(episode_dict)
 
         # Update info
-        self.info["total_episodes"] += 1
-        self.info["total_frames"] += episode_length
-        self.info["total_tasks"] = len(self.tasks)
-        self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}
+        self.info.total_episodes += 1
+        self.info.total_frames += episode_length
+        self.info.total_tasks = len(self.tasks)
+        self.info.splits = {"train": f"0:{self.info.total_episodes}"}
 
         write_info(self.info, self.root)
 
@@ -544,7 +553,7 @@ class LeRobotDatasetMetadata:
         for key in video_keys:
             if not self.features[key].get("info", None):
                 video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
-                self.info["features"][key]["info"] = get_video_info(video_path)
+                self.info.features[key]["info"] = get_video_info(video_path)
 
     def update_chunk_settings(
         self,
@@ -566,17 +575,17 @@ class LeRobotDatasetMetadata:
         if chunks_size is not None:
             if chunks_size <= 0:
                 raise ValueError(f"chunks_size must be positive, got {chunks_size}")
-            self.info["chunks_size"] = chunks_size
+            self.info.chunks_size = chunks_size
 
         if data_files_size_in_mb is not None:
             if data_files_size_in_mb <= 0:
                 raise ValueError(f"data_files_size_in_mb must be positive, got {data_files_size_in_mb}")
-            self.info["data_files_size_in_mb"] = data_files_size_in_mb
+            self.info.data_files_size_in_mb = data_files_size_in_mb
 
         if video_files_size_in_mb is not None:
             if video_files_size_in_mb <= 0:
                 raise ValueError(f"video_files_size_in_mb must be positive, got {video_files_size_in_mb}")
-            self.info["video_files_size_in_mb"] = video_files_size_in_mb
+            self.info.video_files_size_in_mb = video_files_size_in_mb
 
         # Update the info file on disk
         write_info(self.info, self.root)
@@ -672,7 +681,7 @@ class LeRobotDatasetMetadata:
                 f"Features contain video keys {obj.video_keys}, but 'use_videos' is set to False. "
                 "Either remove video features from the features dict, or set 'use_videos=True'."
             )
-        write_json(obj.info, obj.root / INFO_PATH)
+        write_info(obj.info, obj.root)
         obj.revision = None
         obj._pq_writer = None
         obj.latest_episode = None

src/lerobot/datasets/dataset_tools.py

@@ -897,14 +897,10 @@ def _copy_and_reindex_episodes_metadata(
 
     dst_meta.finalize()
 
-    dst_meta.info.update(
-        {
-            "total_episodes": len(episode_mapping),
-            "total_frames": total_frames,
-            "total_tasks": len(dst_meta.tasks) if dst_meta.tasks is not None else 0,
-            "splits": {"train": f"0:{len(episode_mapping)}"},
-        }
-    )
+    dst_meta.info.total_episodes = len(episode_mapping)
+    dst_meta.info.total_frames = total_frames
+    dst_meta.info.total_tasks = len(dst_meta.tasks) if dst_meta.tasks is not None else 0
+    dst_meta.info.splits = {"train": f"0:{len(episode_mapping)}"}
     write_info(dst_meta.info, dst_meta.root)
 
     if not all_stats:
@@ -1069,21 +1065,20 @@ def _copy_episodes_metadata_and_stats(
     if episodes_dir.exists():
         shutil.copytree(episodes_dir, dst_episodes_dir, dirs_exist_ok=True)
 
-    dst_meta.info.update(
-        {
-            "total_episodes": src_dataset.meta.total_episodes,
-            "total_frames": src_dataset.meta.total_frames,
-            "total_tasks": src_dataset.meta.total_tasks,
-            "splits": src_dataset.meta.info.get("splits", {"train": f"0:{src_dataset.meta.total_episodes}"}),
-        }
+    dst_meta.info.total_episodes = src_dataset.meta.total_episodes
+    dst_meta.info.total_frames = src_dataset.meta.total_frames
+    dst_meta.info.total_tasks = src_dataset.meta.total_tasks
+    # Preserve original splits if available, otherwise create default
+    dst_meta.info.splits = (
+        src_dataset.meta.info.splits
+        if src_dataset.meta.info.splits
+        else {"train": f"0:{src_dataset.meta.total_episodes}"}
     )
 
     if dst_meta.video_keys and src_dataset.meta.video_keys:
         for key in dst_meta.video_keys:
             if key in src_dataset.meta.features:
-                dst_meta.info["features"][key]["info"] = src_dataset.meta.info["features"][key].get(
-                    "info", {}
-                )
+                dst_meta.info.features[key]["info"] = src_dataset.meta.info.features[key].get("info", {})
 
     write_info(dst_meta.info, dst_meta.root)
 
@@ -1525,7 +1520,7 @@ def modify_tasks(
     write_tasks(new_task_df, root)
 
     # Update info.json
-    dataset.meta.info["total_tasks"] = len(unique_tasks)
+    dataset.meta.info.total_tasks = len(unique_tasks)
     write_info(dataset.meta.info, root)
 
     # Reload metadata to reflect changes
@@ -1858,10 +1853,10 @@ def convert_image_to_video_dataset(
         episodes_df.to_parquet(episodes_path, index=False)
 
         # Update metadata info
-        new_meta.info["total_episodes"] = len(episode_indices)
-        new_meta.info["total_frames"] = sum(ep["length"] for ep in all_episode_metadata.values())
-        new_meta.info["total_tasks"] = dataset.meta.total_tasks
-        new_meta.info["splits"] = {"train": f"0:{len(episode_indices)}"}
+        new_meta.info.total_episodes = len(episode_indices)
+        new_meta.info.total_frames = sum(ep["length"] for ep in all_episode_metadata.values())
+        new_meta.info.total_tasks = dataset.meta.total_tasks
+        new_meta.info.splits = {"train": f"0:{len(episode_indices)}"}
 
         # Update video info for all image keys (now videos)
         # We need to manually set video info since update_video_info() checks video_keys first
@@ -1870,7 +1865,7 @@ def convert_image_to_video_dataset(
                 video_path = new_meta.root / new_meta.video_path.format(
                     video_key=img_key, chunk_index=0, file_index=0
                 )
-                new_meta.info["features"][img_key]["info"] = get_video_info(video_path)
+                new_meta.info.features[img_key]["info"] = get_video_info(video_path)
 
         write_info(new_meta.info, new_meta.root)
 

src/lerobot/datasets/factory.py

@@ -19,6 +19,7 @@ from pprint import pformat
 import torch
 
 from lerobot.configs import PreTrainedConfig
+from lerobot.configs.rewards import RewardModelConfig
 from lerobot.configs.train import TrainPipelineConfig
 from lerobot.transforms import ImageTransforms
 from lerobot.utils.constants import ACTION, IMAGENET_STATS, OBS_PREFIX, REWARD
@@ -30,12 +31,14 @@ from .streaming_dataset import StreamingLeRobotDataset
 
 
 def resolve_delta_timestamps(
-    cfg: PreTrainedConfig, ds_meta: LeRobotDatasetMetadata
+    cfg: PreTrainedConfig | RewardModelConfig, ds_meta: LeRobotDatasetMetadata
 ) -> dict[str, list] | None:
-    """Resolves delta_timestamps by reading from the 'delta_indices' properties of the PreTrainedConfig.
+    """Resolves delta_timestamps by reading from the 'delta_indices' properties of the config.
 
     Args:
-        cfg (PreTrainedConfig): The PreTrainedConfig to read delta_indices from.
+        cfg (PreTrainedConfig | RewardModelConfig): The config to read delta_indices from. Both
+            ``PreTrainedConfig`` and concrete ``RewardModelConfig`` subclasses expose the
+            ``{observation,action,reward}_delta_indices`` properties used below.
         ds_meta (LeRobotDatasetMetadata): The dataset from which features and fps are used to build
             delta_timestamps against.
 
@@ -82,7 +85,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
         ds_meta = LeRobotDatasetMetadata(
             cfg.dataset.repo_id, root=cfg.dataset.root, revision=cfg.dataset.revision
         )
-        delta_timestamps = resolve_delta_timestamps(cfg.policy, ds_meta)
+        delta_timestamps = resolve_delta_timestamps(cfg.trainable_config, ds_meta)
         if not cfg.dataset.streaming:
             dataset = LeRobotDataset(
                 cfg.dataset.repo_id,

src/lerobot/datasets/feature_utils.py

@@ -34,6 +34,7 @@ from .utils import (
     DEFAULT_DATA_PATH,
     DEFAULT_VIDEO_FILE_SIZE_IN_MB,
     DEFAULT_VIDEO_PATH,
+    DatasetInfo,
 )
 
 
@@ -90,8 +91,8 @@ def create_empty_dataset_info(
     chunks_size: int | None = None,
     data_files_size_in_mb: int | None = None,
     video_files_size_in_mb: int | None = None,
-) -> dict:
-    """Create a template dictionary for a new dataset's `info.json`.
+) -> DatasetInfo:
+    """Create a template ``DatasetInfo`` object for a new dataset's ``meta/info.json``.
 
     Args:
         codebase_version (str): The version of the LeRobot codebase.
@@ -99,25 +100,24 @@ def create_empty_dataset_info(
         features (dict): The LeRobot features dictionary for the dataset.
         use_videos (bool): Whether the dataset will store videos.
         robot_type (str | None): The type of robot used, if any.
+        chunks_size (int | None): Max files per chunk directory. Defaults to ``DEFAULT_CHUNK_SIZE``.
+        data_files_size_in_mb (int | None): Max parquet file size in MB. Defaults to ``DEFAULT_DATA_FILE_SIZE_IN_MB``.
+        video_files_size_in_mb (int | None): Max video file size in MB. Defaults to ``DEFAULT_VIDEO_FILE_SIZE_IN_MB``.
 
     Returns:
-        dict: A dictionary with the initial dataset metadata.
+        DatasetInfo: A typed dataset information object with initial metadata.
     """
-    return {
-        "codebase_version": codebase_version,
-        "robot_type": robot_type,
-        "total_episodes": 0,
-        "total_frames": 0,
-        "total_tasks": 0,
-        "chunks_size": chunks_size or DEFAULT_CHUNK_SIZE,
-        "data_files_size_in_mb": data_files_size_in_mb or DEFAULT_DATA_FILE_SIZE_IN_MB,
-        "video_files_size_in_mb": video_files_size_in_mb or DEFAULT_VIDEO_FILE_SIZE_IN_MB,
-        "fps": fps,
-        "splits": {},
-        "data_path": DEFAULT_DATA_PATH,
-        "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
-        "features": features,
-    }
+    return DatasetInfo(
+        codebase_version=codebase_version,
+        fps=fps,
+        features=features,
+        robot_type=robot_type,
+        chunks_size=chunks_size or DEFAULT_CHUNK_SIZE,
+        data_files_size_in_mb=data_files_size_in_mb or DEFAULT_DATA_FILE_SIZE_IN_MB,
+        video_files_size_in_mb=video_files_size_in_mb or DEFAULT_VIDEO_FILE_SIZE_IN_MB,
+        data_path=DEFAULT_DATA_PATH,
+        video_path=DEFAULT_VIDEO_PATH if use_videos else None,
+    )
 
 
 def check_delta_timestamps(

src/lerobot/datasets/io_utils.py

@@ -38,6 +38,7 @@ from .utils import (
     EPISODES_DIR,
     INFO_PATH,
     STATS_PATH,
+    DatasetInfo,
     serialize_dict,
 )
 
@@ -114,25 +115,21 @@ def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
     return dataset
 
 
-def write_info(info: dict, local_dir: Path) -> None:
-    write_json(info, local_dir / INFO_PATH)
+def write_info(info: DatasetInfo, local_dir: Path) -> None:
+    write_json(info.to_dict(), local_dir / INFO_PATH)
 
 
-def load_info(local_dir: Path) -> dict:
+def load_info(local_dir: Path) -> DatasetInfo:
     """Load dataset info metadata from its standard file path.
 
-    Also converts shape lists to tuples for consistency.
-
     Args:
         local_dir (Path): The root directory of the dataset.
 
     Returns:
-        dict: The dataset information dictionary.
+        DatasetInfo: The typed dataset information object.
     """
-    info = load_json(local_dir / INFO_PATH)
-    for ft in info["features"].values():
-        ft["shape"] = tuple(ft["shape"])
-    return info
+    raw = load_json(local_dir / INFO_PATH)
+    return DatasetInfo.from_dict(raw)
 
 
 def write_stats(stats: dict, local_dir: Path) -> None:

src/lerobot/datasets/language.py

@@ -178,9 +178,7 @@ def validate_camera_field(style: str | None, camera: str | None) -> None:
                 f"field referencing an 'observation.images.*' feature key."
             )
     elif camera is not None:
-        raise ValueError(
-            f"Rows of style {style!r} must have camera=None; got camera={camera!r}."
-        )
+        raise ValueError(f"Rows of style {style!r} must have camera=None; got camera={camera!r}.")
 
 
 # --- Tool registry --------------------------------------------------------

src/lerobot/datasets/language_render.py

@@ -24,12 +24,7 @@ from typing import Any
 
 from lerobot.configs.recipe import DEFAULT_BINDINGS, TrainingRecipe
 
-from .language import (
-    EVENT_ONLY_STYLES,
-    LANGUAGE_PERSISTENT,
-    PERSISTENT_STYLES,
-    column_for_style,
-)
+from .language import LANGUAGE_PERSISTENT, column_for_style
 
 LanguageRow = dict[str, Any]
 RenderedMessages = dict[str, list[Any]]
@@ -42,7 +37,6 @@ def active_at(
     t: float,
     *,
     persistent: Sequence[LanguageRow],
-    events: Sequence[LanguageRow] | None = None,
     style: str | None = None,
     role: str | None = None,
     tool_name: str | None = None,
@@ -52,16 +46,23 @@ def active_at(
 
     A persistent row is "active" at ``t`` when its own ``timestamp`` is the
     most recent one ``<= t`` for the given ``style``/``role``/``tool_name``/
-    ``camera`` selector. ``events`` is accepted for resolver-signature
-    uniformity but is not consulted: only persistent styles are valid here.
+    ``camera`` selector. Only valid for persistent styles.
     """
     _validate_persistent_resolver("active_at", style)
-    matches = _matching_rows(
-        persistent, style=style, role=role, tool_name=tool_name, camera=camera
-    )
-    matches = [row for row in matches if _timestamp(row) <= t]
-    return _select_latest(
-        matches, style=style, role=role, tool_name=tool_name, camera=camera
+    matches = [
+        row
+        for row in _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera)
+        if _timestamp(row) <= t
+    ]
+    if not matches:
+        return None
+    latest_ts = max(_timestamp(row) for row in matches)
+    return _select_one(
+        [row for row in matches if _timestamp(row) == latest_ts],
+        style=style,
+        role=role,
+        tool_name=tool_name,
+        camera=camera,
     )
 
 
@@ -84,41 +85,21 @@ def emitted_at(
     filters by the row's ``camera`` field — required to disambiguate when
     multiple view-dependent rows share ``(t, role)`` across cameras.
     """
-    column = column_for_style(style)
-    if column == LANGUAGE_PERSISTENT:
+    if column_for_style(style) == LANGUAGE_PERSISTENT:
         matches = [
             row
-            for row in _matching_rows(
-                persistent, style=style, role=role, tool_name=tool_name, camera=camera
-            )
+            for row in _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera)
             if _timestamp(row) == t
         ]
-        return _select_one(
-            matches,
-            style=style,
-            role=role,
-            tool_name=tool_name,
-            camera=camera,
-            sort_key=_persistent_sort_key,
-        )
-    matches = _matching_rows(
-        events, style=style, role=role, tool_name=tool_name, camera=camera
-    )
-    return _select_one(
-        matches,
-        style=style,
-        role=role,
-        tool_name=tool_name,
-        camera=camera,
-        sort_key=_event_sort_key,
-    )
+    else:
+        matches = _matching_rows(events, style=style, role=role, tool_name=tool_name, camera=camera)
+    return _select_one(matches, style=style, role=role, tool_name=tool_name, camera=camera)
 
 
 def nth_prev(
     t: float,
     *,
     persistent: Sequence[LanguageRow],
-    events: Sequence[LanguageRow] | None = None,
     style: str | None = None,
     offset: int = 1,
     role: str | None = None,
@@ -132,23 +113,13 @@ def nth_prev(
     one ``offset`` positions before the row active at ``t``. Only valid for
     persistent styles.
     """
-    return _nth_relative(
-        t,
-        persistent=persistent,
-        style=style,
-        offset=-offset,
-        role=role,
-        tool_name=tool_name,
-        camera=camera,
-        resolver_name="nth_prev",
-    )
+    return _nth_relative("nth_prev", t, persistent, style, -offset, role, tool_name, camera)
 
 
 def nth_next(
     t: float,
     *,
     persistent: Sequence[LanguageRow],
-    events: Sequence[LanguageRow] | None = None,
     style: str | None = None,
     offset: int = 1,
     role: str | None = None,
@@ -162,16 +133,7 @@ def nth_next(
     one ``offset`` positions after the row active at ``t``. Only valid for
     persistent styles.
     """
-    return _nth_relative(
-        t,
-        persistent=persistent,
-        style=style,
-        offset=offset,
-        role=role,
-        tool_name=tool_name,
-        camera=camera,
-        resolver_name="nth_next",
-    )
+    return _nth_relative("nth_next", t, persistent, style, offset, role, tool_name, camera)
 
 
 def render_sample(
@@ -239,9 +201,7 @@ def _resolve_bindings(
 ) -> dict[str, LanguageRow | str | None]:
     """Resolve every binding in ``recipe`` (plus ``task``) at time ``t``."""
     bindings: dict[str, LanguageRow | str | None] = {
-        "task": _resolve_task(
-            task, dataset_ctx, persistent=persistent, sample_idx=sample_idx
-        ),
+        "task": _resolve_task(task, dataset_ctx, persistent=persistent, sample_idx=sample_idx),
     }
     specs = {**DEFAULT_BINDINGS, **(recipe.bindings or {})}
     for name, spec in specs.items():
@@ -275,18 +235,12 @@ def _resolve_task(
     if task is not None:
         return task
 
-    aug_rows = [
-        r
-        for r in persistent
-        if r.get("style") == "task_aug" and r.get("role") == "user"
-    ]
+    aug_rows = [r for r in persistent if r.get("style") == "task_aug" and r.get("role") == "user"]
     if aug_rows:
         # Deterministic, blake2b-based pick keyed on sample_idx so the
         # rotation is reproducible across runs (Python's built-in ``hash``
         # is process-randomized).
-        digest = hashlib.blake2b(
-            f"task_aug:{sample_idx}".encode(), digest_size=8
-        ).digest()
+        digest = hashlib.blake2b(f"task_aug:{sample_idx}".encode(), digest_size=8).digest()
         idx = int.from_bytes(digest, "big") % len(aug_rows)
         chosen = aug_rows[idx].get("content")
         if chosen:
@@ -314,15 +268,15 @@ def _resolve_spec(
     kwargs = _parse_resolver_args(match.group("args"))
     kwargs.pop("t_arg", None)
 
