improve script, time saving subtask array

Signed-off-by: Jade Choghari <chogharijade@gmail.com>

.github/workflows/full_tests.yml

@@ -101,11 +101,9 @@ jobs:
     runs-on:
       group: aws-general-8-plus
     if: |
-      github.repository == 'huggingface/lerobot' && (
+      (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
-        (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
+      github.event_name == 'push' ||
-        github.event_name == 'push' ||
+      github.event_name == 'workflow_dispatch'
-        github.event_name == 'workflow_dispatch'
-      )
     outputs:
       image_tag: ${{ steps.set_tag.outputs.image_tag }}
     env:
@@ -173,8 +171,6 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
-      - name: Fix ptxas permissions
-        run: chmod +x /lerobot/.venv/lib/python3.10/site-packages/triton/backends/nvidia/bin/ptxas
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests

.github/workflows/unbound_deps_tests.yml

@@ -91,7 +91,6 @@ jobs:
     name: Build and Push Docker
     runs-on:
       group: aws-general-8-plus
-    if: github.repository == 'huggingface/lerobot'
     outputs:
       image_tag: ${{ env.DOCKER_IMAGE_NAME }}
     env:

MANIFEST.in

@@ -1,3 +1,2 @@
 include src/lerobot/templates/lerobot_modelcard_template.md
 include src/lerobot/datasets/card_template.md
-include src/lerobot/envs/metaworld_config.json

benchmarks/video/README.md

@@ -28,9 +28,9 @@ We don't expect the same optimal settings for a dataset of images from a simulat
 For these reasons, we run this benchmark on four representative datasets:
 
 - `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
-- `lerobot/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
+- `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
-- `lerobot/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
+- `aliberts/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
-- `lerobot/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
+- `aliberts/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
 
 Note: The datasets used for this benchmark need to be image datasets, not video datasets.
 
@@ -179,7 +179,7 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
+        aliberts/aloha_mobile_shrimp_image \
     --vcodec libx264 libx265 \
     --pix-fmt yuv444p yuv420p \
     --g 2 20 None \
@@ -203,9 +203,9 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
+        aliberts/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
+        aliberts/paris_street \
-        lerobot/kitchen \
+        aliberts/kitchen \
     --vcodec libx264 libx265 \
     --pix-fmt yuv444p yuv420p \
     --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -221,9 +221,9 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
+        aliberts/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
+        aliberts/paris_street \
-        lerobot/kitchen \
+        aliberts/kitchen \
     --vcodec libsvtav1 \
     --pix-fmt yuv420p \
     --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -252,37 +252,37 @@ Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_read
 
 These tables show the results for `g=2` and `crf=30`, using `timestamps-modes=6_frames` and `backend=pyav`
 
-| video_images_size_ratio           | vcodec     | pix_fmt |           |           |           |
+| video_images_size_ratio            | vcodec     | pix_fmt |           |           |           |
-| --------------------------------- | ---------- | ------- | --------- | --------- | --------- |
+| ---------------------------------- | ---------- | ------- | --------- | --------- | --------- |
-|                                   | libx264    |         | libx265   |           | libsvtav1 |
+|                                    | libx264    |         | libx265   |           | libsvtav1 |
-| repo_id                           | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
+| repo_id                            | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
-| lerobot/pusht_image               | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
+| lerobot/pusht_image                | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
-| lerobot/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
+| aliberts/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
-| lerobot/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
+| aliberts/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
-| lerobot/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
+| aliberts/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
 
-| video_images_load_time_ratio      | vcodec  | pix_fmt |          |         |           |
+| video_images_load_time_ratio       | vcodec  | pix_fmt |          |         |           |
-| --------------------------------- | ------- | ------- | -------- | ------- | --------- |
+| ---------------------------------- | ------- | ------- | -------- | ------- | --------- |
-|                                   | libx264 |         | libx265  |         | libsvtav1 |
+|                                    | libx264 |         | libx265  |         | libsvtav1 |
-| repo_id                           | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
+| repo_id                            | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
-| lerobot/pusht_image               | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
+| lerobot/pusht_image                | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
-| lerobot/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
+| aliberts/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
-| lerobot/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
+| aliberts/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
-| lerobot/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
+| aliberts/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
 
-|                                   |          | vcodec   | pix_fmt      |          |           |              |
+|                                    |          | vcodec   | pix_fmt      |          |           |              |
-| --------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
+| ---------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
-|                                   |          | libx264  |              | libx265  |           | libsvtav1    |
+|                                    |          | libx264  |              | libx265  |           | libsvtav1    |
-| repo_id                           | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
+| repo_id                            | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
-| lerobot/pusht_image               | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
+| lerobot/pusht_image                | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
-|                                   | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
+|                                    | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
-|                                   | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
+|                                    | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
-| lerobot/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
+| aliberts/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
-|                                   | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
+|                                    | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
-|                                   | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
+|                                    | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
-| lerobot/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
+| aliberts/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
-|                                   | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
+|                                    | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
-|                                   | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
+|                                    | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
-| lerobot/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
+| aliberts/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
-|                                   | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
+|                                    | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
-|                                   | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |
+|                                    | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |

docker/Dockerfile.internal

@@ -85,8 +85,6 @@ RUN if [ "$UNBOUND_DEPS" = "true" ]; then \
 
 RUN uv pip install --no-cache ".[all]"
 
-RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
-
 # Copy the rest of the application source code
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .

docs/source/_toctree.yml

@@ -27,10 +27,10 @@
     title: Porting Large Datasets
   - local: using_dataset_tools
     title: Using the Dataset Tools
+  - local: annotation_tools
+    title: Using the Annotation Tools
   - local: dataset_subtask
     title: Using Subtasks in the Dataset
-  - local: streaming_video_encoding
-    title: Streaming Video Encoding
   title: "Datasets"
 - sections:
   - local: act

docs/source/act.mdx

@@ -88,8 +88,5 @@ lerobot-record \
   --dataset.repo_id=${HF_USER}/eval_act_your_dataset \
   --dataset.num_episodes=10 \
   --dataset.single_task="Your task description" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   --policy.path=${HF_USER}/act_policy
 ```

docs/source/annotation_tools.mdx

@@ -0,0 +1,425 @@
+# Dataset Annotation Tools
+
+This guide explains how to use the automatic annotation tools to add skill labels and synthetic dialogue to your LeRobot datasets.
+
+## Overview
+
+The annotation pipeline consists of two main components:
+
+1. **Subtask Annotation** (`subtask_annotate.py`): Automatically segments robot demonstrations into atomic skills using Vision-Language Models (VLMs)
+2. **High-Level Annotation** (`high_level_annotate.py`): Generates synthetic user prompts and robot utterances for hierarchical policy training
+
+These tools enable you to transform raw robot demonstration data into richly annotated datasets suitable for training hierarchical policies.
+
+## Installation Requirements
+
+Before using the annotation tools, ensure you have the required dependencies:
+
+```bash
+pip install transformers qwen-vl-utils opencv-python rich pandas pyarrow
+```
+
+You'll also need FFmpeg for video processing:
+
+```bash
+# Ubuntu/Debian
+sudo apt-get install ffmpeg
+
+# macOS
+brew install ffmpeg
+```
+
+## Part 1: Subtask Annotation
+
+### What It Does
+
+The subtask annotator segments each episode into short atomic manipulation skills (1-3 seconds each). For example, a "pick and place" episode might be segmented into:
+- "reach towards object" (0.0s - 1.2s)
+- "grasp object" (1.2s - 2.1s)
+- "lift object" (2.1s - 3.5s)
+- "move to target" (3.5s - 5.0s)
+- "release object" (5.0s - 6.2s)
+
+### Usage
+
+#### Basic Example
+
+```bash
+python src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+    --repo-id your-username/your-dataset \
+    --video-key observation.images.base \
+    --output-dir /path/to/output
+```
+
+#### With Local Dataset
+
+```bash
+python src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+    --data-dir /path/to/local/dataset \
+    --video-key observation.images.base \
+    --output-dir /path/to/output
+```
+
+#### Advanced Options
+
+```bash
+python src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+    --repo-id your-username/your-dataset \
+    --video-key observation.images.base \
+    --model Qwen/Qwen2-VL-7B-Instruct \
+    --batch-size 16 \
+    --output-dir /path/to/output \
+    --push-to-hub
+```
+
+### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `--repo-id` | HuggingFace Hub dataset ID | Required (or use --data-dir) |
+| `--data-dir` | Path to local dataset | Required (or use --repo-id) |
+| `--video-key` | Video observation key | Required |
+| `--model` | VLM model to use | `Qwen/Qwen2-VL-7B-Instruct` |
+| `--device` | Device to run model on | `cuda` |
+| `--dtype` | Model dtype | `bfloat16` |
+| `--batch-size` | Episodes per batch | `8` |
+| `--episodes` | Specific episodes to annotate | All episodes |
+| `--output-dir` | Output directory | Auto-generated |
+| `--push-to-hub` | Push to HuggingFace Hub | `False` |
+
+### Supported Models
+
+- **Qwen2-VL**: `Qwen/Qwen2-VL-2B-Instruct`, `Qwen/Qwen2-VL-7B-Instruct`, `Qwen/Qwen2-VL-72B-Instruct`
+- **Qwen3-VL**: `Qwen/Qwen3-VL-30B-A3B-Instruct`
+
+### Output Files
+
+The subtask annotation creates the following files in your dataset:
+
+1. **`meta/subtasks.parquet`**: DataFrame with unique subtask names
+   ```python
+   # Structure:
+   # Index: subtask name (string)
+   # Column: subtask_index (int64)
+   ```
+
+2. **`meta/skills.json`**: Raw skill annotations with timestamps
+   ```json
+   {
+     "coarse_description": "Pick and place the object",
+     "skill_to_subtask_index": {
+       "reach towards object": 0,
+       "grasp object": 1,
+       ...
+     },
+     "episodes": {
+       "0": {
+         "episode_index": 0,
+         "description": "Pick and place the object",
+         "skills": [
+           {"name": "reach towards object", "start": 0.0, "end": 1.2},
+           {"name": "grasp object", "start": 1.2, "end": 2.1},
+           ...
+         ]
+       }
+     }
+   }
+   ```
+
+3. **`subtask_index` feature**: Added to each frame in the dataset
+   - Type: `int64`
+   - Shape: `(1,)`
+   - Maps each frame to its corresponding subtask
+
+### Accessing Subtask Annotations