-    resolvers = {
-        "active_at": active_at,
-        "emitted_at": emitted_at,
-        "nth_prev": nth_prev,
-        "nth_next": nth_next,
-    }
-    if name not in resolvers:
-        raise ValueError(f"Unknown language resolver: {name!r}")
-    return resolvers[name](t, persistent=persistent, events=events, **kwargs)
+    if name == "emitted_at":
+        return emitted_at(t, persistent=persistent, events=events, **kwargs)
+    if name == "active_at":
+        return active_at(t, persistent=persistent, **kwargs)
+    if name == "nth_prev":
+        return nth_prev(t, persistent=persistent, **kwargs)
+    if name == "nth_next":
+        return nth_next(t, persistent=persistent, **kwargs)
+    raise ValueError(f"Unknown language resolver: {name!r}")
 
 
 def _parse_resolver_args(args: str) -> dict[str, Any]:
@@ -444,24 +398,23 @@ def _validate_rendered(rendered: RenderedMessages) -> None:
 
 
 def _nth_relative(
+    name: str,
     t: float,
-    *,
     persistent: Sequence[LanguageRow],
     style: str | None,
     offset: int,
     role: str | None,
     tool_name: str | None,
     camera: str | None,
-    resolver_name: str,
 ) -> LanguageRow | None:
     """Shared body for ``nth_prev`` / ``nth_next`` with signed ``offset``."""
-    _validate_persistent_resolver(resolver_name, style)
+    _validate_persistent_resolver(name, style)
     if abs(offset) < 1:
-        raise ValueError(f"{resolver_name} offset must be non-zero.")
+        raise ValueError(f"{name} offset must be non-zero.")
 
     rows = sorted(
         _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera),
-        key=_persistent_sort_key,
+        key=_row_sort_key,
     )
     if not rows:
         return None
@@ -480,14 +433,12 @@ def _nth_relative(
     return rows[target_idx]
 
 
-def _validate_persistent_resolver(resolver_name: str, style: str | None) -> None:
+def _validate_persistent_resolver(name: str, style: str | None) -> None:
     """Reject calls with missing or event-only ``style`` for persistent resolvers."""
     if style is None:
-        raise ValueError(f"{resolver_name} requires a persistent style.")
-    if style in EVENT_ONLY_STYLES:
-        raise ValueError(f"{resolver_name} cannot be used with event-only style {style!r}.")
-    if style not in PERSISTENT_STYLES:
-        column_for_style(style)
+        raise ValueError(f"{name} requires a persistent style.")
+    if column_for_style(style) != LANGUAGE_PERSISTENT:
+        raise ValueError(f"{name} cannot be used with event-only style {style!r}.")
 
 
 def _matching_rows(
@@ -509,29 +460,6 @@ def _matching_rows(
     ]
 
 
-def _select_latest(
-    rows: Sequence[LanguageRow],
-    *,
-    style: str | None,
-    role: str | None,
-    tool_name: str | None,
-    camera: str | None,
-) -> LanguageRow | None:
-    """Return the row tied for the latest ``timestamp`` (disambiguated by selectors)."""
-    if not rows:
-        return None
-    rows = sorted(rows, key=_persistent_sort_key)
-    latest_ts = _timestamp(rows[-1])
-    return _select_one(
-        [row for row in rows if _timestamp(row) == latest_ts],
-        style=style,
-        role=role,
-        tool_name=tool_name,
-        camera=camera,
-        sort_key=_persistent_sort_key,
-    )
-
-
 def _select_one(
     rows: Sequence[LanguageRow],
     *,
@@ -539,27 +467,38 @@ def _select_one(
     role: str | None,
     tool_name: str | None,
     camera: str | None,
-    sort_key: Any,
 ) -> LanguageRow | None:
-    """Return the single matching row, or raise if the selectors are ambiguous."""
+    """Return the single matching row, or raise if the resolver is ambiguous.
+
+    Multiple matches always raise — even when the caller already passed
+    some selectors — because remaining ambiguity means the data has
+    several rows that look identical to the resolver and the caller
+    needs to pin down a specific one (e.g. add ``camera=...`` for VQA
+    rows shared across cameras).
+    """
     if not rows:
         return None
-    if len(rows) > 1 and role is None and tool_name is None and camera is None:
+    if len(rows) > 1:
         raise ValueError(
-            f"Ambiguous resolver for style={style!r}; add role=..., tool_name=..., "
-            f"or camera=... to disambiguate."
+            f"Ambiguous resolver for style={style!r} role={role!r} "
+            f"tool_name={tool_name!r} camera={camera!r}: {len(rows)} matching rows. "
+            f"Add a selector that distinguishes them."
         )
-    return sorted(rows, key=sort_key)[0]
+    return rows[0]
 
 
-def _persistent_sort_key(row: LanguageRow) -> tuple[float, str, str]:
-    """Sort key for persistent rows: ``(timestamp, style, role)``."""
-    return (_timestamp(row), row.get("style") or "", row.get("role") or "")
+def _row_sort_key(row: LanguageRow) -> tuple[float, str, str]:
+    """Stable sort key for both persistent and event rows.
 
-
-def _event_sort_key(row: LanguageRow) -> tuple[str, str]:
-    """Sort key for event rows: ``(style, role)`` (timestamp is implicit in the frame)."""
-    return (row.get("style") or "", row.get("role") or "")
+    Event rows lack ``timestamp`` (it is implicit in the frame), so default
+    to ``0.0`` — within a single frame all event rows share the same sort
+    bucket and are tiebroken by ``(style, role)``.
+    """
+    timestamp = row.get("timestamp")
+    ts = (
+        float(timestamp.item() if hasattr(timestamp, "item") else timestamp) if timestamp is not None else 0.0
+    )
+    return (ts, row.get("style") or "", row.get("role") or "")
 
 
 def _timestamp(row: LanguageRow) -> float:

src/lerobot/datasets/lerobot_dataset.py

@@ -630,6 +630,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         streaming_encoding: bool = False,
         encoder_queue_maxsize: int = 30,
         encoder_threads: int | None = None,
+        video_files_size_in_mb: int | None = None,
+        data_files_size_in_mb: int | None = None,
     ) -> "LeRobotDataset":
         """Create a new LeRobotDataset from scratch for recording data.
 
@@ -677,6 +679,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
             root=root,
             use_videos=use_videos,
             metadata_buffer_size=metadata_buffer_size,
+            video_files_size_in_mb=video_files_size_in_mb,
+            data_files_size_in_mb=data_files_size_in_mb,
         )
         obj.repo_id = obj.meta.repo_id
         obj._requested_root = obj.meta.root

src/lerobot/datasets/multi_dataset.py

@@ -123,7 +123,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
 
         NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
         """
-        return self._datasets[0].meta.info["fps"]
+        return self._datasets[0].meta.info.fps
 
     @property
     def video(self) -> bool:
@@ -133,7 +133,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
 
         NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
         """
-        return self._datasets[0].meta.info.get("video", False)
+        return len(self._datasets[0].meta.video_keys) > 0
 
     @property
     def features(self) -> datasets.Features:

src/lerobot/datasets/streaming_dataset.py

@@ -434,7 +434,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
 
     def _make_padding_camera_frame(self, camera_key: str):
         """Variable-shape padding frame for given camera keys, given in (H, W, C)"""
-        return torch.zeros(self.meta.info["features"][camera_key]["shape"]).permute(-1, 0, 1)
+        return torch.zeros(self.meta.info.features[camera_key]["shape"]).permute(-1, 0, 1)
 
     def _get_video_frame_padding_mask(
         self,

src/lerobot/datasets/utils.py

@@ -14,9 +14,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import contextlib
+import dataclasses
 import importlib.resources
 import json
 import logging
+from dataclasses import dataclass, field
 from pathlib import Path
 
 import datasets
@@ -70,6 +72,9 @@ class ForwardCompatibilityError(CompatibilityError):
         super().__init__(message)
 
 
+logger = logging.getLogger(__name__)
+
+
 DEFAULT_CHUNK_SIZE = 1000  # Max number of files per chunk
 DEFAULT_DATA_FILE_SIZE_IN_MB = 100  # Max size per file
 DEFAULT_VIDEO_FILE_SIZE_IN_MB = 200  # Max size per file
@@ -93,6 +98,130 @@ LEGACY_EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
 LEGACY_TASKS_PATH = "meta/tasks.jsonl"
 
 
+@dataclass
+class DatasetInfo:
+    """Typed representation of the ``meta/info.json`` file for a LeRobot dataset.
+
+    Replaces the previously untyped ``dict`` returned by ``load_info()`` and
+    created by ``create_empty_dataset_info()``.  Using a dataclass provides
+    explicit field definitions, IDE auto-completion, and validation at
+    construction time.
+    """
+
+    codebase_version: str
+    fps: int
+    features: dict[str, dict]
+
+    # Episode / frame counters — start at zero for new datasets
+    total_episodes: int = 0
+    total_frames: int = 0
+    total_tasks: int = 0
+
+    # Storage settings
+    chunks_size: int = field(default=DEFAULT_CHUNK_SIZE)
+    data_files_size_in_mb: int = field(default=DEFAULT_DATA_FILE_SIZE_IN_MB)
+    video_files_size_in_mb: int = field(default=DEFAULT_VIDEO_FILE_SIZE_IN_MB)
+
+    # File path templates
+    data_path: str = field(default=DEFAULT_DATA_PATH)
+    video_path: str | None = field(default=DEFAULT_VIDEO_PATH)
+
+    # Optional metadata
+    robot_type: str | None = None
+    splits: dict[str, str] = field(default_factory=dict)
+    # OpenAI-style tool schemas declared by the dataset. ``None`` means the
+    # dataset doesn't declare any — readers fall back to ``DEFAULT_TOOLS``.
+    tools: list[dict] | None = None
+
+    def __post_init__(self) -> None:
+        # Coerce feature shapes from list to tuple — JSON deserialisation
+        # returns lists, but the rest of the codebase expects tuples.
+        for ft in self.features.values():
+            if isinstance(ft.get("shape"), list):
+                ft["shape"] = tuple(ft["shape"])
+
+        if self.fps <= 0:
+            raise ValueError(f"fps must be positive, got {self.fps}")
+        if self.chunks_size <= 0:
+            raise ValueError(f"chunks_size must be positive, got {self.chunks_size}")
+        if self.data_files_size_in_mb <= 0:
+            raise ValueError(f"data_files_size_in_mb must be positive, got {self.data_files_size_in_mb}")
+        if self.video_files_size_in_mb <= 0:
+            raise ValueError(f"video_files_size_in_mb must be positive, got {self.video_files_size_in_mb}")
+
+    def to_dict(self) -> dict:
+        """Return a JSON-serialisable dict.
+
+        Converts tuple shapes back to lists so ``json.dump`` can handle them.
+        Drops ``tools`` when unset so existing datasets keep a clean
+        ``info.json``.
+        """
+        d = dataclasses.asdict(self)
+        for ft in d["features"].values():
+            if isinstance(ft.get("shape"), tuple):
+                ft["shape"] = list(ft["shape"])
+        if d.get("tools") is None:
+            d.pop("tools", None)
+        return d
+
+    @classmethod
+    def from_dict(cls, data: dict) -> "DatasetInfo":
+        """Construct from a raw dict (e.g. loaded directly from JSON).
+
+        Unknown keys are ignored for forward compatibility with datasets that
+        carry additional fields (e.g. ``total_videos`` from v2.x). A warning is
+        logged when such fields are present.
+        """
+        known = {f.name for f in dataclasses.fields(cls)}
+        unknown = sorted(k for k in data if k not in known)
+        if unknown:
+            logger.warning(f"Unknown fields in DatasetInfo: {unknown}. These will be ignored.")
+        return cls(**{k: v for k, v in data.items() if k in known})
+
+    # ---------------------------------------------------------------------------
+    # Temporary dict-style compatibility layer
+    # Allows existing ``info["key"]`` call-sites to keep working without changes.
+    # Once all callers have been migrated to attribute access, remove these.
+    # ---------------------------------------------------------------------------
+    def __getitem__(self, key: str):
+        import warnings
+
+        warnings.warn(
+            f"Accessing DatasetInfo with dict-style syntax info['{key}'] is deprecated. "
+            f"Use attribute access info.{key} instead.",
+            DeprecationWarning,
+            stacklevel=2,
+        )
+        try:
+            return getattr(self, key)
+        except AttributeError as err:
+            raise KeyError(key) from err
+
+    def __setitem__(self, key: str, value) -> None:
+        import warnings
+
+        warnings.warn(
+            f"Setting DatasetInfo with dict-style syntax info['{key}'] = ... is deprecated. "
+            f"Use attribute assignment info.{key} = ... instead.",
+            DeprecationWarning,
+            stacklevel=2,
+        )
+        if not hasattr(self, key):
+            raise KeyError(f"DatasetInfo has no field '{key}'")
+        setattr(self, key, value)
+
+    def __contains__(self, key: str) -> bool:
+        """Check if a field exists (dict-like interface)."""
+        return hasattr(self, key)
+
+    def get(self, key: str, default=None):
+        """Get attribute value with default fallback (dict-like interface)."""
+        try:
+            return getattr(self, key)
+        except AttributeError:
+            return default
+
+
 def has_legacy_hub_download_metadata(root: Path) -> bool:
     """Return ``True`` when *root* looks like a legacy Hub ``local_dir`` mirror.
 
@@ -293,7 +422,7 @@ def create_branch(repo_id: str, *, branch: str, repo_type: str | None = None) ->
 
 def create_lerobot_dataset_card(
     tags: list | None = None,
-    dataset_info: dict | None = None,
+    dataset_info: DatasetInfo | None = None,
     **kwargs,
 ) -> DatasetCard:
     """Create a `DatasetCard` for a LeRobot dataset.
@@ -304,7 +433,7 @@ def create_lerobot_dataset_card(
 
     Args:
         tags (list | None): A list of tags to add to the dataset card.
-        dataset_info (dict | None): The dataset's info dictionary, which will
+        dataset_info (DatasetInfo | None): The dataset's info object, which will
             be displayed on the card.
         **kwargs: Additional keyword arguments to populate the card template.
 
@@ -317,7 +446,7 @@ def create_lerobot_dataset_card(
         card_tags += tags
     if dataset_info:
         dataset_structure = "[meta/info.json](meta/info.json):\n"
-        dataset_structure += f"```json\n{json.dumps(dataset_info, indent=4)}\n```\n"
+        dataset_structure += f"```json\n{json.dumps(dataset_info.to_dict(), indent=4)}\n```\n"
         kwargs = {**kwargs, "dataset_structure": dataset_structure}
     card_data = DatasetCardData(
         license=kwargs.get("license"),

src/lerobot/policies/__init__.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from lerobot.utils.action_interpolator import ActionInterpolator as ActionInterpolator
+
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .factory import get_policy_class, make_policy, make_policy_config, make_pre_post_processors
@@ -21,10 +23,7 @@ from .pi0.configuration_pi0 import PI0Config as PI0Config
 from .pi0_fast.configuration_pi0_fast import PI0FastConfig as PI0FastConfig
 from .pi05.configuration_pi05 import PI05Config as PI05Config
 from .pretrained import PreTrainedPolicy as PreTrainedPolicy
-from .rtc import ActionInterpolator as ActionInterpolator
 from .sac.configuration_sac import SACConfig as SACConfig
-from .sac.reward_model.configuration_classifier import RewardClassifierConfig as RewardClassifierConfig
-from .sarm.configuration_sarm import SARMConfig as SARMConfig
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .utils import make_robot_action, prepare_observation_for_inference
@@ -45,9 +44,7 @@ __all__ = [
     "PI0Config",
     "PI0FastConfig",
     "PI05Config",
-    "RewardClassifierConfig",
     "SACConfig",
-    "SARMConfig",
     "SmolVLAConfig",
     "TDMPCConfig",
     "VQBeTConfig",

src/lerobot/policies/act/modeling_act.py

@@ -142,9 +142,10 @@ class ACTPolicy(PreTrainedPolicy):
 
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
 
-        l1_loss = (
-            F.l1_loss(batch[ACTION], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
-        ).mean()
+        abs_err = F.l1_loss(batch[ACTION], actions_hat, reduction="none")
+        valid_mask = ~batch["action_is_pad"].unsqueeze(-1)
+        num_valid = valid_mask.sum() * abs_err.shape[-1]
+        l1_loss = (abs_err * valid_mask).sum() / num_valid.clamp_min(1)
 
         loss_dict = {"l1_loss": l1_loss.item()}
         if self.config.use_vae:

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -380,7 +380,9 @@ class DiffusionModel(nn.Module):
                     f"{self.config.do_mask_loss_for_padding=}."
                 )
             in_episode_bound = ~batch["action_is_pad"]
-            loss = loss * in_episode_bound.unsqueeze(-1)
+            mask = in_episode_bound.unsqueeze(-1)
+            num_valid = mask.sum() * loss.shape[-1]
+            return (loss * mask).sum() / num_valid.clamp_min(1)
 
         return loss.mean()
 

src/lerobot/policies/factory.py

@@ -52,8 +52,6 @@ from .pi0.configuration_pi0 import PI0Config
 from .pi05.configuration_pi05 import PI05Config
 from .pretrained import PreTrainedPolicy
 from .sac.configuration_sac import SACConfig
-from .sac.reward_model.configuration_classifier import RewardClassifierConfig
-from .sarm.configuration_sarm import SARMConfig
 from .smolvla.configuration_smolvla import SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig
 from .utils import validate_visual_features_consistency
@@ -89,7 +87,7 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
 
     Args:
         name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
-            "multi_task_dit", "vqbet", "pi0", "pi05", "sac", "reward_classifier", "smolvla", "wall_x".
+            "multi_task_dit", "vqbet", "pi0", "pi05", "sac", "smolvla", "wall_x".
     Returns:
         The policy class corresponding to the given name.
 
@@ -132,18 +130,10 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from .sac.modeling_sac import SACPolicy
 
         return SACPolicy
-    elif name == "reward_classifier":
-        from .sac.reward_model.modeling_classifier import Classifier
-
-        return Classifier
     elif name == "smolvla":
         from .smolvla.modeling_smolvla import SmolVLAPolicy
 
         return SmolVLAPolicy
-    elif name == "sarm":
-        from .sarm.modeling_sarm import SARMRewardModel
-
-        return SARMRewardModel
     elif name == "groot":
         from .groot.modeling_groot import GrootPolicy
 
@@ -173,7 +163,7 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
     Args:
         policy_type: The type of the policy. Supported types include "tdmpc",
                      "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05", "sac",
-                     "smolvla", "reward_classifier", "wall_x".
+                     "smolvla", "wall_x".
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
     Returns:
@@ -200,8 +190,6 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return SACConfig(**kwargs)
     elif policy_type == "smolvla":
         return SmolVLAConfig(**kwargs)
-    elif policy_type == "reward_classifier":
-        return RewardClassifierConfig(**kwargs)
     elif policy_type == "groot":
         return GrootConfig(**kwargs)
     elif policy_type == "xvla":
@@ -378,14 +366,6 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
-    elif isinstance(policy_cfg, RewardClassifierConfig):
-        from .sac.reward_model.processor_classifier import make_classifier_processor
-
-        processors = make_classifier_processor(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
     elif isinstance(policy_cfg, SmolVLAConfig):
         from .smolvla.processor_smolvla import make_smolvla_pre_post_processors
 
@@ -394,14 +374,6 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
-    elif isinstance(policy_cfg, SARMConfig):
-        from .sarm.processor_sarm import make_sarm_pre_post_processors
-
-        processors = make_sarm_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-            dataset_meta=kwargs.get("dataset_meta"),
-        )
     elif isinstance(policy_cfg, GrootConfig):
         from .groot.processor_groot import make_groot_pre_post_processors
 
@@ -542,7 +514,7 @@ def make_policy(
 
         logging.info("Loading policy's PEFT adapter.")
 
-        peft_pretrained_path = cfg.pretrained_path
+        peft_pretrained_path = str(cfg.pretrained_path)
         peft_config = PeftConfig.from_pretrained(peft_pretrained_path)
 
         kwargs["pretrained_name_or_path"] = peft_config.base_model_name_or_path
@@ -555,7 +527,9 @@ def make_policy(
             )
 
         policy = policy_cls.from_pretrained(**kwargs)
-        policy = PeftModel.from_pretrained(policy, peft_pretrained_path, config=peft_config)
+        policy = PeftModel.from_pretrained(
+            policy, peft_pretrained_path, config=peft_config, is_trainable=True
+        )
 
     else:
         # Make a fresh policy.

src/lerobot/policies/groot/action_head/flow_matching_action_head.py

@@ -13,7 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from dataclasses import dataclass, field
+from dataclasses import field
 from typing import TYPE_CHECKING
 
 import torch
@@ -109,7 +109,6 @@ class MultiEmbodimentActionEncoder(nn.Module):
         return x
 
 
-@dataclass
 class FlowmatchingActionHeadConfig(PretrainedConfig):
     """NOTE: N1.5 uses XEmbFlowmatchingPolicyHeadConfig as action head"""
 

src/lerobot/policies/groot/groot_n1.py

@@ -13,7 +13,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from dataclasses import dataclass, field
 from pathlib import Path
 from typing import TYPE_CHECKING
 
@@ -174,17 +173,14 @@ N_COLOR_CHANNELS = 3
 
 
 # config
-@dataclass
 class GR00TN15Config(PretrainedConfig):
     model_type = "gr00t_n1_5"
-    backbone_cfg: dict = field(init=False, metadata={"help": "Backbone configuration."})
 
-    action_head_cfg: dict = field(init=False, metadata={"help": "Action head configuration."})
-
-    action_horizon: int = field(init=False, metadata={"help": "Action horizon."})
-
-    action_dim: int = field(init=False, metadata={"help": "Action dimension."})
-    compute_dtype: str = field(default="float32", metadata={"help": "Compute dtype."})
+    backbone_cfg: dict
+    action_head_cfg: dict
+    action_horizon: int
+    action_dim: int
+    compute_dtype: str = "float32"
 
     def __init__(self, **kwargs):
         super().__init__(**kwargs)

src/lerobot/policies/multi_task_dit/modeling_multi_task_dit.py

@@ -688,8 +688,9 @@ class DiffusionObjective(nn.Module):
         loss = F.mse_loss(predicted, target, reduction="none")
 
         if self.do_mask_loss_for_padding and "action_is_pad" in batch:
-            valid_actions = ~batch["action_is_pad"]
-            loss = loss * valid_actions.unsqueeze(-1)
+            mask = ~batch["action_is_pad"].unsqueeze(-1)
+            num_valid = mask.sum() * loss.shape[-1]
+            return (loss * mask).sum() / num_valid.clamp_min(1)
 
         return loss.mean()
 
@@ -752,8 +753,9 @@ class FlowMatchingObjective(nn.Module):
         loss = F.mse_loss(predicted_velocity, target_velocity, reduction="none")
 
         if self.do_mask_loss_for_padding and "action_is_pad" in batch:
-            valid_mask = ~batch["action_is_pad"]
-            loss = loss * valid_mask.unsqueeze(-1)
+            mask = ~batch["action_is_pad"].unsqueeze(-1)
+            num_valid = mask.sum() * loss.shape[-1]
+            return (loss * mask).sum() / num_valid.clamp_min(1)
 
         return loss.mean()
 

src/lerobot/policies/pi0/modeling_pi0.py

@@ -444,13 +444,13 @@ class PaliGemmaWithExpertModel(
         if image.dtype != torch.float32:
             image = image.to(torch.float32)
         image_outputs = self.paligemma.model.get_image_features(image)
-        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
+        features = image_outputs.pooler_output
         if features.dtype != out_dtype:
             features = features.to(out_dtype)
         return features
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.model.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.get_input_embeddings()(tokens)
 
     def forward(
         self,
@@ -666,8 +666,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         # Process language tokens
         def lang_embed_func(lang_tokens):
             lang_emb = self.paligemma_with_expert.embed_language_tokens(lang_tokens)
-            lang_emb_dim = lang_emb.shape[-1]
-            return lang_emb * math.sqrt(lang_emb_dim)
+            return lang_emb
 
         lang_emb = self._apply_checkpoint(lang_embed_func, lang_tokens)
         embs.append(lang_emb)
@@ -748,16 +747,8 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         return embs, pad_masks, att_masks, adarms_cond
 
-    def forward(
-        self, images, img_masks, lang_tokens, lang_masks, state, actions, noise=None, time=None
-    ) -> Tensor:
+    def forward(self, images, img_masks, lang_tokens, lang_masks, state, actions, noise, time) -> Tensor:
         """Do a full training forward pass and compute the loss."""
-        if noise is None:
-            noise = self.sample_noise(actions.shape, actions.device)
-
-        if time is None:
-            time = self.sample_time(actions.shape[0], actions.device)
-
         time_expanded = time[:, None, None]
         x_t = time_expanded * noise + (1 - time_expanded) * actions
         u_t = noise - actions
@@ -1292,8 +1283,11 @@ class PI0Policy(PreTrainedPolicy):
         state = self.prepare_state(batch)
         actions = self.prepare_action(batch)
 
+        noise = self.model.sample_noise(actions.shape, actions.device)
+        time = self.model.sample_time(actions.shape[0], actions.device)
+
         # Compute loss
-        losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions)
+        losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions, noise, time)
 
         # Truncate losses to actual action dimensions
         original_action_dim = self.config.output_features[ACTION].shape[0]

src/lerobot/policies/pi05/modeling_pi05.py

@@ -728,14 +728,8 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         return embs, pad_masks, att_masks, adarms_cond
 
-    def forward(self, images, img_masks, tokens, masks, actions, noise=None, time=None) -> Tensor:
+    def forward(self, images, img_masks, tokens, masks, actions, noise, time) -> Tensor:
         """Do a full training forward pass and compute the loss."""
-        if noise is None:
-            noise = self.sample_noise(actions.shape, actions.device)
-
-        if time is None:
-            time = self.sample_time(actions.shape[0], actions.device)
-
         time_expanded = time[:, None, None]
         x_t = time_expanded * noise + (1 - time_expanded) * actions
         u_t = noise - actions
@@ -1262,8 +1256,11 @@ class PI05Policy(PreTrainedPolicy):
 
         actions = self.prepare_action(batch)
 
+        noise = self.model.sample_noise(actions.shape, actions.device)
+        time = self.model.sample_time(actions.shape[0], actions.device)
+
         # Compute loss (no separate state needed for PI05)
-        losses = self.model.forward(images, img_masks, tokens, masks, actions)
+        losses = self.model.forward(images, img_masks, tokens, masks, actions, noise, time)
 
         # Truncate losses to actual action dimensions
         original_action_dim = self.config.output_features[ACTION].shape[0]

src/lerobot/policies/pi0_fast/modeling_pi0_fast.py

@@ -16,7 +16,6 @@
 
 import builtins
 import logging
-import math
 from collections import deque
 from pathlib import Path
 from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
@@ -227,6 +226,7 @@ class PI0FastPaliGemma(nn.Module):
         # forward(..., adarms_cond=...) is supported (same as pi0/pi05).
         if use_adarms[0]:
             text_config = self.paligemma.config.text_config
+            del self.paligemma.model.language_model
             self.paligemma.model.language_model = PiGemmaModel(text_config)
 
         self.to_bfloat16_for_selected_params(precision)
@@ -260,13 +260,15 @@ class PI0FastPaliGemma(nn.Module):
         if image.dtype != torch.float32:
             image = image.to(torch.float32)
         image_outputs = self.paligemma.model.get_image_features(image)
-        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
+        features = image_outputs.pooler_output
+        norm = 2048**0.5
+        features = features / norm * norm
         if features.dtype != out_dtype:
             features = features.to(out_dtype)
         return features
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.model.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.get_input_embeddings()(tokens)
 
     def forward(
         self,
@@ -416,8 +418,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
         # Process language instruction tokens
         def lang_embed_func(tokens):
             lang_emb = self.paligemma_with_expert.embed_language_tokens(tokens)
-            lang_emb_dim = lang_emb.shape[-1]
-            return lang_emb * math.sqrt(lang_emb_dim)
+            return lang_emb
 
         lang_emb = self._apply_checkpoint(lang_embed_func, tokens)
         embs.append(lang_emb)
@@ -431,8 +432,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
 
             def fast_action_embed_func(fast_action_tokens):
                 fast_emb = self.paligemma_with_expert.embed_language_tokens(fast_action_tokens)
-                fast_emb_dim = fast_emb.shape[-1]
-                return fast_emb * math.sqrt(fast_emb_dim)
+                return fast_emb
 
             fast_action_emb = self._apply_checkpoint(fast_action_embed_func, fast_action_tokens)
             embs.append(fast_action_emb)
@@ -665,7 +665,6 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
             if t < max_decoding_steps - 1:
                 # embed the newly generated token
                 next_token_emb = self.paligemma_with_expert.embed_language_tokens(next_token)
-                next_token_emb = next_token_emb * math.sqrt(next_token_emb.shape[-1])
                 if prefix_embs.dtype == torch.bfloat16:
                     next_token_emb = next_token_emb.to(dtype=torch.bfloat16)
 
@@ -770,7 +769,6 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
             # Embed the single previous token
             # We use embed_language_tokens directly to avoid overhead of full prefix embedding
             next_token_emb = self.paligemma_with_expert.embed_language_tokens(next_token)
-            next_token_emb = next_token_emb * math.sqrt(next_token_emb.shape[-1])
             if prefix_embs.dtype == torch.bfloat16:
                 next_token_emb = next_token_emb.to(dtype=torch.bfloat16)
 

src/lerobot/policies/pi_gemma.py

@@ -197,6 +197,9 @@ class PiGemmaModel(GemmaModel):  # type: ignore[misc]
 
     def __init__(self, config: GemmaConfig, **kwargs):
         super().__init__(config, **kwargs)
+        # Free parent-allocated layers/norm before replacing to avoid ~2x peak memory.
+        del self.layers
+        del self.norm
         # if not getattr(config, "use_adarms", False):
         #     return
         cond_dim = getattr(config, "adarms_cond_dim", None)
@@ -328,6 +331,7 @@ class PiGemmaForCausalLM(GemmaForCausalLM):  # type: ignore[misc]
 
     def __init__(self, config: GemmaConfig, **kwargs):
         super().__init__(config, **kwargs)
+        del self.model
         self.model = PiGemmaModel(config)
 
 
@@ -336,6 +340,7 @@ class PaliGemmaModelWithPiGemma(PaliGemmaModel):
 
     def __init__(self, config):
         super().__init__(config)
+        del self.language_model
         self.language_model = PiGemmaModel(config.text_config)
 
 
@@ -344,6 +349,7 @@ class PaliGemmaForConditionalGenerationWithPiGemma(PaliGemmaForConditionalGenera
 
     def __init__(self, config):
         super().__init__(config)
+        del self.model
         self.model = PaliGemmaModelWithPiGemma(config)
 
     # Make modules available through conditional class for BC

src/lerobot/policies/rtc/__init__.py

@@ -19,6 +19,7 @@ from .action_queue import ActionQueue
 from .configuration_rtc import RTCConfig
 from .latency_tracker import LatencyTracker
 from .modeling_rtc import RTCProcessor
+from .relative import reanchor_relative_rtc_prefix
 
 __all__ = [
     "ActionInterpolator",
@@ -26,4 +27,5 @@ __all__ = [
     "LatencyTracker",
     "RTCConfig",
     "RTCProcessor",
+    "reanchor_relative_rtc_prefix",
 ]

src/lerobot/policies/rtc/action_interpolator.py

@@ -1,116 +1,4 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
+# Moved to lerobot.utils.action_interpolator — re-exported for backwards compatibility.
+from lerobot.utils.action_interpolator import ActionInterpolator
 
-"""Action interpolation for smoother robot control.
-
-Provides configurable Nx control rate by interpolating between consecutive actions.
-Useful with RTC and action-chunking policies to reduce jerkiness.
-"""
-
-from torch import Tensor
-
-
-class ActionInterpolator:
-    """Interpolates between consecutive actions for smoother control.
-
-    When enabled with multiplier N, produces N actions per policy action
-    by linearly interpolating between the previous and current action.
-
-    Example with multiplier=3:
-        prev_action -> [1/3 interpolated, 2/3 interpolated, current_action]
-
-    This effectively multiplies the control rate for smoother motion.
-
-    Usage:
-        interpolator = ActionInterpolator(multiplier=2)  # 2x control rate
-
-        # In control loop:
-        if interpolator.needs_new_action():
-            new_action = queue.get()
-            if new_action:
-                interpolator.add(new_action.cpu())
-
-        action = interpolator.get()
-        if action:
-            robot.send_action(action)
-    """
-
-    def __init__(self, multiplier: int = 1):
-        """Initialize the interpolator.
-
-        Args:
-            multiplier: Control rate multiplier (1 = no interpolation, 2 = 2x, 3 = 3x, etc.)
-        """
-        if multiplier < 1:
-            raise ValueError(f"multiplier must be >= 1, got {multiplier}")
-        self.multiplier = multiplier
-        self._prev: Tensor | None = None
-        self._buffer: list[Tensor] = []
-        self._idx = 0
-
-    @property
-    def enabled(self) -> bool:
-        """Whether interpolation is active (multiplier > 1)."""
-        return self.multiplier > 1
-
-    def reset(self):
-        """Reset interpolation state (call between episodes)."""
-        self._prev = None
-        self._buffer = []
-        self._idx = 0
-
-    def needs_new_action(self) -> bool:
-        """Check if a new action is needed from the queue."""
-        return self._idx >= len(self._buffer)
-
-    def add(self, action: Tensor) -> None:
-        """Add a new action and compute interpolated sequence.
-
-        Args:
-            action: New action tensor from policy/queue (already on CPU).
-        """
-        if self.multiplier > 1 and self._prev is not None:
-            self._buffer = []
-            for i in range(1, self.multiplier + 1):
-                t = i / self.multiplier
-                interp = self._prev + t * (action - self._prev)
-                self._buffer.append(interp)
-        else:
-            # First step: no previous action yet, so run at base FPS without interpolation.
-            self._buffer = [action.clone()]
-        self._prev = action.clone()
-        self._idx = 0
-
-    def get(self) -> Tensor | None:
-        """Get the next interpolated action.
-
-        Returns:
-            Next action tensor, or None if buffer is exhausted.
-        """
-        if self._idx >= len(self._buffer):
-            return None
-        action = self._buffer[self._idx]
-        self._idx += 1
-        return action
-
-    def get_control_interval(self, fps: float) -> float:
-        """Get the control interval based on interpolation multiplier.
-
-        Args:
-            fps: Base frames per second.
-
-        Returns:
-            Control interval in seconds (divided by multiplier).
-        """
-        return 1.0 / (fps * self.multiplier)
+__all__ = ["ActionInterpolator"]

src/lerobot/policies/rtc/action_queue.py

@@ -92,10 +92,10 @@ class ActionQueue:
         Returns:
             int: Number of unconsumed actions.
         """
-        if self.queue is None:
-            return 0
-        length = len(self.queue)
-        return length - self.last_index
+        with self.lock:
+            if self.queue is None:
+                return 0
+            return len(self.queue) - self.last_index
 
     def empty(self) -> bool:
         """Check if the queue is empty.
@@ -103,11 +103,10 @@ class ActionQueue:
         Returns:
             bool: True if no actions remain, False otherwise.
         """
-        if self.queue is None:
-            return True
-
-        length = len(self.queue)
-        return length - self.last_index <= 0
+        with self.lock:
+            if self.queue is None:
+                return True
+            return len(self.queue) - self.last_index <= 0
 
     def get_action_index(self) -> int:
         """Get the current action consumption index.
@@ -115,7 +114,8 @@ class ActionQueue:
         Returns:
             int: Index of the next action to be consumed.
         """
-        return self.last_index
+        with self.lock:
+            return self.last_index
 
     def get_left_over(self) -> Tensor | None:
         """Get leftover original actions for RTC prev_chunk_left_over.

src/lerobot/policies/rtc/configuration_rtc.py

@@ -35,7 +35,7 @@ class RTCConfig:
     """
 