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Load annotated dataset
+dataset = LeRobotDataset(repo_id="your/dataset_with_subtasks")
+
+# Get a frame
+frame = dataset[100]
+
+# Get the subtask for this frame
+subtask_idx = frame["subtask_index"].item()
+subtask_name = dataset.meta.subtasks.iloc[subtask_idx].name
+
+print(f"Frame 100 is performing: {subtask_name}")
+
+# Load all subtasks
+subtasks_df = dataset.meta.subtasks
+print(subtasks_df)
+```
+
+## Part 2: High-Level Annotation
+
+### What It Does
+
+The high-level annotator generates synthetic dialogue for hierarchical policy training. For each skill, it creates:
+- **User Prompt** (`ℓ_t`): A natural language request from the user
+- **Robot Utterance** (`u_t`): A natural language response from the robot
+
+This enables training policies that can understand and respond to human instructions in natural dialogue.
+
+### Prerequisites
+
+**Important**: You must run subtask annotation first! High-level annotation requires the `skills.json` file generated by subtask annotation.
+
+### Usage
+
+#### Image Mode (Default)
+
+Samples frames at regular intervals and passes images to the VLM:
+
+```bash
+python src/lerobot/policies/pi05_full/annotate/high_level_annotate.py \
+    --repo-id your/dataset_with_subtasks \
+    --model Qwen/Qwen2-VL-7B-Instruct \
+    --image-key observation.images.base \
+    --output-dir /path/to/output
+```
+
+#### Video Mode
+
+Passes entire episode videos to the VLM for better temporal understanding:
+
+```bash
+python src/lerobot/policies/pi05_full/annotate/high_level_annotate.py \
+    --repo-id your/dataset_with_subtasks \
+    --model Qwen/Qwen2-VL-7B-Instruct \
+    --video-mode \
+    --video-key observation.images.base \
+    --video-batch-size 4 \
+    --output-dir /path/to/output
+```
+
+### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `--repo-id` | HuggingFace Hub dataset ID | Required (or use --data-dir) |
+| `--data-dir` | Path to local dataset | Required (or use --repo-id) |
+| `--model` | VLM model to use | `Qwen/Qwen2-VL-7B-Instruct` |
+| `--image-key` | Image observation key (image mode) | First camera key |
+| `--video-mode` | Use video instead of images | `False` |
+| `--video-key` | Video observation key (video mode) | Auto-detected |
+| `--video-batch-size` | Episodes per batch (video mode) | `1` |
+| `--sample-interval` | Sampling interval in seconds | `1.0` |
+| `--temperature` | Sampling temperature | `0.7` |
+| `--output-dir` | Output directory | Auto-generated |
+| `--push-to-hub` | Push to HuggingFace Hub | `False` |
+
+### Output Files
+
+The high-level annotation creates:
+
+1. **`meta/tasks_high_level.parquet`**: DataFrame with high-level tasks
+   ```python
+   # Structure:
+   # Index: task string (concatenated user_prompt | robot_utterance)
+   # Columns:
+   #   - task_index: int64
+   #   - user_prompt: string
+   #   - robot_utterance: string
+   #   - skill: string (associated subtask)
+   #   - scenario_type: string
+   #   - response_type: string
+   ```
+
+2. **`meta/syn_annotations.jsonl`**: Debug annotations (JSONL format)
+   ```json
+   {"episode_id": 0, "timestamp": 1.5, "skill_current": "grasp object", "user_prompt": "Can you pick that up?", "robot_utterance": "Sure, I'll grasp it now", ...}
+   ```
+
+3. **`task_index_high_level` feature**: Added to each frame
+   - Type: `int64`
+   - Shape: `(1,)`
+   - Maps each frame to its high-level task
+
+### Dialogue Types Generated
+
+The system generates diverse interaction types:
+
+**Scenario Types:**
+- `specific_object`: "Pick up the red block"
+- `negative_task`: "Don't touch the blue one"
+- `situated_correction`: "Actually, move to the other box instead"
+- `implicit_request`: "I need something red for the tower"
+- `constraint_based`: "Make sure to handle it gently"
+
+**Response Types:**
+- `confirmation`: "OK, I'll pick it up"
+- `clarification`: "Just to confirm, you want me to pick up the red block?"
+- `acknowledgment`: "Got it, picking up the red block"
+- `constraint_acknowledgment`: "Sure, I'll pick it up gently"
+
+### Accessing High-Level Annotations
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+import pandas as pd
+
+# Load annotated dataset
+dataset = LeRobotDataset(repo_id="your/dataset_with_high_level_tasks")
+
+# Get a frame
+frame = dataset[100]
+
+# Get the high-level task
+task_idx = frame["task_index_high_level"].item()
+
+# Load tasks metadata
+tasks_df = pd.read_parquet(dataset.root / "meta" / "tasks_high_level.parquet")
+task_row = tasks_df[tasks_df["task_index"] == task_idx].iloc[0]
+
+print(f"User: {task_row['user_prompt']}")
+print(f"Robot: {task_row['robot_utterance']}")
+print(f"Skill: {task_row['skill']}")
+
+# Use in a DataLoader
+import torch
+from torch.utils.data import DataLoader