     # Infrastructure
-    enabled: bool = False
+    enabled: bool = True
 
     # Core RTC settings
     # Todo change to exp

src/lerobot/policies/rtc/relative.py

@@ -0,0 +1,58 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Relative-action helpers for Real-Time Chunking (RTC)."""
+
+from __future__ import annotations
+
+import torch
+
+from lerobot.processor import (
+    NormalizerProcessorStep,
+    RelativeActionsProcessorStep,
+    TransitionKey,
+    create_transition,
+    to_relative_actions,
+)
+
+
+def reanchor_relative_rtc_prefix(
+    prev_actions_absolute: torch.Tensor,
+    current_state: torch.Tensor,
+    relative_step: RelativeActionsProcessorStep,
+    normalizer_step: NormalizerProcessorStep | None,
+    policy_device: torch.device | str,
+) -> torch.Tensor:
+    """Convert absolute leftover actions into model-space for relative-action RTC policies.
+
+    When using relative actions, the RTC prefix (previous chunk's unexecuted tail)
+    is stored in absolute coordinates. Before feeding it back to the policy, this
+    helper re-expresses those actions relative to the robot's current joint state
+    and optionally normalizes them so the policy receives correctly scaled inputs.
+    """
+    state = current_state.detach().cpu()
+    if state.dim() == 1:
+        state = state.unsqueeze(0)
+
+    action_cpu = prev_actions_absolute.detach().cpu()
+    mask = relative_step._build_mask(action_cpu.shape[-1])
+    relative_actions = to_relative_actions(action_cpu, state, mask)
+
+    transition = create_transition(action=relative_actions)
+    if normalizer_step is not None:
+        transition = normalizer_step(transition)
+
+    return transition[TransitionKey.ACTION].to(policy_device)

src/lerobot/policies/sarm/README.md

@@ -1 +0,0 @@
-../../../../docs/source/policy_sarm_README.md

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -394,13 +394,21 @@ class SmolVLAPolicy(PreTrainedPolicy):
         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
-            per_sample_loss = losses.mean(dim=(1, 2))
+            # Return per-sample losses (B,) by averaging over valid (time, action) entries
+            if actions_is_pad is None:
+                per_sample_loss = losses.mean(dim=(1, 2))
+            else:
+                num_valid = ((~actions_is_pad).sum(dim=1) * losses.shape[-1]).clamp_min(1)
+                per_sample_loss = losses.sum(dim=(1, 2)) / num_valid
             loss_dict["loss"] = per_sample_loss.mean().item()
             return per_sample_loss, loss_dict
         else:
-            # Default: return scalar mean loss
-            loss = losses.mean()
+            # Default: return scalar mean loss over valid (time, action) entries
+            if actions_is_pad is None:
+                loss = losses.mean()
+            else:
+                num_valid = ((~actions_is_pad).sum() * losses.shape[-1]).clamp_min(1)
+                loss = losses.sum() / num_valid
             loss_dict["loss"] = loss.item()
             return loss, loss_dict
 

src/lerobot/policies/wall_x/qwen_model/qwen2_5_vl_moe.py

@@ -22,7 +22,7 @@ from transformers.utils import (
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
     is_flash_attn_2_available,
-    is_flash_attn_greater_or_equal_2_10,
+    is_flash_attn_greater_or_equal,
     is_torchdynamo_compiling,
     logging,
     replace_return_docstrings,
@@ -890,7 +890,7 @@ class Qwen2_5_VLFlashAttention2(Qwen2_5_VLAttention):
         # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
         # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
         # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
-        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal("2.1.0")
 
     def forward(
         self,

src/lerobot/policies/xvla/modeling_florence2.py

@@ -45,7 +45,7 @@ from transformers.utils import (
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
     is_flash_attn_2_available,
-    is_flash_attn_greater_or_equal_2_10,
+    is_flash_attn_greater_or_equal,
     logging,
     replace_return_docstrings,
 )
@@ -909,7 +909,7 @@ class Florence2FlashAttention2(Florence2Attention):
         # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
         # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
         # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
-        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal("2.1.0")
 
     def _reshape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
         return tensor.view(bsz, seq_len, self.num_heads, self.head_dim)

src/lerobot/processor/__init__.py

@@ -93,9 +93,15 @@ from .relative_action_processor import (
     to_relative_actions,
 )
 from .rename_processor import RenameObservationsProcessorStep, rename_stats
-from .render_messages_processor import RenderMessagesStep
 from .tokenizer_processor import ActionTokenizerProcessorStep, TokenizerProcessorStep
 
+# RenderMessagesStep is intentionally NOT re-exported here: it pulls in
+# `lerobot.datasets.language`, which requires the `[dataset]` extra
+# (`datasets`, `pyarrow`). Importing it from the processor package would
+# break every base-install consumer of `lerobot.processor`. Users that
+# need it import directly:
+#   from lerobot.processor.render_messages_processor import RenderMessagesStep
+
 __all__ = [
     "ActionProcessorStep",
     "AddTeleopActionAsComplimentaryDataStep",
@@ -129,7 +135,6 @@ __all__ = [
     "make_default_robot_observation_processor",
     "AbsoluteActionsProcessorStep",
     "RelativeActionsProcessorStep",
-    "RenderMessagesStep",
     "MapDeltaActionToRobotActionStep",
     "MapTensorToDeltaActionDictStep",
     "NewLineTaskProcessorStep",

src/lerobot/processor/hil_processor.py

@@ -557,7 +557,7 @@ class RewardClassifierProcessorStep(ProcessorStep):
     def __post_init__(self):
         """Initializes the reward classifier model after the dataclass is created."""
         if self.pretrained_path is not None:
-            from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
+            from lerobot.rewards.classifier.modeling_classifier import Classifier
 
             self.reward_classifier = Classifier.from_pretrained(self.pretrained_path)
             self.reward_classifier.to(self.device)

src/lerobot/processor/relative_action_processor.py

@@ -142,6 +142,10 @@ class RelativeActionsProcessorStep(ProcessorStep):
         new_transition[TransitionKey.ACTION] = to_relative_actions(action, state, mask)
         return new_transition
 
+    def get_cached_state(self) -> torch.Tensor | None:
+        """Return the cached ``observation.state`` used as the reference point for relative/absolute action conversions."""
+        return self._last_state
+
     def get_config(self) -> dict[str, Any]:
         return {
             "enabled": self.enabled,
@@ -182,7 +186,8 @@ class AbsoluteActionsProcessorStep(ProcessorStep):
                 "but relative_step is None. Ensure relative_step is set when constructing the postprocessor."
             )
 
-        if self.relative_step._last_state is None:
+        cached_state = self.relative_step.get_cached_state()
+        if cached_state is None:
             raise RuntimeError(
                 "AbsoluteActionsProcessorStep requires state from RelativeActionsProcessorStep "
                 "but no state has been cached. Ensure the preprocessor runs before the postprocessor."
@@ -194,9 +199,7 @@ class AbsoluteActionsProcessorStep(ProcessorStep):
             return new_transition
 
         mask = self.relative_step._build_mask(action.shape[-1])
-        new_transition[TransitionKey.ACTION] = to_absolute_actions(
-            action, self.relative_step._last_state, mask
-        )
+        new_transition[TransitionKey.ACTION] = to_absolute_actions(action, cached_state, mask)
         return new_transition
 
     def get_config(self) -> dict[str, Any]:

src/lerobot/rewards/__init__.py

@@ -0,0 +1,36 @@
+# 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 .classifier.configuration_classifier import RewardClassifierConfig as RewardClassifierConfig
+from .factory import (
+    get_reward_model_class as get_reward_model_class,
+    make_reward_model as make_reward_model,
+    make_reward_model_config as make_reward_model_config,
+    make_reward_pre_post_processors as make_reward_pre_post_processors,
+)
+from .pretrained import PreTrainedRewardModel as PreTrainedRewardModel
+from .sarm.configuration_sarm import SARMConfig as SARMConfig
+
+__all__ = [
+    # Configuration classes
+    "RewardClassifierConfig",
+    "SARMConfig",
+    # Base class
+    "PreTrainedRewardModel",
+    # Factory functions
+    "get_reward_model_class",
+    "make_reward_model",
+    "make_reward_model_config",
+    "make_reward_pre_post_processors",
+]

src/lerobot/rewards/classifier/configuration_classifier.py

@@ -1,5 +1,3 @@
-# !/usr/bin/env python
-
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -15,14 +13,15 @@
 # limitations under the License.
 from dataclasses import dataclass, field
 
-from lerobot.configs import NormalizationMode, PreTrainedConfig
+from lerobot.configs import NormalizationMode
+from lerobot.configs.rewards import RewardModelConfig
 from lerobot.optim import AdamWConfig, LRSchedulerConfig, OptimizerConfig
 from lerobot.utils.constants import OBS_IMAGE
 
 
-@PreTrainedConfig.register_subclass(name="reward_classifier")
+@RewardModelConfig.register_subclass(name="reward_classifier")
 @dataclass
-class RewardClassifierConfig(PreTrainedConfig):
+class RewardClassifierConfig(RewardModelConfig):
     """Configuration for the Reward Classifier model."""
 
     name: str = "reward_classifier"

src/lerobot/rewards/classifier/modeling_classifier.py

@@ -1,5 +1,3 @@
-# !/usr/bin/env python
-
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -19,11 +17,10 @@ import logging
 import torch
 from torch import Tensor, nn
 
+from lerobot.rewards.classifier.configuration_classifier import RewardClassifierConfig
+from lerobot.rewards.pretrained import PreTrainedRewardModel
 from lerobot.utils.constants import OBS_IMAGE, REWARD
 
-from ...pretrained import PreTrainedPolicy
-from .configuration_classifier import RewardClassifierConfig
-
 
 class ClassifierOutput:
     """Wrapper for classifier outputs with additional metadata."""
@@ -99,7 +96,7 @@ class SpatialLearnedEmbeddings(nn.Module):
         return output
 
 
-class Classifier(PreTrainedPolicy):
+class Classifier(PreTrainedRewardModel):
     """Image classifier built on top of a pre-trained encoder."""
 
     name = "reward_classifier"
@@ -235,6 +232,16 @@ class Classifier(PreTrainedPolicy):
 
         return ClassifierOutput(logits=logits, probabilities=probabilities, hidden_states=encoder_outputs)
 
+    def compute_reward(self, batch: dict[str, Tensor]) -> Tensor:
+        """Returns 1.0 for success, 0.0 for failure based on image observations."""
+        images = [batch[key] for key in self.config.input_features if key.startswith(OBS_IMAGE)]
+        output = self.predict(images)
+
+        if self.config.num_classes == 2:
+            return (output.probabilities > 0.5).float()
+        else:
+            return torch.argmax(output.probabilities, dim=1).float()
+
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict[str, Tensor]]:
         """Standard forward pass for training compatible with train.py."""
         # Extract images and labels
@@ -269,10 +276,6 @@ class Classifier(PreTrainedPolicy):
 
     def predict_reward(self, batch, threshold=0.5):
         """Eval method. Returns predicted reward with the decision threshold as argument."""
-        # Check for both OBS_IMAGE and OBS_IMAGES prefixes
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-
         # Extract images from batch dict
         images = [batch[key] for key in self.config.input_features if key.startswith(OBS_IMAGE)]
 
@@ -282,28 +285,3 @@ class Classifier(PreTrainedPolicy):
             return (probs > threshold).float()
         else:
             return torch.argmax(self.predict(images).probabilities, dim=1)
-
-    def get_optim_params(self):
-        """Return optimizer parameters for the policy."""
-        return self.parameters()
-
-    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
-        """
-        This method is required by PreTrainedPolicy but not used for reward classifiers.
-        The reward classifier is not an actor and does not select actions.
-        """
-        raise NotImplementedError("Reward classifiers do not select actions")
-
-    def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
-        """
-        This method is required by PreTrainedPolicy but not used for reward classifiers.
-        The reward classifier is not an actor and does not produce action chunks.
-        """
-        raise NotImplementedError("Reward classifiers do not predict action chunks")
-
-    def reset(self):
-        """
-        This method is required by PreTrainedPolicy but not used for reward classifiers.
-        The reward classifier is not an actor and does not select actions.
-        """
-        pass

src/lerobot/rewards/classifier/processor_classifier.py

@@ -1,5 +1,3 @@
-# !/usr/bin/env python
-
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -27,8 +25,7 @@ from lerobot.processor import (
     policy_action_to_transition,
     transition_to_policy_action,
 )
-
-from .configuration_classifier import RewardClassifierConfig
+from lerobot.rewards.classifier.configuration_classifier import RewardClassifierConfig
 
 
 def make_classifier_processor(
@@ -52,8 +49,6 @@ def make_classifier_processor(
     Args:
         config: The configuration object for the RewardClassifier.
         dataset_stats: A dictionary of statistics for normalization.
-        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
-        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
 
     Returns:
         A tuple containing the configured pre-processor and post-processor pipelines.

src/lerobot/rewards/factory.py

@@ -0,0 +1,238 @@
+#!/usr/bin/env python
+
+# 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.
+
+import importlib
+import logging
+from typing import Any
+
+import torch
+
+from lerobot.configs.rewards import RewardModelConfig
+from lerobot.processor import PolicyAction, PolicyProcessorPipeline
+from lerobot.rewards.classifier.configuration_classifier import RewardClassifierConfig
+from lerobot.rewards.pretrained import PreTrainedRewardModel
+from lerobot.rewards.sarm.configuration_sarm import SARMConfig
+
+
+def get_reward_model_class(name: str) -> type[PreTrainedRewardModel]:
+    """
+    Retrieves a reward model class by its registered name.
+
+    This function uses dynamic imports to avoid loading all reward model classes into
+    memory at once, improving startup time and reducing dependencies.
+
+    Args:
+        name: The name of the reward model. Supported names are "reward_classifier",
+              "sarm".
+
+    Returns:
+        The reward model class corresponding to the given name.
+
+    Raises:
+        ValueError: If the reward model name is not recognized.
+    """
+    if name == "reward_classifier":
+        from lerobot.rewards.classifier.modeling_classifier import Classifier
+
+        return Classifier
+    elif name == "sarm":
+        from lerobot.rewards.sarm.modeling_sarm import SARMRewardModel
+
+        return SARMRewardModel
+    else:
+        try:
+            return _get_reward_model_cls_from_name(name=name)
+        except Exception as e:
+            raise ValueError(f"Reward model type '{name}' is not available.") from e
+
+
+def make_reward_model_config(reward_type: str, **kwargs) -> RewardModelConfig:
+    """
+    Instantiates a reward model configuration object based on the reward type.
+
+    This factory function simplifies the creation of reward model configuration objects
+    by mapping a string identifier to the corresponding config class.
+
+    Args:
+        reward_type: The type of the reward model. Supported types include
+                     "reward_classifier", "sarm".
+        **kwargs: Keyword arguments to be passed to the configuration class constructor.
+
+    Returns:
+        An instance of a `RewardModelConfig` subclass.
+
+    Raises:
+        ValueError: If the `reward_type` is not recognized.
+    """
+    if reward_type == "reward_classifier":
+        return RewardClassifierConfig(**kwargs)
+    elif reward_type == "sarm":
+        return SARMConfig(**kwargs)
+    else:
+        try:
+            config_cls = RewardModelConfig.get_choice_class(reward_type)
+            return config_cls(**kwargs)
+        except Exception as e:
+            raise ValueError(f"Reward model type '{reward_type}' is not available.") from e
+
+
+def make_reward_model(cfg: RewardModelConfig, **kwargs) -> PreTrainedRewardModel:
+    """
+    Instantiate a reward model from its configuration.
+
+    Args:
+        cfg: The configuration for the reward model to be created. If
+             `cfg.pretrained_path` is set, the model will be loaded with weights
+             from that path.
+        **kwargs: Additional keyword arguments forwarded to the model constructor
+            (e.g., ``dataset_stats``, ``dataset_meta``).
+
+    Returns:
+        An instantiated and device-placed reward model.
+    """
+    reward_cls = get_reward_model_class(cfg.type)
+
+    kwargs["config"] = cfg
+
+    if cfg.pretrained_path:
+        kwargs["pretrained_name_or_path"] = cfg.pretrained_path
+        reward_model = reward_cls.from_pretrained(**kwargs)
+    else:
+        reward_model = reward_cls(**kwargs)
+
+    reward_model.to(cfg.device)
+    assert isinstance(reward_model, torch.nn.Module)
+
+    return reward_model
+
+
+def make_reward_pre_post_processors(
+    reward_cfg: RewardModelConfig,
+    **kwargs,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Create pre- and post-processor pipelines for a given reward model.
+
+    Each reward model type has a dedicated factory function for its processors.
+
+    Args:
+        reward_cfg: The configuration of the reward model for which to create processors.
+        **kwargs: Additional keyword arguments passed to the processor factory
+            (e.g., ``dataset_stats``, ``dataset_meta``).
+
+    Returns:
+        A tuple containing the input (pre-processor) and output (post-processor) pipelines.
+
+    Raises:
+        ValueError: If a processor factory is not implemented for the given reward
+            model configuration type.
+    """
+    # Create a new processor based on reward model type
+    if isinstance(reward_cfg, RewardClassifierConfig):
+        from lerobot.rewards.classifier.processor_classifier import make_classifier_processor
+
+        return make_classifier_processor(
+            config=reward_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
+    elif isinstance(reward_cfg, SARMConfig):
+        from lerobot.rewards.sarm.processor_sarm import make_sarm_pre_post_processors
+
+        return make_sarm_pre_post_processors(
+            config=reward_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            dataset_meta=kwargs.get("dataset_meta"),
+        )
+
+    else:
+        try:
+            processors = _make_processors_from_reward_model_config(
+                config=reward_cfg,
+                dataset_stats=kwargs.get("dataset_stats"),
+            )
+        except Exception as e:
+            raise ValueError(
+                f"Processor for reward model type '{reward_cfg.type}' is not implemented."
+            ) from e
+        return processors
+
+
+def _get_reward_model_cls_from_name(name: str) -> type[PreTrainedRewardModel]:
+    """Get reward model class from its registered name using dynamic imports.
+
+    This is used as a helper function to import reward models from 3rd party lerobot
+    plugins.
+
+    Args:
+        name: The name of the reward model.
+
+    Returns:
+        The reward model class corresponding to the given name.
+    """
+    if name not in RewardModelConfig.get_known_choices():
+        raise ValueError(
+            f"Unknown reward model name '{name}'. "
+            f"Available reward models: {RewardModelConfig.get_known_choices()}"
+        )
+
+    config_cls = RewardModelConfig.get_choice_class(name)
+    config_cls_name = config_cls.__name__
+
+    model_name = config_cls_name.removesuffix("Config")
+    if model_name == config_cls_name:
+        raise ValueError(
+            f"The config class name '{config_cls_name}' does not follow the expected naming convention. "
+            f"Make sure it ends with 'Config'!"
+        )
+
+    cls_name = model_name + "RewardModel"
+    module_path = config_cls.__module__.replace("configuration_", "modeling_")
+
+    module = importlib.import_module(module_path)
+    reward_cls = getattr(module, cls_name)
+    return reward_cls
+
+
+def _make_processors_from_reward_model_config(
+    config: RewardModelConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[Any, Any]:
+    """Create pre- and post-processors from a reward model configuration using dynamic imports.
+
+    This is used as a helper function to import processor factories from 3rd party
+    lerobot reward model plugins.
+
+    Args:
+        config: The reward model configuration object.
+        dataset_stats: Dataset statistics for normalization.
+
+    Returns:
+        A tuple containing the input (pre-processor) and output (post-processor) pipelines.
+    """
+    reward_type = config.type
+    function_name = f"make_{reward_type}_pre_post_processors"
+    module_path = config.__class__.__module__.replace("configuration_", "processor_")
+    logging.debug(
+        f"Instantiating reward pre/post processors using function '{function_name}' "
+        f"from module '{module_path}'"
+    )
+    module = importlib.import_module(module_path)
+    function = getattr(module, function_name)
+    return function(config, dataset_stats=dataset_stats)

src/lerobot/rewards/pretrained.py

@@ -0,0 +1,244 @@
+# 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.
+
+import abc
+import builtins
+import logging
+import os
+from importlib.resources import files
+from pathlib import Path
+from tempfile import TemporaryDirectory
+from typing import TYPE_CHECKING, Any, TypeVar
+
+import packaging
+import safetensors
+from huggingface_hub import HfApi, ModelCard, ModelCardData, hf_hub_download
+from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
+from huggingface_hub.errors import HfHubHTTPError
+from safetensors.torch import load_model as load_model_as_safetensor, save_model as save_model_as_safetensor
+from torch import Tensor, nn
+
+from lerobot.configs.rewards import RewardModelConfig
+from lerobot.utils.hub import HubMixin
+
+if TYPE_CHECKING:
+    from lerobot.configs.train import TrainPipelineConfig
+
+T = TypeVar("T", bound="PreTrainedRewardModel")
+
+
+class PreTrainedRewardModel(nn.Module, HubMixin, abc.ABC):
+    """Base class for reward models."""
+
+    config_class: None
+    name: None
+
+    def __init__(self, config: RewardModelConfig, *inputs, **kwargs):
+        super().__init__()
+        if not isinstance(config, RewardModelConfig):
+            raise ValueError(
+                f"Parameter config in `{self.__class__.__name__}(config)` should be an instance of class "
+                "`RewardModelConfig`. To create a model from a pretrained model use "
+                f"`model = {self.__class__.__name__}.from_pretrained(PRETRAINED_MODEL_NAME)`"
+            )
+        self.config = config
+
+    def __init_subclass__(cls, **kwargs):
+        super().__init_subclass__(**kwargs)
+        if not getattr(cls, "config_class", None):
+            raise TypeError(f"Class {cls.__name__} must define 'config_class'")
+        if not getattr(cls, "name", None):
+            raise TypeError(f"Class {cls.__name__} must define 'name'")
+
+    def _save_pretrained(self, save_directory: Path) -> None:
+        self.config._save_pretrained(save_directory)
+        model_to_save = self.module if hasattr(self, "module") else self
+        save_model_as_safetensor(model_to_save, str(save_directory / SAFETENSORS_SINGLE_FILE))
+
+    @classmethod
+    def from_pretrained(
+        cls: builtins.type[T],
+        pretrained_name_or_path: str | Path,
+        *,
+        config: RewardModelConfig | None = None,
+        force_download: bool = False,
+        resume_download: bool | None = None,
+        proxies: dict | None = None,
+        token: str | bool | None = None,
+        cache_dir: str | Path | None = None,
+        local_files_only: bool = False,
+        revision: str | None = None,
+        strict: bool = False,
+        **kwargs,
+    ) -> T:
+        """
+        The reward model is set in evaluation mode by default using `reward.eval()` (dropout modules are
+        deactivated). To train it, you should first set it back in training mode with `reward.train()`.
+        """
+        if config is None:
+            config = RewardModelConfig.from_pretrained(
+                pretrained_name_or_path=pretrained_name_or_path,
+                force_download=force_download,
+                resume_download=resume_download,
+                proxies=proxies,
+                token=token,
+                cache_dir=cache_dir,
+                local_files_only=local_files_only,
+                revision=revision,
+                **kwargs,
+            )
+        model_id = str(pretrained_name_or_path)
+        instance = cls(config, **kwargs)
+        if os.path.isdir(model_id):
+            print("Loading weights from local directory")
+            model_file = os.path.join(model_id, SAFETENSORS_SINGLE_FILE)
+            reward = cls._load_as_safetensor(instance, model_file, config.device or "cpu", strict)
+        else:
+            try:
+                model_file = hf_hub_download(
+                    repo_id=model_id,
+                    filename=SAFETENSORS_SINGLE_FILE,
+                    revision=revision,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    resume_download=resume_download,
+                    token=token,
+                    local_files_only=local_files_only,
+                )
+                reward = cls._load_as_safetensor(instance, model_file, config.device or "cpu", strict)
+            except HfHubHTTPError as e:
+                raise FileNotFoundError(
+                    f"{SAFETENSORS_SINGLE_FILE} not found on the HuggingFace Hub in {model_id}"
+                ) from e
+
+        reward.to(config.device)
+        reward.eval()
+        return reward
+
+    @classmethod
+    def _load_as_safetensor(cls, model: T, model_file: str, map_location: str, strict: bool) -> T:
+        # Create base kwargs
+        kwargs = {"strict": strict}
+
+        # Add device parameter for newer versions that support it
+        if packaging.version.parse(safetensors.__version__) >= packaging.version.parse("0.4.3"):
+            kwargs["device"] = map_location
+
+        # Load the model with appropriate kwargs
+        missing_keys, unexpected_keys = load_model_as_safetensor(model, model_file, **kwargs)
+        if missing_keys:
+            logging.warning(f"Missing key(s) when loading model: {missing_keys}")
+        if unexpected_keys:
+            logging.warning(f"Unexpected key(s) when loading model: {unexpected_keys}")
+
+        # For older versions, manually move to device if needed
+        if "device" not in kwargs and map_location != "cpu":
+            logging.warning(
+                "Loading model weights on other devices than 'cpu' is not supported natively in your version of safetensors."
+                " This means that the model is loaded on 'cpu' first and then copied to the device."
+                " This leads to a slower loading time."
+                " Please update safetensors to version 0.4.3 or above for improved performance."
+            )
+            model.to(map_location)
+        return model
+
+    def get_optim_params(self):
+        """
+        Returns the reward-model-specific parameters dict to be passed on to the optimizer.
+        """
+        return self.parameters()
+
+    def reset(self) -> None:
+        """Reset any internal state."""
+        pass
+
+    @abc.abstractmethod
+    def compute_reward(self, batch: dict[str, Tensor]) -> Tensor:
+        """Compute a scalar reward signal for a batch of observations.
+
+        Args:
+            batch: Dictionary containing at minimum observation tensors.
+                   May also contain "action", "next_observation.*", etc.
+
+        Returns:
+            Tensor of shape ``(batch_size,)`` with reward values.
+        """
+        ...
+
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict[str, Any]]:
+        """Training forward pass — override for trainable reward models."""
+        raise NotImplementedError(
+            f"{self.__class__.__name__} is not trainable. Only use compute_reward() for inference."
+        )
+
+    @property
+    def is_trainable(self) -> bool:
+        """Whether this reward model can be trained via ``lerobot-train``.
+
+        Trainable reward models override :meth:`forward`; zero-shot models
+        inherit the base implementation that raises ``NotImplementedError``.
+        """
+        return type(self).forward is not PreTrainedRewardModel.forward
+
+    def push_model_to_hub(self, cfg: "TrainPipelineConfig"):
+        api = HfApi()
+        repo_id = api.create_repo(
+            repo_id=self.config.repo_id, private=self.config.private, exist_ok=True
+        ).repo_id
+
+        # Push the files to the repo in a single commit
+        with TemporaryDirectory(ignore_cleanup_errors=True) as tmp:
+            saved_path = Path(tmp) / repo_id
+
+            self.save_pretrained(saved_path)  # Calls _save_pretrained and stores model tensors
+
+            card = self.generate_model_card(
+                cfg.dataset.repo_id, self.config.type, self.config.license, self.config.tags
+            )
+            card.save(str(saved_path / "README.md"))
+
+            cfg.save_pretrained(saved_path)  # Calls _save_pretrained and stores train config
+
+            commit_info = api.upload_folder(
+                repo_id=repo_id,
+                repo_type="model",
+                folder_path=saved_path,
+                commit_message="Upload reward model weights, train config and readme",
+                allow_patterns=["*.safetensors", "*.json", "*.yaml", "*.md"],
+                ignore_patterns=["*.tmp", "*.log"],
+            )
+
+            logging.info(f"Model pushed to {commit_info.repo_url.url}")
+
+    def generate_model_card(
+        self, dataset_repo_id: str, model_type: str, license: str | None, tags: list[str] | None
+    ) -> ModelCard:
+        card_data = ModelCardData(
+            license=license or "apache-2.0",
+            library_name="lerobot",
+            pipeline_tag="robotics",
+            tags=list(set(tags or []).union({"robotics", "lerobot", "reward-model", model_type})),
+            model_name=model_type,
+            datasets=dataset_repo_id,
+        )
+
+        template_card = (
+            files("lerobot.templates")
+            .joinpath("lerobot_rewardmodel_modelcard_template.md")
+            .read_text(encoding="utf-8")
+        )
+        card = ModelCard.from_template(card_data, template_str=template_card)
+        card.validate()
+        return card

src/lerobot/rewards/sarm/__init__.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+# 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.
@@ -14,5 +14,6 @@
 
 from .configuration_sarm import SARMConfig
 from .modeling_sarm import SARMRewardModel
+from .processor_sarm import make_sarm_pre_post_processors
 
-__all__ = ["SARMConfig", "SARMRewardModel"]
+__all__ = ["SARMConfig", "SARMRewardModel", "make_sarm_pre_post_processors"]

src/lerobot/rewards/sarm/compute_rabc_weights.py

@@ -25,18 +25,18 @@ need ~num_frames/30 queries instead of one per frame (~30x speedup).
 
 Usage:
     # Full RA-BC computation with visualizations
-    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
+    python src/lerobot/rewards/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
         --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 \\
+    python src/lerobot/rewards/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
         --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 \\
+    python src/lerobot/rewards/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
         --reward-model-path <USER>/sarm_single_uni4 \\
         --visualize-only \\
@@ -58,10 +58,9 @@ import torch
 from tqdm import tqdm
 
 from lerobot.datasets import LeRobotDataset
-
-from .modeling_sarm import SARMRewardModel
-from .processor_sarm import make_sarm_pre_post_processors
-from .sarm_utils import normalize_stage_tau
+from lerobot.rewards.sarm.modeling_sarm import SARMRewardModel
+from lerobot.rewards.sarm.processor_sarm import make_sarm_pre_post_processors
+from lerobot.rewards.sarm.sarm_utils import normalize_stage_tau
 
 
 def get_reward_model_path_from_parquet(parquet_path: Path) -> str | None:
@@ -713,12 +712,12 @@ def main():
         epilog="""
 Examples:
     # Full RA-BC computation with visualizations
-    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
+    python src/lerobot/rewards/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
         --reward-model-path <USER>/sarm_single_uni4
 
     # Visualize predictions only (no RA-BC computation)
-    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
+    python src/lerobot/rewards/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
         --reward-model-path <USER>/sarm_single_uni4 \\
         --visualize-only \\

src/lerobot/rewards/sarm/configuration_sarm.py

@@ -1,5 +1,3 @@
-#!/usr/bin/env python
-
 # Copyright 2025 Qianzhong Chen, Justin Yu, Mac Schwager, Pieter Abbeel, Yide Shentu, Philipp Wu
 # and The HuggingFace Inc. team. All rights reserved.
 #
@@ -22,14 +20,15 @@ Paper: https://arxiv.org/abs/2509.25358
 
 from dataclasses import dataclass, field
 
-from lerobot.configs import FeatureType, NormalizationMode, PolicyFeature, PreTrainedConfig
+from lerobot.configs import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.configs.rewards import RewardModelConfig
 from lerobot.optim import AdamWConfig, CosineDecayWithWarmupSchedulerConfig
 from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
 
 
-@PreTrainedConfig.register_subclass("sarm")
+@RewardModelConfig.register_subclass("sarm")
 @dataclass
-class SARMConfig(PreTrainedConfig):
+class SARMConfig(RewardModelConfig):
     """Configuration class for SARM (Stage-Aware Reward Modeling).
 
     Supports three annotation modes:
@@ -110,7 +109,6 @@ class SARMConfig(PreTrainedConfig):
 
     def __post_init__(self):
         super().__post_init__()
-
         if self.annotation_mode not in ["single_stage", "dense_only", "dual"]:
             raise ValueError(
                 f"annotation_mode must be 'single_stage', 'dense_only', or 'dual', got {self.annotation_mode}"

src/lerobot/rewards/sarm/modeling_sarm.py

@@ -1,5 +1,3 @@
-#!/usr/bin/env python
-
 # Copyright 2025 Qianzhong Chen, Justin Yu, Mac Schwager, Pieter Abbeel, Yide Shentu, Philipp Wu
 # and The HuggingFace Inc. team. All rights reserved.
 #
@@ -34,14 +32,13 @@ import torch.nn as nn
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 
-from lerobot.utils.constants import OBS_STR
-
-from ..pretrained import PreTrainedPolicy
-from .configuration_sarm import SARMConfig
-from .sarm_utils import (
+from lerobot.rewards.pretrained import PreTrainedRewardModel
+from lerobot.rewards.sarm.configuration_sarm import SARMConfig
+from lerobot.rewards.sarm.sarm_utils import (
     normalize_stage_tau,
     pad_state_to_max_dim,
 )
+from lerobot.utils.constants import OBS_STR
 
 
 class StageTransformer(nn.Module):
@@ -353,7 +350,7 @@ def gen_stage_emb(num_classes: int, targets: torch.Tensor) -> torch.Tensor:
     return stage_onehot
 
 
-class SARMRewardModel(PreTrainedPolicy):
+class SARMRewardModel(PreTrainedRewardModel):
     """
     SARM Reward Model for stage-aware task completion rewards.
 
@@ -471,6 +468,23 @@ class SARMRewardModel(PreTrainedPolicy):
         self.subtask_model.to(device)
         return self
 
+    def compute_reward(self, batch: dict[str, Tensor]) -> Tensor:
+        """Compute dense progress reward in [0, 1] from batch.
+
+        Expects batch to contain:
+        - "observation_features" or video embeddings: (B, T, 512)
+        - "language_embedding" or text embeddings: (B, 512)
+        - optionally "observation.state": (B, T, state_dim)
+        """
+        text_emb = batch.get("language_embedding", batch.get("text_features"))
+        video_emb = batch.get("observation_features", batch.get("video_features"))
+        state = batch.get("observation.state", batch.get("state_features"))
+
+        rewards = self.calculate_rewards(text_emb, video_emb, state)
+        if isinstance(rewards, np.ndarray):
+            rewards = torch.from_numpy(rewards).float()
+        return rewards
+
     @torch.no_grad()
     def calculate_rewards(
         self,
@@ -631,17 +645,9 @@ class SARMRewardModel(PreTrainedPolicy):
         return self.parameters()
 
     def reset(self):
-        """Required by PreTrainedPolicy but not used for reward models."""
+        """SARM has no episode-level state to reset."""
         pass
 
-    def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
-        """Required by PreTrainedPolicy but not used for reward models."""
-        raise NotImplementedError("SARM model does not predict action chunks")
-
-    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
-        """Required by PreTrainedPolicy but not used for SARM."""
-        raise NotImplementedError("SARM model does not select actions")
-
     def _train_step(
         self,
         img_emb: torch.Tensor,  # (B, N, T, D)

src/lerobot/rewards/sarm/processor_sarm.py

@@ -1,5 +1,3 @@
-#!/usr/bin/env python
-
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -60,16 +58,15 @@ from lerobot.processor import (
     policy_action_to_transition,
     transition_to_policy_action,
 )
-from lerobot.types import EnvTransition, PolicyAction, TransitionKey
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
-
-from .configuration_sarm import SARMConfig
-from .sarm_utils import (
+from lerobot.rewards.sarm.configuration_sarm import SARMConfig
+from lerobot.rewards.sarm.sarm_utils import (
     apply_rewind_augmentation,
     compute_absolute_indices,
     find_stage_and_tau,
     pad_state_to_max_dim,
 )
+from lerobot.types import EnvTransition, PolicyAction, TransitionKey
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 
 class SARMEncodingProcessorStep(ProcessorStep):
@@ -455,7 +452,13 @@ class SARMEncodingProcessorStep(ProcessorStep):
             inputs = {k: v.to(self.device) for k, v in inputs.items()}
 
             # Get image embeddings
-            embeddings = self.clip_model.get_image_features(**inputs).detach().cpu()
+            # transformers 5.x returns BaseModelOutputWithPooling instead of a plain tensor
+            output = self.clip_model.get_image_features(**inputs)
+            if not isinstance(output, torch.Tensor):
+                output = output.pooler_output
+                if output is None:
+                    raise ValueError("pooler_output should not be None for CLIP models.")
+            embeddings = output.detach().cpu()
 
             # Handle single frame case
             if embeddings.dim() == 1:
@@ -482,7 +485,13 @@ class SARMEncodingProcessorStep(ProcessorStep):
         inputs = self.clip_processor.tokenizer([text], return_tensors="pt", padding=True, truncation=True)
         inputs = {k: v.to(self.device) for k, v in inputs.items()}
 
-        text_embedding = self.clip_model.get_text_features(**inputs).detach().cpu()
+        # transformers 5.x returns BaseModelOutputWithPooling instead of a plain tensor
+        output = self.clip_model.get_text_features(**inputs)
+        if not isinstance(output, torch.Tensor):
+            output = output.pooler_output
+            if output is None:
+                raise ValueError("pooler_output should not be None for CLIP models.")
+        text_embedding = output.detach().cpu()
         text_embedding = text_embedding.expand(batch_size, -1)
 
         return text_embedding

src/lerobot/rewards/sarm/rabc.py

@@ -1,5 +1,3 @@
-#!/usr/bin/env python
-
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,14 +12,38 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+"""
+RA-BC (Reward-Aligned Behavior Cloning) sample weighting implementation.
+
+This module implements the SampleWeighter protocol for RA-BC training,
+which weights training samples based on their task progress as measured
+by the SARM reward model.
+
+The weights are computed based on progress deltas:
+    delta = progress[t + chunk_size] - progress[t]
+
+High-quality samples (positive progress) get higher weights, while
+samples with negative progress (going backwards) get zero weight.
+
+See: https://arxiv.org/abs/2509.25358 for the SARM paper.
+"""
+
 import logging
 from pathlib import Path
+from typing import TYPE_CHECKING
 
 import numpy as np
-import pandas as pd
 import torch
 from huggingface_hub import hf_hub_download
 
+from lerobot.utils.import_utils import _pandas_available
+from lerobot.utils.sample_weighting import SampleWeighter
+
+if TYPE_CHECKING or _pandas_available:
+    import pandas as pd
+else:
+    pd = None  # type: ignore[assignment]
+
 
 def resolve_hf_path(path: str | Path) -> Path:
     """Resolve a path that may be a HuggingFace URL (hf://datasets/...) to a local path."""
@@ -34,23 +56,27 @@ def resolve_hf_path(path: str | Path) -> Path:
     return Path(path)
 
 
-class RABCWeights:
+class RABCWeights(SampleWeighter):
     """
     Load precomputed SARM progress values and compute RA-BC weights during training.
 
+    This class implements the SampleWeighter ABC for use with the generic
+    sample weighting infrastructure in lerobot.
+
     Progress values are loaded from a parquet file (generated by compute_rabc_weights.py).
     During training, computes:
         - progress_delta = progress[t + chunk_size] - progress[t]
         - rabc_weight based on the delta (paper Eq. 8-9)
 
     Args:
-        progress_path: Path to parquet file with precomputed progress values
-        chunk_size: Number of frames ahead for computing progress delta
-        head_mode: Which SARM head to use ("sparse" or "dense")
-        kappa: Hard threshold for high-quality samples (default: 0.01)
-        epsilon: Small constant for numerical stability (default: 1e-6)
-        fallback_weight: Weight to use for frames without valid delta (default: 1.0)
-        device: Device to return tensors on
+        progress_path: Path to parquet file with precomputed progress values.
+                      Supports HuggingFace URLs (hf://datasets/...).
+        chunk_size: Number of frames ahead for computing progress delta.
+        head_mode: Which SARM head to use ("sparse" or "dense").
+        kappa: Hard threshold for high-quality samples (default: 0.01).
+        epsilon: Small constant for numerical stability (default: 1e-6).
+        fallback_weight: Weight to use for frames without valid delta (default: 1.0).
+        device: Device to return tensors on.
     """
 
     def __init__(
@@ -61,7 +87,7 @@ class RABCWeights:
         kappa: float = 0.01,
         epsilon: float = 1e-6,
         fallback_weight: float = 1.0,
-        device: torch.device = None,
+        device: torch.device | None = None,
     ):
         self.progress_path = resolve_hf_path(progress_path)
         self.chunk_size = chunk_size
@@ -87,8 +113,8 @@ class RABCWeights:
 
         logging.info(f"Using progress column: {self.progress_column}")
 
-        self.progress_lookup = {}
-        self.episode_lookup = {}
+        self.progress_lookup: dict[int, float] = {}
+        self.episode_lookup: dict[int, int] = {}
 
         for _, row in self.df.iterrows():
             global_idx = int(row["index"])
@@ -100,7 +126,7 @@ class RABCWeights:
             self.episode_lookup[global_idx] = episode_idx
 
         # Build episode boundaries for delta computation
-        self.episode_boundaries = {}
+        self.episode_boundaries: dict[int, dict[str, int]] = {}
         for episode_idx in self.df["episode_index"].unique():
             ep_df = self.df[self.df["episode_index"] == episode_idx]
             self.episode_boundaries[int(episode_idx)] = {
@@ -114,7 +140,7 @@ class RABCWeights:
         # Compute global statistics for weight computation
         self._compute_global_stats()
 
-    def _compute_global_stats(self):
+    def _compute_global_stats(self) -> None:
         """Compute global mean and std of progress deltas for weight calculation."""
         all_deltas = []
 
@@ -138,8 +164,8 @@ class RABCWeights:
                 all_deltas.append(delta)
 
         if all_deltas:
-            self.delta_mean = max(np.mean(all_deltas), 0.0)
-            self.delta_std = max(np.std(all_deltas), self.epsilon)
+            self.delta_mean = max(float(np.mean(all_deltas)), 0.0)
+            self.delta_std = max(float(np.std(all_deltas)), self.epsilon)
             logging.info(f"Progress delta stats: mean={self.delta_mean:.4f}, std={self.delta_std:.4f}")
         else:
             self.delta_mean = 0.0
@@ -157,18 +183,19 @@ class RABCWeights:
         4. Compute weight using paper Eq. 8-9
 
         Args:
-            batch: Training batch containing "index" key with global frame indices
+            batch: Training batch containing "index" key with global frame indices.
 