+
+dataloader = DataLoader(dataset, batch_size=32, shuffle=True)
+batch = next(iter(dataloader))
+
+print(f"Task indices: {batch['task_index_high_level']}")
+print(f"User prompts: {batch['user_prompt'][0]}")
+print(f"Robot utterances: {batch['robot_utterance'][0]}")
+```
+
+## Complete Pipeline Example
+
+Here's how to run both annotation stages:
+
+```bash
+#!/bin/bash
+
+REPO_ID="your-username/your-dataset"
+MODEL="Qwen/Qwen2-VL-7B-Instruct"
+OUTPUT_DIR="/path/to/output"
+
+# Step 1: Subtask Annotation
+python src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+    --repo-id "$REPO_ID" \
+    --video-key observation.images.base \
+    --model "$MODEL" \
+    --batch-size 8 \
+    --output-dir "${OUTPUT_DIR}/subtasks"
+
+# Step 2: High-Level Annotation (Image Mode)
+python src/lerobot/policies/pi05_full/annotate/high_level_annotate.py \
+    --data-dir "${OUTPUT_DIR}/subtasks" \
+    --model "$MODEL" \
+    --image-key observation.images.base \
+    --sample-interval 1.0 \
+    --output-dir "${OUTPUT_DIR}/final"
+
+# Or Step 2: High-Level Annotation (Video Mode - Recommended)
+python src/lerobot/policies/pi05_full/annotate/high_level_annotate.py \
+    --data-dir "${OUTPUT_DIR}/subtasks" \
+    --model "$MODEL" \
+    --video-mode \
+    --video-key observation.images.base \
+    --video-batch-size 4 \
+    --output-dir "${OUTPUT_DIR}/final"
+```
+
+## Performance Tips
+
+### For Faster Processing
+
+1. **Increase batch size**: Use `--batch-size 16` or higher (subtask annotation)
+2. **Increase video batch size**: Use `--video-batch-size 8` (high-level annotation in video mode)
+3. **Larger sampling interval**: Use `--sample-interval 5.0` for testing (samples every 5 seconds instead of 1)
+4. **Use smaller models**: `Qwen/Qwen2-VL-2B-Instruct` is faster than `Qwen2-VL-7B-Instruct`
+5. **Process specific episodes**: Use `--episodes 0 1 2 3` to annotate only a subset
+
+### For Better Quality
+
+1. **Use larger models**: `Qwen/Qwen3-VL-30B-A3B-Instruct` or `Qwen/Qwen2-VL-72B-Instruct`
+2. **Use video mode**: Provides better temporal context
+3. **Smaller sampling intervals**: `--sample-interval 0.5` for dense annotations
+4. **Adjust temperature**: Use `--temperature 0.9` for more diverse dialogue
+
+## Memory Requirements
+
+| Model | GPU Memory | Recommended Batch Size |
+|-------|------------|------------------------|
+| Qwen2-VL-2B | ~8 GB | 16-32 |
+| Qwen2-VL-7B | ~16 GB | 8-16 |
+| Qwen2-VL-72B | ~80 GB | 1-2 |
+| Qwen3-VL-30B | ~40 GB | 4-8 |
+
+## Troubleshooting
+
+### "FFmpeg not found"
+```bash
+# Install FFmpeg
+sudo apt-get install ffmpeg  # Ubuntu/Debian
+brew install ffmpeg          # macOS
+```
+
+### "CUDA out of memory"
+- Reduce batch size: `--batch-size 1` or `--video-batch-size 1`
+- Use smaller model: `Qwen/Qwen2-VL-2B-Instruct`
+- Use CPU: `--device cpu` (much slower)
+
+### "No skills.json found"
+Run subtask annotation first before high-level annotation.
+
+### "Video key not found"
+List available keys:
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+dataset = LeRobotDataset(repo_id="your/dataset")
+print("Video keys:", dataset.meta.video_keys)
+print("Camera keys:", dataset.meta.camera_keys)
+```
+
+## Dataset Structure After Annotation
+
+```
+your_dataset_with_high_level_tasks/
+├── meta/
+│   ├── info.json                      # Original metadata
+│   ├── tasks.parquet                  # Original tasks (preserved)
+│   ├── subtasks.parquet              # NEW: Subtask names and indices
+│   ├── skills.json                    # NEW: Raw skill annotations with timestamps
+│   ├── tasks_high_level.parquet      # NEW: High-level tasks with dialogue
+│   └── syn_annotations.jsonl         # NEW: Debug annotations
+├── data/
+│   └── chunk-000/
+│       ├── observation.images.base.mp4
+│       ├── action.safetensors
+│       ├── subtask_index.safetensors # NEW: Subtask per frame
+│       └── task_index_high_level.safetensors # NEW: High-level task per frame
+└── videos/
+    └── ...
+```
+
+## Citation
+
+If you use these annotation tools in your research, please cite:
+
+```bibtex
+@article{lerobot2024,
+  title={LeRobot: State-of-the-art Machine Learning for Real-World Robotics},
+  author={LeRobot Contributors},
+  year={2024},
+  url={https://github.com/huggingface/lerobot}
+}
+```
+
+## Next Steps
+
+After annotation, you can:
+1. Train hierarchical policies using the subtask and high-level annotations
+2. Use the synthetic dialogue for instruction-following policy training
+3. Analyze skill distributions and dialogue patterns
+4. Share your annotated dataset on HuggingFace Hub with `--push-to-hub`
+
+For training examples, see the [training documentation](../training/).
+

docs/source/earthrover_mini_plus.mdx

@@ -185,16 +185,13 @@ echo $HF_USER
 Use the standard recording command:
 