         Returns:
             Tuple of:
-            - Weights tensor (batch_size,) normalized to sum to batch_size
-            - Stats dict with raw_mean_weight, num_zero_weight, num_full_weight
+            - Weights tensor (batch_size,) normalized to sum to batch_size.
+            - Stats dict with weighting statistics for logging.
         """
         indices = batch.get("index")
         if indices is None:
             logging.warning("RA-BC: Batch missing 'index' key, using uniform weights")
             batch_size = self._get_batch_size(batch)
-            return torch.ones(batch_size, device=self.device), {"raw_mean_weight": 1.0}
+            stats = {"mean_weight": 1.0, "num_zero_weight": 0, "num_full_weight": batch_size}
+            return torch.ones(batch_size, device=self.device), stats
 
         # Convert to list of ints
         if isinstance(indices, torch.Tensor):
@@ -183,29 +210,29 @@ class RABCWeights:
             delta = self._compute_delta(idx)
             deltas.append(delta)
 
-        deltas = np.array(deltas, dtype=np.float32)
+        deltas_array = np.array(deltas, dtype=np.float32)
 
         # Compute weights from deltas
-        weights = self._compute_weights(deltas)
+        weights = self._compute_weights(deltas_array)
 
         # Compute stats before normalization for logging
         raw_mean_weight = float(np.nanmean(weights))
         num_zero_weight = int(np.sum(weights == 0))
         num_full_weight = int(np.sum(weights == 1.0))
         batch_stats = {
-            "raw_mean_weight": raw_mean_weight,
+            "mean_weight": raw_mean_weight,
             "num_zero_weight": num_zero_weight,
             "num_full_weight": num_full_weight,
         }
 
-        weights = torch.tensor(weights, device=self.device, dtype=torch.float32)
+        weights_tensor = torch.tensor(weights, device=self.device, dtype=torch.float32)
 
         # Normalize to sum to batch_size
-        batch_size = len(weights)
-        weight_sum = weights.sum() + self.epsilon
-        weights = weights * batch_size / weight_sum
+        batch_size = len(weights_tensor)
+        weight_sum = weights_tensor.sum() + self.epsilon
+        weights_tensor = weights_tensor * batch_size / weight_sum
 
-        return weights, batch_stats
+        return weights_tensor, batch_stats
 
     def _compute_delta(self, global_idx: int) -> float:
         """Compute progress delta for a single frame."""
@@ -241,7 +268,7 @@ class RABCWeights:
         - Final weight: wi = 1{ri > κ} + 1{0 ≤ ri ≤ κ}˜wi
 
         Returns:
-            Array of weights
+            Array of weights.
         """
         valid_mask = ~np.isnan(deltas)
 
@@ -273,12 +300,13 @@ class RABCWeights:
             if key in batch:
                 val = batch[key]
                 if isinstance(val, (torch.Tensor, np.ndarray)):
-                    return val.shape[0]
+                    return int(val.shape[0])
         return 1
 
     def get_stats(self) -> dict:
-        """Get statistics."""
+        """Get global statistics about the RA-BC weighting."""
         return {
+            "type": "rabc",
             "num_frames": len(self.progress_lookup),
             "chunk_size": self.chunk_size,
             "head_mode": self.head_mode,

src/lerobot/rewards/sarm/sarm_utils.py

@@ -1,5 +1,3 @@
-#!/usr/bin/env python
-
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");

src/lerobot/rl/actor.py

@@ -76,6 +76,7 @@ from lerobot.transport.utils import (
 )
 from lerobot.types import TransitionKey
 from lerobot.utils.device_utils import get_safe_torch_device
+from lerobot.utils.process import ProcessSignalHandler
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.transition import (
@@ -94,7 +95,6 @@ from .gym_manipulator import (
     make_robot_env,
     step_env_and_process_transition,
 )
-from .process import ProcessSignalHandler
 from .queue import get_last_item_from_queue
 
 # Main entry point

src/lerobot/rl/crop_dataset_roi.py

@@ -193,15 +193,15 @@ def convert_lerobot_dataset_to_cropped_lerobot_dataset(
         fps=int(original_dataset.fps),
         root=new_dataset_root,
         robot_type=original_dataset.meta.robot_type,
-        features=original_dataset.meta.info["features"],
+        features=original_dataset.meta.info.features,
         use_videos=len(original_dataset.meta.video_keys) > 0,
     )
 
     # Update the metadata for every image key that will be cropped:
     # (Here we simply set the shape to be the final resize_size.)
     for key in crop_params_dict:
-        if key in new_dataset.meta.info["features"]:
-            new_dataset.meta.info["features"][key]["shape"] = [3] + list(resize_size)
+        if key in new_dataset.meta.info.features:
+            new_dataset.meta.info.features[key]["shape"] = (3, *resize_size)
 
     # TODO:  Directly modify the mp4 video + meta info features, instead of recreating a dataset
     prev_episode_index = 0

src/lerobot/rl/learner.py

@@ -90,6 +90,7 @@ from lerobot.utils.constants import (
     TRAINING_STATE_DIR,
 )
 from lerobot.utils.device_utils import get_safe_torch_device
+from lerobot.utils.process import ProcessSignalHandler
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.transition import move_state_dict_to_device, move_transition_to_device
 from lerobot.utils.utils import (
@@ -99,7 +100,6 @@ from lerobot.utils.utils import (
 
 from .buffer import ReplayBuffer, concatenate_batch_transitions
 from .learner_service import MAX_WORKERS, SHUTDOWN_TIMEOUT, LearnerService
-from .process import ProcessSignalHandler
 
 
 @parser.wrap()

src/lerobot/rollout/__init__.py

@@ -0,0 +1,87 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Policy deployment engine with pluggable rollout strategies."""
+
+from lerobot.utils.import_utils import require_package
+
+require_package("datasets", extra="dataset")
+
+from .configs import (
+    BaseStrategyConfig,
+    DAggerKeyboardConfig,
+    DAggerPedalConfig,
+    DAggerStrategyConfig,
+    HighlightStrategyConfig,
+    RolloutConfig,
+    RolloutStrategyConfig,
+    SentryStrategyConfig,
+)
+from .context import (
+    DatasetContext,
+    HardwareContext,
+    PolicyContext,
+    ProcessorContext,
+    RolloutContext,
+    RuntimeContext,
+    build_rollout_context,
+)
+from .inference import (
+    InferenceEngine,
+    InferenceEngineConfig,
+    RTCInferenceConfig,
+    RTCInferenceEngine,
+    SyncInferenceConfig,
+    SyncInferenceEngine,
+    create_inference_engine,
+)
+from .strategies import (
+    BaseStrategy,
+    DAggerStrategy,
+    HighlightStrategy,
+    RolloutStrategy,
+    SentryStrategy,
+    create_strategy,
+)
+
+__all__ = [
+    "BaseStrategy",
+    "BaseStrategyConfig",
+    "DAggerKeyboardConfig",
+    "DAggerPedalConfig",
+    "DAggerStrategy",
+    "DAggerStrategyConfig",
+    "DatasetContext",
+    "HardwareContext",
+    "HighlightStrategy",
+    "HighlightStrategyConfig",
+    "InferenceEngine",
+    "InferenceEngineConfig",
+    "PolicyContext",
+    "ProcessorContext",
+    "RTCInferenceConfig",
+    "RTCInferenceEngine",
+    "RolloutConfig",
+    "RolloutContext",
+    "RolloutStrategy",
+    "RolloutStrategyConfig",
+    "RuntimeContext",
+    "SentryStrategy",
+    "SentryStrategyConfig",
+    "SyncInferenceConfig",
+    "SyncInferenceEngine",
+    "build_rollout_context",
+    "create_inference_engine",
+    "create_strategy",
+]

src/lerobot/rollout/configs.py

@@ -0,0 +1,323 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Configuration dataclasses for the rollout deployment engine."""
+
+from __future__ import annotations
+
+import abc
+import logging
+from dataclasses import dataclass, field
+
+import draccus
+
+from lerobot.configs import PreTrainedConfig, parser
+from lerobot.configs.dataset import DatasetRecordConfig
+from lerobot.robots.config import RobotConfig
+from lerobot.teleoperators.config import TeleoperatorConfig
+from lerobot.utils.device_utils import auto_select_torch_device, is_torch_device_available
+
+from .inference import InferenceEngineConfig, SyncInferenceConfig
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Strategy configs (polymorphic dispatch via draccus ChoiceRegistry)
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class RolloutStrategyConfig(draccus.ChoiceRegistry, abc.ABC):
+    """Abstract base for rollout strategy configurations.
+
+    Use ``--strategy.type=<name>`` on the CLI to select a strategy.
+    """
+
+    @property
+    def type(self) -> str:
+        return self.get_choice_name(self.__class__)
+
+
+@RolloutStrategyConfig.register_subclass("base")
+@dataclass
+class BaseStrategyConfig(RolloutStrategyConfig):
+    """Autonomous rollout with no data recording."""
+
+    pass
+
+
+@RolloutStrategyConfig.register_subclass("sentry")
+@dataclass
+class SentryStrategyConfig(RolloutStrategyConfig):
+    """Continuous autonomous rollout with always-on recording.
+
+    Episode duration is derived from camera resolution, FPS, and
+    ``target_video_file_size_mb`` so that each saved episode produces a
+    video file that has crossed the target size.  This aligns episode
+    boundaries with the dataset's video file chunking, so each
+    ``push_to_hub`` call uploads complete video files rather than
+    re-uploading a growing file that hasn't crossed the chunk boundary.
+    """
+
+    upload_every_n_episodes: int = 5
+    # Target video file size in MB for episode rotation.  Episodes are
+    # saved once the estimated video duration would exceed this limit.
+    # Defaults to DEFAULT_VIDEO_FILE_SIZE_IN_MB when set to None.
+    target_video_file_size_mb: int | None = None
+
+
+@RolloutStrategyConfig.register_subclass("highlight")
+@dataclass
+class HighlightStrategyConfig(RolloutStrategyConfig):
+    """Autonomous rollout with on-demand recording via ring buffer.
+
+    A memory-bounded ring buffer continuously captures telemetry.  When
+    the user presses the save key, the buffer contents are flushed to
+    the dataset and live recording continues until the key is pressed
+    again.
+    """
+
+    ring_buffer_seconds: float = 10.0
+    ring_buffer_max_memory_mb: int = 1024
+    save_key: str = "s"
+    push_key: str = "h"
+
+
+@dataclass
+class DAggerKeyboardConfig:
+    """Keyboard key bindings for DAgger controls.
+
+    Keys are specified as single characters (e.g. ``"c"``, ``"h"``) or
+    special key names (``"space"``).
+    """
+
+    pause_resume: str = "space"
+    correction: str = "tab"
+    upload: str = "enter"
+
+
+@dataclass
+class DAggerPedalConfig:
+    """Foot pedal configuration for DAgger controls.
+
+    Pedal codes are evdev key code strings (e.g. ``"KEY_A"``).
+    """
+
+    device_path: str = "/dev/input/by-id/usb-PCsensor_FootSwitch-event-kbd"
+    pause_resume: str = "KEY_A"
+    correction: str = "KEY_B"
+    upload: str = "KEY_C"
+
+
+@RolloutStrategyConfig.register_subclass("dagger")
+@dataclass
+class DAggerStrategyConfig(RolloutStrategyConfig):
+    """Human-in-the-loop data collection (DAgger / RaC).
+
+    Alternates between autonomous policy execution and human intervention.
+    Intervention frames are tagged with ``intervention=True``.
+
+    Input is controlled via either a keyboard or foot pedal, selected by
+    ``input_device``.  Each device exposes three actions:
+
+    1. **pause_resume** — toggle policy execution on/off.
+    2. **correction** — toggle human correction recording.
+    3. **upload** — push dataset to hub on demand (corrections-only mode).
+
+    When ``record_autonomous=False`` (default) only human-correction windows
+    are recorded — each correction becomes its own episode.  Set to ``True``
+    to record both autonomous and correction frames with size-based episode
+    rotation (same as Sentry) and background uploading.  ``push_to_hub`` is
+    blocked while a correction is in progress.
+    """
+
+    # Number of correction episodes to collect (corrections-only mode).
+    # When None, falls back to ``--dataset.num_episodes``.
+    num_episodes: int | None = None
+    record_autonomous: bool = False
+    upload_every_n_episodes: int = 5
+    # Target video file size in MB for episode rotation (record_autonomous
+    # mode only).  Defaults to DEFAULT_VIDEO_FILE_SIZE_IN_MB when None.
+    target_video_file_size_mb: int | None = None
+    input_device: str = "keyboard"
+    keyboard: DAggerKeyboardConfig = field(default_factory=DAggerKeyboardConfig)
+    pedal: DAggerPedalConfig = field(default_factory=DAggerPedalConfig)
+
+    def __post_init__(self):
+        if self.input_device not in ("keyboard", "pedal"):
+            raise ValueError(f"DAgger input_device must be 'keyboard' or 'pedal', got '{self.input_device}'")
+
+
+# ---------------------------------------------------------------------------
+# Top-level rollout config
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class RolloutConfig:
+    """Top-level configuration for the ``lerobot-rollout`` CLI.
+
+    Combines hardware, policy, strategy, and runtime settings.  The
+    ``__post_init__`` method performs fail-fast validation to reject
+    invalid flag combinations early.
+    """
+
+    # Hardware
+    robot: RobotConfig | None = None
+    teleop: TeleoperatorConfig | None = None
+
+    # Policy (loaded from --policy.path via __post_init__)
+    policy: PreTrainedConfig | None = None
+
+    # Strategy (polymorphic: --strategy.type=base|sentry|highlight|dagger)
+    strategy: RolloutStrategyConfig = field(default_factory=BaseStrategyConfig)
+
+    # Inference backend (polymorphic: --inference.type=sync|rtc)
+    inference: InferenceEngineConfig = field(default_factory=SyncInferenceConfig)
+
+    # Dataset (required for sentry, highlight, dagger; None for base)
+    dataset: DatasetRecordConfig | None = None
+
+    # Runtime
+    fps: float = 30.0
+    duration: float = 0.0  # 0 = infinite (24/7 mode)
+    interpolation_multiplier: int = 1
+    device: str | None = None
+    task: str = ""
+    display_data: bool = False
+    # Display data on a remote Rerun server
+    display_ip: str | None = None
+    # Port of the remote Rerun server
+    display_port: int | None = None
+    # Whether to display compressed images in Rerun
+    display_compressed_images: bool = False
+    # Use vocal synthesis to read events
+    play_sounds: bool = True
+    resume: bool = False
+    # Rename map for mapping robot/dataset observation keys to policy keys
+    rename_map: dict[str, str] = field(default_factory=dict)
+
+    # Hardware teardown
+    # When True (default), smoothly interpolate the robot back to the joint
+    # positions captured at startup before disconnecting.  Set to False to
+    # leave the robot in its final achieved pose at shutdown.
+    return_to_initial_position: bool = True
+
+    # Torch compile
+    use_torch_compile: bool = False
+    torch_compile_backend: str = "inductor"
+    torch_compile_mode: str = "default"
+    compile_warmup_inferences: int = 2
+
+    def __post_init__(self):
+        """Validate config invariants and load the policy config from ``--policy.path``."""
+        # --- Strategy-specific validation ---
+        if isinstance(self.strategy, DAggerStrategyConfig) and self.teleop is None:
+            raise ValueError("DAgger strategy requires --teleop.type to be set")
+
+        # TODO(Steven): DAgger shouldn't require a dataset (user may want to just rollout+intervene without recording), but for now we require it to simplify the implementation.
+        needs_dataset = isinstance(
+            self.strategy, (SentryStrategyConfig, HighlightStrategyConfig, DAggerStrategyConfig)
+        )
+        if needs_dataset and (self.dataset is None or not self.dataset.repo_id):
+            raise ValueError(f"{self.strategy.type} strategy requires --dataset.repo_id to be set")
+
+        if isinstance(self.strategy, BaseStrategyConfig) and self.dataset is not None:
+            raise ValueError(
+                "Base strategy does not record data. Use sentry, highlight, or dagger for recording."
+            )
+
+        # Sentry MUST use streaming encoding to avoid disk I/O blocking the control loop
+        if (
+            isinstance(self.strategy, SentryStrategyConfig)
+            and self.dataset is not None
+            and not self.dataset.streaming_encoding
+        ):
+            logger.warning("Sentry mode forces streaming_encoding=True")
+            self.dataset.streaming_encoding = True
+
+        # Highlight writes frames while the policy is still running, so streaming is mandatory.
+        if (
+            isinstance(self.strategy, HighlightStrategyConfig)
+            and self.dataset is not None
+            and not self.dataset.streaming_encoding
+        ):
+            logger.warning("Highlight mode forces streaming_encoding=True")
+            self.dataset.streaming_encoding = True
+
+        # DAgger: streaming is mandatory only when the autonomous phase is also recorded.
+        if isinstance(self.strategy, DAggerStrategyConfig) and self.dataset is not None:
+            if self.strategy.record_autonomous and not self.dataset.streaming_encoding:
+                logger.warning("DAgger with record_autonomous=True forces streaming_encoding=True")
+                self.dataset.streaming_encoding = True
+            elif not self.strategy.record_autonomous and not self.dataset.streaming_encoding:
+                logger.info(
+                    "Streaming encoding is disabled for DAgger corrections-only mode. "
+                    "Consider enabling it for faster episode saving: "
+                    "--dataset.streaming_encoding=true --dataset.encoder_threads=2"
+                )
+
+        # DAgger: resolve num_episodes from dataset config when not explicitly set.
+        if isinstance(self.strategy, DAggerStrategyConfig) and self.strategy.num_episodes is None:
+            if self.dataset is not None:
+                self.strategy.num_episodes = self.dataset.num_episodes
+                logger.info(
+                    "DAgger num_episodes not set — using --dataset.num_episodes=%d",
+                    self.strategy.num_episodes,
+                )
+            else:
+                raise ValueError(
+                    "DAgger num_episodes must be set either via --strategy.num_episodes or --dataset.num_episodes"
+                )
+
+        # --- Policy loading ---
+        if self.robot is None:
+            raise ValueError("--robot.type is required for rollout")
+
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        if self.policy is None:
+            raise ValueError("--policy.path is required for rollout")
+
+        # --- Task resolution ---
+        # When any --dataset.* flag is passed, draccus creates a DatasetRecordConfig with single_task="".
+        # If the user set the task via the top-level --task flag, propagate it so that all
+        # downstream consumers (inference engine, dataset frame builders) see it.
+        if self.dataset is not None and not self.dataset.single_task and self.task:
+            logger.info("Propagating top-level task '%s' to dataset config", self.task)
+            self.dataset.single_task = self.task
+        elif self.dataset is not None and self.dataset.single_task and not self.task:
+            logger.info("Propagating dataset single_task '%s' to top-level task", self.dataset.single_task)
+            self.task = self.dataset.single_task
+
+        # --- Device resolution ---
+        # Resolve device from the policy config when not explicitly set so all
+        # components (policy.to, preprocessor, inference engine) use the same
+        # device string instead of inconsistent fallbacks.
+        if self.device is None or not is_torch_device_available(self.device):
+            resolved = self.policy.device
+            if resolved:
+                self.device = resolved
+                logger.info("Resolved device from policy config: %s", self.device)
+            else:
+                self.device = auto_select_torch_device().type
+                logger.info("No policy config to resolve device from; auto-selected device: %s", self.device)
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        return ["policy"]