 ```bash
-lerobot-record \
+python src/lerobot/scripts/lerobot_record.py \
     --robot.type=earthrover_mini_plus \
     --teleop.type=keyboard_rover \
     --dataset.repo_id=your_username/dataset_name \
     --dataset.num_episodes=2 \
     --dataset.fps=10 \
     --dataset.single_task="Navigate around obstacles" \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 

docs/source/groot.mdx

@@ -120,12 +120,9 @@ lerobot-record \
   --display_data=true \
   --dataset.repo_id=<user>/eval_groot-bimanual  \
   --dataset.num_episodes=10 \
-  --dataset.single_task="Grab and handover the red cube to the other arm" \
+  --dataset.single_task="Grab and handover the red cube to the other arm"
-  --dataset.streaming_encoding=true \
+  --policy.path=<user>/groot-bimanual # your trained model
-  --dataset.encoder_threads=2 \
+  --dataset.episode_time_s=30
-  # --dataset.vcodec=auto \
-  --policy.path=<user>/groot-bimanual \ # your trained model
-  --dataset.episode_time_s=30 \
   --dataset.reset_time_s=10
 ```
 

docs/source/hope_jr.mdx

@@ -224,15 +224,12 @@ lerobot-record \
     --teleop.port=/dev/tty.usbmodem1201 \
     --teleop.id=right \
     --teleop.side=right \
-    --dataset.repo_id=<USER>/hand_record_test_with_video_data \
+    --dataset.repo_id=nepyope/hand_record_test_with_video_data \
     --dataset.single_task="Hand recording test with video data" \
     --dataset.num_episodes=1 \
     --dataset.episode_time_s=5 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -244,7 +241,7 @@ lerobot-replay \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=right \
     --robot.side=right \
-    --dataset.repo_id=<USER>/hand_record_test_with_camera \
+    --dataset.repo_id=nepyope/hand_record_test_with_camera \
     --dataset.episode=0
 ```
 
@@ -252,13 +249,13 @@ lerobot-replay \
 
 ```bash
 lerobot-train \
-  --dataset.repo_id=<USER>/hand_record_test_with_video_data \
+  --dataset.repo_id=nepyope/hand_record_test_with_video_data \
   --policy.type=act \
   --output_dir=outputs/train/hopejr_hand \
   --job_name=hopejr \
   --policy.device=mps \
   --wandb.enable=true \
-  --policy.repo_id=<USER>/hand_test_policy
+  --policy.repo_id=nepyope/hand_test_policy
 ```
 
 ### Evaluate
@@ -273,11 +270,8 @@ lerobot-record \
   --robot.side=right \
   --robot.cameras='{"main": {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}}' \
   --display_data=false \
-  --dataset.repo_id=<USER>/eval_hopejr \
+  --dataset.repo_id=nepyope/eval_hopejr \
   --dataset.single_task="Evaluate hopejr hand policy" \
   --dataset.num_episodes=10 \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
 ```

docs/source/il_robots.mdx

@@ -165,7 +165,7 @@ huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(hf auth whoami | head -n 1)
 echo $HF_USER
 ```
 
@@ -185,10 +185,7 @@ lerobot-record \
     --display_data=true \
     --dataset.repo_id=${HF_USER}/record-test \
     --dataset.num_episodes=5 \
-    --dataset.single_task="Grab the black cube" \
+    --dataset.single_task="Grab the black cube"
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
 ```
 </hfoption>
 <hfoption id="API example">
@@ -518,9 +515,6 @@ lerobot-record  \
   --display_data=false \
   --dataset.repo_id=${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 \

docs/source/installation.mdx

@@ -1,15 +1,13 @@
 # Installation
 
-This guide uses conda (via miniforge) to manage environments. If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.10 and ffmpeg installed with the `libsvtav1` encoder, then skip ahead to [Install LeRobot](#step-3-install-lerobot-).
+## Install [`miniforge`](https://conda-forge.org/download/)
-
-## Step 1: Install [`miniforge`](https://conda-forge.org/download/)
 
 ```bash
 wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
 bash Miniforge3-$(uname)-$(uname -m).sh
 ```
 
-## Step 2: Environment Setup
+## Environment Setup
 
 Create a virtual environment with Python 3.10, using conda:
 
@@ -40,14 +38,7 @@ conda install ffmpeg -c conda-forge
 >
 > - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
 
-> [!NOTE]
+## Install LeRobot 🤗
-> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to install `evdev` with the following command:
->
-> ```bash
-> conda install evdev -c conda-forge
-> ```
-
-## Step 3: Install LeRobot 🤗
 
 ### From Source
 

docs/source/lerobot-dataset-v3.mdx

@@ -41,10 +41,7 @@ lerobot-record \
   --display_data=true \
   --dataset.repo_id=${HF_USER}/record-test \
   --dataset.num_episodes=5 \
-  --dataset.single_task="Grab the black cube" \
+  --dataset.single_task="Grab the black cube"
-  --dataset.streaming_encoding=true \
-  # --dataset.vcodec=auto \
-  --dataset.encoder_threads=2
 ```
 
 See the [recording guide](./il_robots#record-a-dataset) for more details.

docs/source/phone_teleop.mdx

@@ -66,13 +66,12 @@ Run on of the examples scripts to teleoperate, record a dataset, replay a datase
 
 All scripts assume you configured your robot (e.g., SO-100 follower) and set the correct serial port.
 
-Additionally you need to **copy the URDF of the robot into the examples folder**. For the examples in this tutorial (using SO100/SO101), copy the `SO101` folder from the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101) into the `examples/phone_to_so100/` directory, so that the URDF file path becomes `examples/phone_to_so100/SO101/so101_new_calib.urdf`.
+Additionally you need to **copy the urdf of the robot to the examples folder**. For the examples in this tutorial (Using SO100/SO101) it is highly recommended to use the urdf in the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf)
 
 - Run this example to teleoperate:
 
   ```bash
-  cd examples/phone_to_so100
+  python examples/phone_to_so100/teleoperate.py
-  python teleoperate.py
   ```
 
 After running the example:
@@ -85,22 +84,19 @@ Additionally you can customize mapping or safety limits by editing the processor
 - Run this example to record a dataset, which saves absolute end effector observations and actions:
 
   ```bash
-  cd examples/phone_to_so100
+  python examples/phone_to_so100/record.py
-  python record.py
   ```
 
 - Run this example to replay recorded episodes:
 
   ```bash
-  cd examples/phone_to_so100
+  python examples/phone_to_so100/replay.py
-  python replay.py
   ```
 
 - Run this example to evaluate a pretrained policy:
 
   ```bash
-  cd examples/phone_to_so100
+  python examples/phone_to_so100/evaluate.py
-  python evaluate.py
   ```
 
 ### Important pipeline steps and options

docs/source/pi0.mdx

@@ -60,7 +60,7 @@ policy.type=pi0
 For training π₀, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
     --dataset.repo_id=your_dataset \
     --policy.type=pi0 \
     --output_dir=./outputs/pi0_training \

docs/source/pi05.mdx

@@ -56,7 +56,7 @@ policy.type=pi05
 Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py\
     --dataset.repo_id=your_dataset \
     --policy.type=pi05 \
     --output_dir=./outputs/pi05_training \

docs/source/reachy2.mdx

@@ -159,9 +159,6 @@ lerobot-record \
     --dataset.fps=15 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -201,9 +198,6 @@ lerobot-record \
     --dataset.fps=15 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 

docs/source/sarm.mdx

@@ -269,7 +269,7 @@ This generates visualizations showing video frames with subtask boundaries overl
 Train with **no annotations** - uses linear progress from 0 to 1:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=single_stage \
@@ -288,7 +288,7 @@ lerobot-train \
 Train with **dense annotations only** (sparse auto-generated):
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=dense_only \
@@ -307,7 +307,7 @@ lerobot-train \
 Train with **both sparse and dense annotations**:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=dual \
@@ -468,7 +468,7 @@ This script:
 Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`). Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
   --use_rabc=true \

docs/source/smolvla.mdx

@@ -106,9 +106,6 @@ lerobot-record \
   --dataset.repo_id=${HF_USER}/eval_DATASET_NAME_test \  # <- This will be the dataset name on HF Hub
   --dataset.episode_time_s=50 \
   --dataset.num_episodes=10 \
-  --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 \