src/lerobot/rollout/context.py

@@ -0,0 +1,454 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Rollout context: shared state created once before strategy dispatch.
+
+Grouped into five topical sub-contexts — :class:`RuntimeContext`,
+:class:`HardwareContext`, :class:`PolicyContext`, :class:`ProcessorContext`,
+and :class:`DatasetContext` — assembled into :class:`RolloutContext`.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, field
+from threading import Event
+
+import torch
+
+from lerobot.configs import FeatureType
+from lerobot.datasets import (
+    LeRobotDataset,
+    aggregate_pipeline_dataset_features,
+    create_initial_features,
+)
+from lerobot.policies import get_policy_class, make_pre_post_processors
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.processor import (
+    PolicyProcessorPipeline,
+    RobotAction,
+    RobotObservation,
+    RobotProcessorPipeline,
+    make_default_processors,
+    rename_stats,
+)
+from lerobot.processor.relative_action_processor import RelativeActionsProcessorStep
+from lerobot.robots import make_robot_from_config
+from lerobot.teleoperators import Teleoperator, make_teleoperator_from_config
+from lerobot.utils.feature_utils import combine_feature_dicts, hw_to_dataset_features
+
+from .configs import BaseStrategyConfig, DAggerStrategyConfig, RolloutConfig
+from .inference import (
+    InferenceEngine,
+    RTCInferenceConfig,
+    SyncInferenceConfig,
+    create_inference_engine,
+)
+from .robot_wrapper import ThreadSafeRobot
+
+logger = logging.getLogger(__name__)
+
+
+def _resolve_action_key_order(
+    policy_action_names: list[str] | None, dataset_action_names: list[str]
+) -> list[str]:
+    """Choose action name ordering for mapping policy tensor outputs to robot action dicts."""
+    if not policy_action_names:
+        return dataset_action_names
+    policy_action_names = list(policy_action_names)
+    if len(policy_action_names) != len(dataset_action_names):
+        logger.warning(
+            "policy.action_feature_names length (%d) != dataset action dim (%d); using dataset order",
+            len(policy_action_names),
+            len(dataset_action_names),
+        )
+        return dataset_action_names
+    if set(dataset_action_names) != set(policy_action_names):
+        logger.warning("policy.action_feature_names keys don't match dataset; using dataset order")
+        return dataset_action_names
+    return policy_action_names
+
+
+# ---------------------------------------------------------------------------
+# Sub-contexts
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class RuntimeContext:
+    """Runtime knobs shared with every strategy."""
+
+    cfg: RolloutConfig
+    shutdown_event: Event
+
+
+@dataclass
+class HardwareContext:
+    """Connected hardware.
+
+    The raw robot is available via ``robot_wrapper.inner`` when needed
+    (e.g. for disconnect); strategies should otherwise go through the
+    thread-safe wrapper.
+
+    ``initial_position`` stores the robot's joint positions at connect
+    time.  Strategies use it to return the robot to a safe pose before
+    shutting down.
+    """
+
+    robot_wrapper: ThreadSafeRobot
+    teleop: Teleoperator | None
+    initial_position: dict | None = None
+
+
+@dataclass
+class PolicyContext:
+    """Loaded policy and its inference engine."""
+
+    policy: PreTrainedPolicy
+    preprocessor: PolicyProcessorPipeline
+    postprocessor: PolicyProcessorPipeline
+    inference: InferenceEngine
+
+
+@dataclass
+class ProcessorContext:
+    """Robot-side pipelines (run outside the policy)."""
+
+    teleop_action_processor: RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction]
+    robot_action_processor: RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction]
+    robot_observation_processor: RobotProcessorPipeline[RobotObservation, RobotObservation]
+
+
+@dataclass
+class DatasetContext:
+    """Dataset and feature bookkeeping."""
+
+    dataset: LeRobotDataset | None
+    dataset_features: dict = field(default_factory=dict)
+    hw_features: dict = field(default_factory=dict)
+    ordered_action_keys: list[str] = field(default_factory=list)
+
+
+@dataclass
+class RolloutContext:
+    """Bundle of sub-contexts passed to every rollout strategy.
+
+    Built once by :func:`build_rollout_context` before strategy dispatch.
+    """
+
+    runtime: RuntimeContext
+    hardware: HardwareContext
+    policy: PolicyContext
+    processors: ProcessorContext
+    data: DatasetContext
+
+
+# ---------------------------------------------------------------------------
+# Build
+# ---------------------------------------------------------------------------
+
+
+def build_rollout_context(
+    cfg: RolloutConfig,
+    shutdown_event: Event,
+    teleop_action_processor: RobotProcessorPipeline | None = None,
+    robot_action_processor: RobotProcessorPipeline | None = None,
+    robot_observation_processor: RobotProcessorPipeline | None = None,
+) -> RolloutContext:
+    """Wire up policy, processors, hardware, dataset, and inference engine.
+
+    The order is policy-first / hardware-last so a bad ``--policy.path``
+    fails fast without touching the robot.
+    """
+    is_rtc = isinstance(cfg.inference, RTCInferenceConfig)
+
+    # --- 1. Policy (heavy I/O, but no hardware yet) -------------------
+    logger.info("Loading policy from '%s'...", cfg.policy.pretrained_path)
+    policy_config = cfg.policy
+    policy_class = get_policy_class(policy_config.type)
+
+    if hasattr(policy_config, "compile_model"):
+        policy_config.compile_model = cfg.use_torch_compile
+
+    if policy_config.type == "vqbet" and cfg.device == "mps":
+        raise NotImplementedError(
+            "Current implementation of VQBeT does not support `mps` backend. "
+            "Please use `cpu` or `cuda` backend."
+        )
+
+    if policy_config.use_peft:
+        from peft import PeftConfig, PeftModel
+
+        peft_path = policy_config.pretrained_path
+        peft_config = PeftConfig.from_pretrained(peft_path)
+        policy = policy_class.from_pretrained(
+            pretrained_name_or_path=peft_config.base_model_name_or_path, config=policy_config
+        )
+        policy = PeftModel.from_pretrained(policy, peft_path, config=peft_config)
+    else:
+        policy = policy_class.from_pretrained(policy_config.pretrained_path, config=policy_config)
+
+    if is_rtc:
+        policy.config.rtc_config = cfg.inference.rtc
+        if hasattr(policy, "init_rtc_processor"):
+            policy.init_rtc_processor()
+
+    policy = policy.to(cfg.device)
+    policy.eval()
+    logger.info("Policy loaded: type=%s, device=%s", policy_config.type, cfg.device)
+
+    if cfg.use_torch_compile and policy.type not in ("pi0", "pi05"):
+        try:
+            if hasattr(torch, "compile"):
+                compile_kwargs = {
+                    "backend": cfg.torch_compile_backend,
+                    "mode": cfg.torch_compile_mode,
+                    "options": {"triton.cudagraphs": False},
+                }
+                policy.predict_action_chunk = torch.compile(policy.predict_action_chunk, **compile_kwargs)
+                logger.info("torch.compile applied to predict_action_chunk")
+        except Exception as e:
+            logger.warning("Failed to apply torch.compile: %s", e)
+
+    # --- 2. Robot-side processors (user-supplied or defaults) --------
+    if (
+        teleop_action_processor is None
+        or robot_action_processor is None
+        or robot_observation_processor is None
+    ):
+        _t, _r, _o = make_default_processors()
+        teleop_action_processor = teleop_action_processor or _t
+        robot_action_processor = robot_action_processor or _r
+        robot_observation_processor = robot_observation_processor or _o
+
+    # --- 3. Hardware (heaviest side-effect, deferred) -----------------
+    logger.info("Connecting robot (%s)...", cfg.robot.type if cfg.robot else "?")
+    robot = make_robot_from_config(cfg.robot)
+    robot.connect()
+    logger.info("Robot connected: %s", robot.name)
+
+    # Store the initial joint positions so we can return to a safe pose on shutdown.
+    initial_obs = robot.get_observation()
+    initial_position = {k: v for k, v in initial_obs.items() if k.endswith(".pos")}
+    logger.info("Captured initial robot position (%d keys)", len(initial_position))
+
+    robot_wrapper = ThreadSafeRobot(robot)
+
+    teleop = None
+    if cfg.teleop is not None:
+        logger.info("Connecting teleoperator (%s)...", cfg.teleop.type if cfg.teleop else "?")
+        teleop = make_teleoperator_from_config(cfg.teleop)
+        teleop.connect()
+        logger.info("Teleoperator connected")
+
+    # TODO(Steven): once Teleoperator motor-control methods are standardised
+    # (``enable_torque`` / ``disable_torque`` / ``write_goal_positions``), gate
+    # the DAgger strategy on their presence here and fail fast with a helpful
+    # message instead of relying on the operator to pre-align the leader by
+    # hand.  See :func:`DAggerStrategy._apply_transition` for the matching
+    # disabled call sites.
+    # if isinstance(cfg.strategy, DAggerStrategyConfig) and teleop is not None:
+    #     required_teleop_methods = ("enable_torque", "disable_torque", "write_goal_positions")
+    #     missing = [m for m in required_teleop_methods if not callable(getattr(teleop, m, None))]
+    #     if missing:
+    #         teleop.disconnect()
+    #         raise ValueError(
+    #             f"DAgger strategy requires a teleoperator with motor control methods "
+    #             f"{required_teleop_methods}. '{type(teleop).__name__}' is missing: {missing}"
+    #         )
+
+    # --- 4. Features + action-key reconciliation ---------------------
+    # TODO(Steven):Only ``.pos`` joint features are routed to the policy as state and as the
+    # action target; velocity and torque channels (when present) are kept in
+    # the raw observation but excluded from the policy-facing tensors.
+    all_obs_features = robot.observation_features
+    # ``observation_features`` values are either a tuple (camera shape) or the
+    # ``float`` type itself used as a sentinel for scalar motor features —
+    # see ``dict[str, type | tuple]`` annotation on ``Robot.observation_features``.
+    observation_features_hw = {
+        k: v
+        for k, v in all_obs_features.items()
+        if isinstance(v, tuple) or (v is float and k.endswith(".pos"))
+    }
+    action_features_hw = {k: v for k, v in robot.action_features.items() if k.endswith(".pos")}
+
+    # The action side is always needed: sync inference reads action names from
+    # ``dataset_features[ACTION]`` to map policy tensors back to robot actions.
+    action_dataset_features = aggregate_pipeline_dataset_features(
+        pipeline=teleop_action_processor,
+        initial_features=create_initial_features(action=action_features_hw),
+        use_videos=cfg.dataset.video if cfg.dataset else True,
+    )
+    # Observation-side aggregation is needed because of build_dataset_frame
+    observation_dataset_features = aggregate_pipeline_dataset_features(
+        pipeline=robot_observation_processor,
+        initial_features=create_initial_features(observation=observation_features_hw),
+        use_videos=cfg.dataset.video if cfg.dataset else True,
+    )
+    dataset_features = combine_feature_dicts(action_dataset_features, observation_dataset_features)
+    hw_features = hw_to_dataset_features(observation_features_hw, "observation")
+    raw_action_keys = list(action_features_hw.keys())
+    policy_action_names = getattr(policy_config, "action_feature_names", None)
+    ordered_action_keys = _resolve_action_key_order(
+        list(policy_action_names) if policy_action_names else None,
+        raw_action_keys,
+    )
+
+    # Validate visual features if no rename_map is active
+    rename_map = cfg.rename_map
+    if not rename_map:
+        expected_visuals = {
+            k for k, v in policy_config.input_features.items() if v.type == FeatureType.VISUAL
+        }
+        provided_visuals = {
+            f"observation.images.{k}" for k, v in robot.observation_features.items() if isinstance(v, tuple)
+        }
+        policy_subset = expected_visuals.issubset(provided_visuals)
+        hw_subset = provided_visuals.issubset(expected_visuals)
+        if not (policy_subset or hw_subset):
+            raise ValueError(
+                f"Visual feature mismatch between policy and robot hardware.\n"
+                f"Policy expects: {expected_visuals}\n"
+                f"Robot provides: {provided_visuals}"
+            )
+
+    # --- 5. Dataset -------------
+    dataset = None
+    if cfg.dataset is not None and not isinstance(cfg.strategy, BaseStrategyConfig):
+        logger.info("Setting up dataset (repo_id=%s)...", cfg.dataset.repo_id)
+        if cfg.resume:
+            dataset = LeRobotDataset.resume(
+                cfg.dataset.repo_id,
+                root=cfg.dataset.root,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+                vcodec=cfg.dataset.vcodec,
+                streaming_encoding=cfg.dataset.streaming_encoding,
+                encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
+                encoder_threads=cfg.dataset.encoder_threads,
+                image_writer_processes=cfg.dataset.num_image_writer_processes,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera
+                * len(robot.cameras if hasattr(robot, "cameras") else []),
+            )
+        else:
+            if isinstance(cfg.strategy, DAggerStrategyConfig):
+                dataset_features["intervention"] = {
+                    "dtype": "bool",
+                    "shape": (1,),
+                    "names": None,
+                }
+
+            repo_name = cfg.dataset.repo_id.split("/", 1)[-1]
+            if not repo_name.startswith("rollout_"):
+                raise ValueError(
+                    "Dataset names for rollout must start with 'rollout_'. "
+                    "Use --dataset.repo_id=<user>/rollout_<name> for policy deployment datasets."
+                )
+            cfg.dataset.stamp_repo_id()
+            target_video_mb = getattr(cfg.strategy, "target_video_file_size_mb", None)
+            dataset = LeRobotDataset.create(
+                cfg.dataset.repo_id,
+                cfg.dataset.fps,
+                root=cfg.dataset.root,
+                robot_type=robot.name,
+                features=dataset_features,
+                use_videos=cfg.dataset.video,
+                image_writer_processes=cfg.dataset.num_image_writer_processes,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera
+                * len(robot.cameras if hasattr(robot, "cameras") else []),
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+                vcodec=cfg.dataset.vcodec,
+                streaming_encoding=cfg.dataset.streaming_encoding,
+                encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
+                encoder_threads=cfg.dataset.encoder_threads,
+                video_files_size_in_mb=target_video_mb,
+            )
+
+    if dataset is not None:
+        logger.info("Dataset ready: %s (%d existing episodes)", dataset.repo_id, dataset.num_episodes)
+
+    # --- 6. Policy pre/post processors (needs dataset stats if any) ---
+    dataset_stats = None
+    if dataset is not None:
+        dataset_stats = rename_stats(
+            dataset.meta.stats,
+            cfg.rename_map,
+        )
+
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=policy_config,
+        pretrained_path=cfg.policy.pretrained_path,
+        dataset_stats=dataset_stats,
+        preprocessor_overrides={
+            "device_processor": {"device": cfg.device},
+            "rename_observations_processor": {"rename_map": cfg.rename_map},
+        },
+    )
+
+    if isinstance(cfg.inference, SyncInferenceConfig) and any(
+        isinstance(step, RelativeActionsProcessorStep) and step.enabled
+        for step in getattr(preprocessor, "steps", ())
+    ):
+        raise NotImplementedError(
+            "SyncInferenceEngine does not support policies with relative actions for now."
+            "Use --inference.type=rtc or remove relative action processor steps from the policy pipeline."
+        )
+
+    # --- 7. Inference strategy (needs policy + pre/post + hardware) --
+    logger.info(
+        "Creating inference engine (type=%s)...",
+        cfg.inference.type if hasattr(cfg.inference, "type") else "sync",
+    )
+    task_str = cfg.dataset.single_task if cfg.dataset else cfg.task
+    inference_strategy = create_inference_engine(
+        cfg.inference,
+        policy=policy,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
+        robot_wrapper=robot_wrapper,
+        hw_features=hw_features,
+        dataset_features=dataset_features,
+        ordered_action_keys=ordered_action_keys,
+        task=task_str,
+        fps=cfg.fps,
+        device=cfg.device,
+        use_torch_compile=cfg.use_torch_compile,
+        compile_warmup_inferences=cfg.compile_warmup_inferences,
+        shutdown_event=shutdown_event,
+    )
+
+    # --- 8. Assemble ---------------------------------------------------
+    logger.info("Rollout context assembled successfully")
+    return RolloutContext(
+        runtime=RuntimeContext(cfg=cfg, shutdown_event=shutdown_event),
+        hardware=HardwareContext(
+            robot_wrapper=robot_wrapper, teleop=teleop, initial_position=initial_position
+        ),
+        policy=PolicyContext(
+            policy=policy,
+            preprocessor=preprocessor,
+            postprocessor=postprocessor,
+            inference=inference_strategy,
+        ),
+        processors=ProcessorContext(
+            teleop_action_processor=teleop_action_processor,
+            robot_action_processor=robot_action_processor,
+            robot_observation_processor=robot_observation_processor,
+        ),
+        data=DatasetContext(
+            dataset=dataset,
+            dataset_features=dataset_features,
+            hw_features=hw_features,
+            ordered_action_keys=ordered_action_keys,
+        ),
+    )

src/lerobot/rollout/inference/__init__.py

@@ -0,0 +1,39 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Inference engine package — backend-agnostic action production.
+
+Concrete backends (``sync``, ``rtc``, ...) expose the same small interface so
+rollout strategies never branch on which backend is in use.
+"""
+
+from .base import InferenceEngine
+from .factory import (
+    InferenceEngineConfig,
+    RTCInferenceConfig,
+    SyncInferenceConfig,
+    create_inference_engine,
+)
+from .rtc import RTCInferenceEngine
+from .sync import SyncInferenceEngine
+
+__all__ = [
+    "InferenceEngine",
+    "InferenceEngineConfig",
+    "RTCInferenceConfig",
+    "RTCInferenceEngine",
+    "SyncInferenceConfig",
+    "SyncInferenceEngine",
+    "create_inference_engine",
+]

src/lerobot/rollout/inference/base.py

@@ -0,0 +1,89 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Inference engine ABC.
+
+Rollout strategies consume actions through this small interface so they
+do not need to know whether inference happens inline on the control thread
+or asynchronously in a background thread (RTC).
+"""
+
+from __future__ import annotations
+
+import abc
+
+import torch
+
+
+class InferenceEngine(abc.ABC):
+    """Abstract backend for producing actions during rollout.
+
+    Subclasses decide whether inference happens inline on the control
+    thread or asynchronously in a background thread.  The contract is
+    minimal so additional backends can be plugged in without touching
+    rollout strategies.
+
+    Lifecycle
+    ---------
+    ``start`` — prepare the backend (e.g. launch a background thread).
+    ``stop`` — shut the backend down cleanly.
+    ``reset`` — clear episode-scoped state (policy hidden state, queues…).
+
+    Action production
+    -----------------
+    ``get_action(obs_frame)`` — return the next action tensor, or
+    ``None`` if none is available (e.g. async queue empty).  Sync
+    backends always compute from ``obs_frame``; async backends ignore
+    it (they receive observations via ``notify_observation``).
+
+    Optional hooks
+    --------------
+    ``notify_observation`` / ``pause`` / ``resume`` have a no-op default
+    so rollout strategies can invoke them unconditionally.
+    """
+
+    @abc.abstractmethod
+    def start(self) -> None:
+        """Initialise the backend."""
+
+    @abc.abstractmethod
+    def stop(self) -> None:
+        """Tear the backend down."""
+
+    @abc.abstractmethod
+    def reset(self) -> None:
+        """Clear episode-scoped state."""
+
+    @abc.abstractmethod
+    def get_action(self, obs_frame: dict | None) -> torch.Tensor | None:
+        """Return the next action tensor, or ``None`` if unavailable."""
+
+    def notify_observation(self, obs: dict) -> None:  # noqa: B027
+        """Publish the latest processed observation.  Default: no-op."""
+
+    def pause(self) -> None:  # noqa: B027
+        """Pause background inference.  Default: no-op."""
+
+    def resume(self) -> None:  # noqa: B027
+        """Resume background inference.  Default: no-op."""
+
+    @property
+    def ready(self) -> bool:
+        """True once the backend can produce actions (e.g. warmup done)."""
+        return True
+
+    @property
+    def failed(self) -> bool:
+        """True if an unrecoverable error occurred in the backend."""
+        return False

src/lerobot/rollout/inference/factory.py

@@ -0,0 +1,128 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Inference engine configs and factory.
+
+Selection is explicit via ``--inference.type=sync|rtc``.  Adding a new
+backend requires registering its config subclass and dispatching it in
+:func:`create_inference_engine`.
+"""
+
+from __future__ import annotations
+
+import abc
+import logging
+from dataclasses import dataclass, field
+from threading import Event
+
+import draccus
+
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.processor import PolicyProcessorPipeline
+
+from ..robot_wrapper import ThreadSafeRobot
+from .base import InferenceEngine
+from .rtc import RTCInferenceEngine
+from .sync import SyncInferenceEngine
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Configs
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class InferenceEngineConfig(draccus.ChoiceRegistry, abc.ABC):
+    """Abstract base for inference backend configuration.
+
+    Use ``--inference.type=<name>`` on the CLI to select a backend.
+    """
+
+    @property
+    def type(self) -> str:
+        return self.get_choice_name(self.__class__)
+
+
+@InferenceEngineConfig.register_subclass("sync")
+@dataclass
+class SyncInferenceConfig(InferenceEngineConfig):
+    """Inline synchronous inference (one policy call per control tick)."""
+
+
+@InferenceEngineConfig.register_subclass("rtc")
+@dataclass
+class RTCInferenceConfig(InferenceEngineConfig):
+    """Real-Time Chunking: async policy inference in a background thread."""
+
+    # Eagerly constructed so draccus exposes nested fields directly on the CLI
+    # (e.g. ``--inference.rtc.execution_horizon=...``).
+    rtc: RTCConfig = field(default_factory=RTCConfig)
+    queue_threshold: int = 30
+
+
+# ---------------------------------------------------------------------------
+# Factory
+# ---------------------------------------------------------------------------
+
+
+def create_inference_engine(
+    config: InferenceEngineConfig,
+    *,
+    policy: PreTrainedPolicy,
+    preprocessor: PolicyProcessorPipeline,
+    postprocessor: PolicyProcessorPipeline,
+    robot_wrapper: ThreadSafeRobot,
+    hw_features: dict,
+    dataset_features: dict,
+    ordered_action_keys: list[str],
+    task: str,
+    fps: float,
+    device: str | None,
+    use_torch_compile: bool = False,
+    compile_warmup_inferences: int = 2,
+    shutdown_event: Event | None = None,
+) -> InferenceEngine:
+    """Instantiate the appropriate inference engine from a config object."""
+    logger.info("Creating inference engine: %s", config.type)
+    if isinstance(config, SyncInferenceConfig):
+        return SyncInferenceEngine(
+            policy=policy,
+            preprocessor=preprocessor,
+            postprocessor=postprocessor,
+            dataset_features=dataset_features,
+            ordered_action_keys=ordered_action_keys,
+            task=task,
+            device=device,
+            robot_type=robot_wrapper.robot_type,
+        )
+    if isinstance(config, RTCInferenceConfig):
+        return RTCInferenceEngine(
+            policy=policy,
+            preprocessor=preprocessor,
+            postprocessor=postprocessor,
+            robot_wrapper=robot_wrapper,
+            rtc_config=config.rtc,
+            hw_features=hw_features,
+            task=task,
+            fps=fps,
+            device=device,
+            use_torch_compile=use_torch_compile,
+            compile_warmup_inferences=compile_warmup_inferences,
+            rtc_queue_threshold=config.queue_threshold,
+            shutdown_event=shutdown_event,
+        )
+    raise ValueError(f"Unknown inference engine type: {type(config).__name__}")

src/lerobot/rollout/inference/rtc.py

@@ -0,0 +1,360 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Real-Time Chunking inference engine.
+
+A background thread produces action chunks asynchronously via
+:meth:`policy.predict_action_chunk`.  The main control loop polls
+``get_action`` for the next ready action; observations flow the other
+way via ``notify_observation``.
+"""
+
+from __future__ import annotations
+
+import logging
+import math
+import time
+import traceback
+from threading import Event, Lock, Thread
+from typing import Any
+
+import torch
+
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rtc import ActionQueue, LatencyTracker, reanchor_relative_rtc_prefix
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.utils import prepare_observation_for_inference
+from lerobot.processor import (
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    RelativeActionsProcessorStep,
+)
+from lerobot.utils.feature_utils import build_dataset_frame
+
+from ..robot_wrapper import ThreadSafeRobot
+from .base import InferenceEngine
+
+logger = logging.getLogger(__name__)
+
+# How long the RTC loop sleeps when paused, idle, or backpressured by a full queue.
+_RTC_IDLE_SLEEP_S: float = 0.01
+# Backoff between transient inference errors (per consecutive failure).
+_RTC_ERROR_RETRY_DELAY_S: float = 0.5
+# Consecutive transient errors tolerated before giving up and propagating shutdown.
+_RTC_MAX_CONSECUTIVE_ERRORS: int = 10
+# Hard timeout for joining the RTC thread on stop().
+_RTC_JOIN_TIMEOUT_S: float = 3.0
+
+
+# ---------------------------------------------------------------------------
+# RTC helpers
+# ---------------------------------------------------------------------------
+
+
+def _normalize_prev_actions_length(prev_actions: torch.Tensor, target_steps: int) -> torch.Tensor:
+    """Pad or truncate RTC prefix actions to a fixed length for stable compiled inference."""
+    if prev_actions.ndim != 2:
+        raise ValueError(f"Expected 2D [T, A] tensor, got shape={tuple(prev_actions.shape)}")
+    steps, action_dim = prev_actions.shape
+    if steps == target_steps:
+        return prev_actions
+    if steps > target_steps:
+        return prev_actions[:target_steps]
+    padded = torch.zeros((target_steps, action_dim), dtype=prev_actions.dtype, device=prev_actions.device)
+    padded[:steps] = prev_actions
+    return padded
+
+
+# ---------------------------------------------------------------------------
+# RTCInferenceEngine
+# ---------------------------------------------------------------------------
+
+
+class RTCInferenceEngine(InferenceEngine):
+    """Async RTC inference: a background thread produces action chunks.
+
+    ``get_action`` pops the next action from the shared queue (or
+    returns ``None`` if the queue is empty).  The main loop should call
+    ``notify_observation`` every tick and ``pause``/``resume`` around
+    human-intervention phases.
+    """
+
+    def __init__(
+        self,
+        policy: PreTrainedPolicy,
+        preprocessor: PolicyProcessorPipeline,
+        postprocessor: PolicyProcessorPipeline,
+        robot_wrapper: ThreadSafeRobot,
+        rtc_config: RTCConfig,
+        hw_features: dict,
+        task: str,
+        fps: float,
+        device: str | None,
+        use_torch_compile: bool = False,
+        compile_warmup_inferences: int = 2,
+        rtc_queue_threshold: int = 30,
+        shutdown_event: Event | None = None,
+    ) -> None:
+        self._policy = policy
+        self._preprocessor = preprocessor
+        self._postprocessor = postprocessor
+        self._robot = robot_wrapper
+        self._rtc_config = rtc_config
+        self._hw_features = hw_features
+        self._task = task
+        self._fps = fps
+        self._device = device or "cpu"
+        self._use_torch_compile = use_torch_compile
+        self._compile_warmup_inferences = compile_warmup_inferences
+        self._rtc_queue_threshold = rtc_queue_threshold
+
+        self._action_queue: ActionQueue | None = None
+        self._obs_holder: dict[str, Any] = {}
+        self._obs_lock = Lock()
+        self._policy_active = Event()
+        self._compile_warmup_done = Event()
+        self._shutdown_event = Event()
+        self._rtc_error = Event()
+        self._global_shutdown_event = shutdown_event
+        self._rtc_thread: Thread | None = None
+
+        if not self._use_torch_compile:
+            self._compile_warmup_done.set()
+            logger.info("RTCInferenceEngine initialized (torch.compile disabled, no warmup needed)")
+        else:
+            logger.info(
+                "RTCInferenceEngine initialized (torch.compile enabled, %d warmup inferences)",
+                compile_warmup_inferences,
+            )
+
+        # Processor introspection for relative-action re-anchoring.
+        self._relative_step = next(
+            (s for s in preprocessor.steps if isinstance(s, RelativeActionsProcessorStep) and s.enabled),
+            None,
+        )
+        self._normalizer_step = next(
+            (s for s in preprocessor.steps if isinstance(s, NormalizerProcessorStep)),
+            None,
+        )
+        if self._relative_step is not None:
+            if self._relative_step.action_names is None:
+                cfg_names = getattr(policy.config, "action_feature_names", None)
+                if cfg_names:
+                    self._relative_step.action_names = list(cfg_names)
+                else:
+                    self._relative_step.action_names = [
+                        k for k in robot_wrapper.action_features if k.endswith(".pos")
+                    ]
+            logger.info("Relative actions enabled: RTC prefix will be re-anchored")
+
+    # ------------------------------------------------------------------
+    # Lifecycle
+    # ------------------------------------------------------------------
+
+    @property
+    def ready(self) -> bool:
+        """True once torch.compile warmup is complete (or immediately if compile is disabled)."""
+        return self._compile_warmup_done.is_set()
+
+    @property
+    def failed(self) -> bool:
+        """True if the RTC background thread exited due to an unrecoverable error."""
+        return self._rtc_error.is_set()
+
+    @property
+    def action_queue(self) -> ActionQueue | None:
+        """The shared action queue between the RTC thread and the main loop."""
+        return self._action_queue
+
+    def start(self) -> None:
+        """Launch the RTC background thread."""
+        self._action_queue = ActionQueue(self._rtc_config)
+        self._obs_holder = {
+            "obs": None,
+            "robot_type": self._robot.robot_type,
+        }
+        self._shutdown_event.clear()
+        self._rtc_thread = Thread(
+            target=self._rtc_loop,
+            daemon=True,
+            name="RTCInference",
+        )
+        self._rtc_thread.start()
+        logger.info("RTC inference thread started")
+
+    def stop(self) -> None:
+        """Signal the RTC thread to stop and wait for it."""
+        logger.info("Stopping RTC inference thread...")
+        self._shutdown_event.set()
+        self._policy_active.clear()
+        if self._rtc_thread is not None and self._rtc_thread.is_alive():
+            self._rtc_thread.join(timeout=_RTC_JOIN_TIMEOUT_S)
+            if self._rtc_thread.is_alive():
+                logger.warning("RTC thread did not join within %.1fs", _RTC_JOIN_TIMEOUT_S)
+            else:
+                logger.info("RTC inference thread stopped")
+            self._rtc_thread = None
+
+    def pause(self) -> None:
+        """Pause the RTC background thread."""
+        logger.info("Pausing RTC inference thread")
+        self._policy_active.clear()
+
+    def resume(self) -> None:
+        """Resume the RTC background thread."""
+        logger.info("Resuming RTC inference thread")
+        self._policy_active.set()
+
+    def reset(self) -> None:
+        """Reset the policy, processors, and action queue."""
+        logger.info("Resetting RTC inference state (policy + processors + queue)")
+        self._policy.reset()
+        self._preprocessor.reset()
+        self._postprocessor.reset()
+        if self._action_queue is not None:
+            self._action_queue.clear()
+
+    # ------------------------------------------------------------------
+    # Action production (called from main thread)
+    # ------------------------------------------------------------------
+
+    def get_action(self, obs_frame: dict | None) -> torch.Tensor | None:
+        """Pop the next action from the RTC queue (ignores ``obs_frame``)."""
+        if self._action_queue is None:
+            return None
+        return self._action_queue.get()
+
+    def notify_observation(self, obs: dict) -> None:
+        """Publish the latest observation for the RTC thread to consume."""
+        with self._obs_lock:
+            self._obs_holder["obs"] = obs
+
+    # ------------------------------------------------------------------
+    # RTC: background inference thread
+    # ------------------------------------------------------------------
+
+    def _rtc_loop(self) -> None:
+        """Background thread that generates action chunks via RTC."""
+        try:
+            latency_tracker = LatencyTracker()
+            time_per_chunk = 1.0 / self._fps
+            policy_device = torch.device(self._device)
+
+            warmup_required = max(1, self._compile_warmup_inferences) if self._use_torch_compile else 0
+            inference_count = 0
+            consecutive_errors = 0
+
+            while not self._shutdown_event.is_set():
+                if not self._policy_active.is_set():
+                    time.sleep(_RTC_IDLE_SLEEP_S)
+                    continue
+
+                queue = self._action_queue
+                with self._obs_lock:
+                    obs = self._obs_holder.get("obs")
+                if queue is None or obs is None:
+                    time.sleep(_RTC_IDLE_SLEEP_S)
+                    continue
+
+                if queue.qsize() <= self._rtc_queue_threshold:
+                    try:
+                        current_time = time.perf_counter()
+                        idx_before = queue.get_action_index()
+                        prev_actions = queue.get_left_over()
+
+                        latency = latency_tracker.max()
+                        delay = math.ceil(latency / time_per_chunk) if latency else 0
+
+                        obs_batch = build_dataset_frame(self._hw_features, obs, prefix="observation")
+                        obs_batch = prepare_observation_for_inference(
+                            obs_batch, policy_device, self._task, self._robot.robot_type
+                        )
+                        obs_batch["task"] = [self._task]
+
+                        preprocessed = self._preprocessor(obs_batch)
+
+                        if prev_actions is not None and self._relative_step is not None:
+                            # Rebase against the raw cached state so the leftover tail stays in
+                            # the training-time coordinate frame.
+                            raw_state = self._relative_step.get_cached_state()
+                            if raw_state is not None:
+                                prev_abs = queue.get_processed_left_over()
+                                if prev_abs is not None and prev_abs.numel() > 0:
+                                    prev_actions = reanchor_relative_rtc_prefix(
+                                        prev_actions_absolute=prev_abs,
+                                        current_state=raw_state,
+                                        relative_step=self._relative_step,
+                                        normalizer_step=self._normalizer_step,
+                                        policy_device=policy_device,
+                                    )
+
+                        if prev_actions is not None:
+                            prev_actions = _normalize_prev_actions_length(
+                                prev_actions, target_steps=self._rtc_config.execution_horizon
+                            )
+
+                        actions = self._policy.predict_action_chunk(
+                            preprocessed, inference_delay=delay, prev_chunk_left_over=prev_actions
+                        )
+
+                        original = actions.squeeze(0).clone()
+                        processed = self._postprocessor(actions).squeeze(0)
+                        new_latency = time.perf_counter() - current_time
+                        new_delay = math.ceil(new_latency / time_per_chunk)
+
+                        inference_count += 1
+                        consecutive_errors = 0
+                        is_warmup = self._use_torch_compile and inference_count <= warmup_required
+                        if is_warmup:
+                            latency_tracker.reset()
+                        else:
+                            latency_tracker.add(new_latency)
+
+                        queue.merge(original, processed, new_delay, idx_before)
+
+                        if (
+                            is_warmup
+                            and inference_count >= warmup_required
+                            and not self._compile_warmup_done.is_set()
+                        ):
+                            self._compile_warmup_done.set()
+                            logger.info("Compile warmup complete (%d inferences)", inference_count)
+
+                        logger.debug("RTC inference latency=%.2fs, queue=%d", new_latency, queue.qsize())
+
+                    except Exception as e:
+                        consecutive_errors += 1
+                        logger.error(
+                            "RTC inference error (%d/%d): %s",
+                            consecutive_errors,
+                            _RTC_MAX_CONSECUTIVE_ERRORS,
+                            e,
+                        )
+                        logger.debug(traceback.format_exc())
+                        if consecutive_errors >= _RTC_MAX_CONSECUTIVE_ERRORS:
+                            # Persistent failure: stop retrying and propagate shutdown.
+                            raise
+                        time.sleep(_RTC_ERROR_RETRY_DELAY_S)
+                else:
+                    time.sleep(_RTC_IDLE_SLEEP_S)
+
+        except Exception as e:
+            logger.error("Fatal error in RTC thread: %s", e)
+            logger.error(traceback.format_exc())
+            self._rtc_error.set()
+            # Unblock any warmup waiters so the main loop doesn't spin forever
+            self._compile_warmup_done.set()
+            # Signal the top-level shutdown so strategies exit their control loops
+            if self._global_shutdown_event is not None:
+                self._global_shutdown_event.set()