docs/source/streaming_video_encoding.mdx

@@ -1,155 +0,0 @@
-# Streaming Video Encoding Guide
-
-## 1. Overview
-
-Streaming video encoding eliminates the traditional PNG round-trip during video dataset recording. Instead of:
-
-1. Capture frame -> write PNG to disk -> (at episode end) read PNG's -> encode to MP4 -> delete PNG's
-
-Frames can be encoded in real-time during capture:
-
-1. Capture frame -> queue to encoder thread -> encode to MP4 directly
-
-This makes `save_episode()` near-instant (the video is already encoded by the time the episode ends) and removes the blocking wait that previously occurred between episodes, especially with multiple cameras in long episodes.
-
-## 2. Tuning Parameters
-
-| Parameter               | CLI Flag                          | Type          | Default       | Description                                                       |
-| ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- |
-| `streaming_encoding`    | `--dataset.streaming_encoding`    | `bool`        | `True`        | Enable real-time encoding during capture                          |
-| `vcodec`                | `--dataset.vcodec`                | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
-| `encoder_threads`       | `--dataset.encoder_threads`       | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide |
-| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `60`          | Max buffered frames per camera (~2s at 30fps). Consumes RAM       |
-
-## 3. Performance Considerations
-
-Streaming encoding means the CPU is encoding video **during** the capture loop, not after. This creates a CPU budget that must be shared between:
-
-- **Control loop** (reading cameras, control the robot, writing non-video data)
-- **Encoder threads** (one pool per camera)
-- **Rerun visualization** (if enabled)
-- **OS and other processes**
-
-### Resolution & Number of Cameras Impact
-
-| Setup                     | Throughput (px/sec) | CPU Encoding Load | Notes                          |
-| ------------------------- | ------------------- | ----------------- | ------------------------------ |
-| 2camsx 640x480x3 @30fps   | 55M                 | Low               | Works on most systems          |
-| 2camsx 1280x720x3 @30fps  | 165M                | Moderate          | Comfortable on modern systems  |
-| 2camsx 1920x1080x3 @30fps | 373M                | High              | Requires powerful high-end CPU |
-
-### `encoder_threads` Tuning
-
-This parameter controls how many threads each encoder instance uses internally:
-
-- **Higher values** (e.g., 4-5): Faster encoding, but uses more CPU cores per camera. Good for high-end systems with many cores.
-- **Lower values** (e.g., 1-2): Less CPU per camera, freeing cores for capture and visualization. Good for low-res images and capable CPUs.
-- **`None` (default)**: Lets the codec decide. Information available in the codec logs.
-
-### Backpressure and Frame Dropping
-
-Each camera has a bounded queue (`encoder_queue_maxsize`, default 60 frames). When the encoder can't keep up:
-
-1. The queue fills up (consuming RAM)
-2. New frames are **dropped** (not blocked) — the capture loop continues uninterrupted
-3. A warning is logged: `"Encoder queue full for {camera}, dropped N frame(s)"`
-4. At episode end, total dropped frames per camera are reported
-
-### Symptoms of Encoder Falling Behind
-
-- **System feels laggy and freezes**: all CPUs are at 100%
-- **Dropped frame warnings** in the log or lower frames/FPS than expected in the recorded dataset
-- **Choppy robot movement**: If CPU is severely overloaded, even the capture loop may be affected
-- **Accumulated rerun lag**: Visualization falls behind real-time
-
-## 4. Hardware-Accelerated Encoding
-
-### When to Use
-
-Use HW encoding when:
-
-- CPU is the bottleneck (dropped frames, choppy robot, rerun lag)
-- You have compatible hardware (GPU or dedicated encoder)
-- You're recording at high throughput (high resolution or with many cameras)
-
-### Choosing a Codec
-
-| Codec                 | CPU Usage | File Size      | Quality | Notes                                                            |
-| --------------------- | --------- | -------------- | ------- | ---------------------------------------------------------------- |
-| `libsvtav1` (default) | High      | Smallest       | Best    | Default. Best compression but most CPU-intensive                 |
-| `h264`                | Medium    | ~30-50% larger | Good    | Software H.264. Lower CPU                                        |
-| HW encoders           | Very Low  | Largest        | Good    | Offloads to dedicated hardware. Best for CPU-constrained systems |
-
-### Available HW Encoders
-
-| Encoder             | Platform      | Hardware                                                                                         | CLI Value                            |
-| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------ |
-| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=h264_videotoolbox` |
-| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=hevc_videotoolbox` |
-| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=h264_nvenc`        |
-| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=hevc_nvenc`        |
-| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.vcodec=h264_vaapi`        |
-| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.vcodec=h264_qsv`          |
-| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.vcodec=auto`              |
-
-> [!NOTE]
-> In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
-
-> [!NOTE]
-> `libsvtav1` is the default because it provides the best training performance; other vcodecs can reduce CPU usage and be faster, but they typically produce larger files and may affect training time.
-
-## 5. Troubleshooting
-
-| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                  |
-| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.vcodec=auto`) |
-| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.vcodec=auto`).                                                                                                                                                    |
-| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                     |
-| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                              |
-| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                              |
-| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.vcodec=auto`                                                                                                                                                                                                                             |
-| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                   |
-
-## 6. Recommended Configurations
-
-These estimates are conservative; we recommend testing them on your setup—start with a low load and increase it gradually.
-
-### High-End Systems: modern 12+ cores (24+ threads)
-
-A throughput between ~250-500M px/sec should be comfortable in CPU. For even better results try HW encoding if available.
-
-```bash
-# 3camsx 1280x720x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
-# 2camsx 1920x1080x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
-lerobot-record --dataset.encoder_threads=5 ...
-
-# 3camsx 1920x1080x3 @30fps: Might require some tuning.
-```
-
-### Mid-Range Systems: modern 8+ cores (16+ threads) or Apple Silicon
-
-A throughput between ~80-300M px/sec should be possible in CPU.
-
-```bash
-# 3camsx 640x480x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
-# 2camsx 1280x720x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
-lerobot-record --dataset.encoder_threads=2 ...
-
-# 2camsx 1920x1080x3 @30fps: Might require some tuning.
-```
-
-### Low-Resource Systems: modern 4+ cores (8+ threads) or Raspberry Pi 5
-
-On very constrained systems, streaming encoding may compete too heavily with the capture loop. Disabling it falls back to the PNG-based approach where encoding happens between episodes (blocking, but doesn't interfere with capture). Alternatively, record at a lower throughput to reduce both capture and encoding load. Consider also changing codec to `h264` and using batch encoding.
-
-```bash
-# 2camsx 640x480x3 @30fps: Requires some tuning.
-
-# Use H.264, disable streaming, consider batching encoding
-lerobot-record --dataset.vcodec=h264 --dataset.streaming_encoding=false ...
-```
-
-## 7. Closing note
-
-Performance ultimately depends on your exact setup — frames-per-second, resolution, CPU cores and load, available memory, episode length, and the encoder you choose. Always test with your target workload, be mindful about your CPU & system capabilities and tune `encoder_threads`, `encoder_queue_maxsize`, and
-`vcodec` reasonably. That said, a common practical configuration (for many applications) is three cameras at 640×480x3 @30fps; this usually runs fine with the default streaming video encoding settings in modern systems. Always verify your recorded dataset is healthy by comparing the video duration to the CLI episode duration and confirming the row count equals FPS × CLI duration.

docs/source/unitree_g1.mdx

@@ -216,7 +216,7 @@ lerobot-teleoperate \
 ### Record Dataset in Simulation
 
 ```bash
-lerobot-record \
+python -m lerobot.scripts.lerobot_record \
     --robot.type=unitree_g1 \
     --robot.is_simulation=true \
     --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
@@ -229,10 +229,7 @@ lerobot-record \
     --dataset.num_episodes=2 \
     --dataset.episode_time_s=5 \
     --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true \
+    --dataset.push_to_hub=true
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
 ```
 
 Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)
@@ -269,7 +266,7 @@ lerobot-teleoperate \
 ### Record Dataset on Real Robot
 
 ```bash
-lerobot-record \
+python -m lerobot.scripts.lerobot_record \
     --robot.type=unitree_g1 \
     --robot.is_simulation=false \
     --robot.cameras='{"global_view": {"type": "zmq", "server_address": "172.18.129.215", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
@@ -282,10 +279,7 @@ lerobot-record \
     --dataset.num_episodes=2 \
     --dataset.episode_time_s=5 \
     --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true \
+    --dataset.push_to_hub=true
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
 ```
 
 **Note**: Update `server_address` to match your robot's camera server IP.

docs/source/using_dataset_tools.mdx

@@ -12,7 +12,6 @@ LeRobot provides several utilities for manipulating datasets:
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
 6. **Convert to Video** - Convert image-based datasets to video format for efficient storage
-7. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
 
 The core implementation is in `lerobot.datasets.dataset_tools`.
 An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
@@ -157,30 +156,6 @@ lerobot-edit-dataset \
 
 **Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). All episodes, stats, and tasks are preserved.
 
-### Show the information of datasets
-
-Show the information of datasets such as number of episode, number of frame, File size and so on.
-No change will be made to the dataset
-
-```bash
-
-# Show dataset information without feature details
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht_image \
-    --operation.type info \
-
-# Show dataset information with feature details
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht_image \
-    --operation.type info \
-    --operation.show_features true
-
-```
-
-**Parameters:**
-
-- `parameters`: The flag to control show or no show dataset information with feature details.(default=false)
-
 ### Push to Hub
 
 Add the `--push_to_hub true` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:

docs/source/walloss.mdx

@@ -45,7 +45,7 @@ policy.type=wall_x
 For training WallX, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
     --dataset.repo_id=your_dataset \
     --policy.type=wall_x \
     --output_dir=./outputs/wallx_training \

docs/source/xvla.mdx

@@ -154,7 +154,7 @@ lerobot-train \
 
 ```bash
 lerobot-train \
-  --dataset.repo_id=<USER>/bimanual-so100-handover-cube \
+  --dataset.repo_id=pepijn223/bimanual-so100-handover-cube \
   --output_dir=./outputs/xvla_bimanual \
   --job_name=xvla_so101_training \
   --policy.path="lerobot/xvla-base" \

examples/backward_compatibility/replay.py

@@ -22,7 +22,7 @@ lerobot-replay \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=black \
-    --dataset.repo_id=<USER>/record-test \
+    --dataset.repo_id=aliberts/record-test \
     --dataset.episode=2
 ```
 """
@@ -57,7 +57,7 @@ class DatasetReplayConfig:
     repo_id: str
     # Episode to replay.
     episode: int
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | Path | None = None
     # Limit the frames per second. By default, uses the policy fps.
     fps: int = 30

examples/rtc/eval_dataset.py

@@ -27,8 +27,8 @@ measuring consistency and ground truth alignment.
 Usage:
     # Basic usage with smolvla policy
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --rtc.max_guidance_weight=10.0 \
@@ -58,16 +58,16 @@ Usage:
         --device=cuda
 
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/reuben_pi0 \
+        --policy.path=lipsop/reuben_pi0 \
-        --dataset.repo_id=<USER>/so101_cube_in_cup \
+        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
         --rtc.execution_horizon=8 \
         --device=cuda
 
     # With torch.compile for faster inference (PyTorch 2.0+)
     # Note: CUDA graphs disabled by default due to in-place ops in denoising loop
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --use_torch_compile=true \
@@ -75,8 +75,8 @@ Usage:
 
     # With torch.compile on CUDA (CUDA graphs disabled by default)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=cuda \
         --use_torch_compile=true \
@@ -84,8 +84,8 @@ Usage:
 
     # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --dataset.repo_id=helper2424/check_rtc \
         --use_torch_compile=true \
         --torch_compile_backend=inductor \
         --torch_compile_mode=max-autotune \

examples/rtc/eval_with_real_robot.py

@@ -28,7 +28,7 @@ For simulation environments, see eval_with_simulation.py
 Usage:
     # Run RTC with Real robot with RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
@@ -41,7 +41,7 @@ Usage:
 
     # Run RTC with Real robot without RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=false \
         --robot.type=so100_follower \
@@ -53,7 +53,7 @@ Usage:
 
     # Run RTC with Real robot with pi0.5 policy
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/pi05_check_rtc \
+        --policy.path=helper2424/pi05_check_rtc \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \

pyproject.toml

@@ -59,7 +59,7 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=4.0.0,<5.0.0",
+    "datasets>=4.0.0,<4.2.0",
     "diffusers>=0.27.2,<0.36.0",
     "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
     "accelerate>=1.10.0,<2.0.0",
@@ -76,9 +76,9 @@ dependencies = [
     "pyserial>=3.5,<4.0",
     "wandb>=0.24.0,<0.25.0",
 
-    "torch>=2.2.1,<2.11.0", # TODO: Bump dependency
+    "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
-    "torchcodec>=0.2.1,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bump dependency
+    "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
-    "torchvision>=0.21.0,<0.26.0", # TODO: Bump dependency
+    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
 
     "draccus==0.10.0", # TODO: Remove ==
     "gymnasium>=1.1.1,<2.0.0",
@@ -98,13 +98,11 @@ pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.10.0"]
 transformers-dep = ["transformers>=4.57.1,<5.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
-can-dep = ["python-can>=4.2.0,<5.0.0"]
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
-damiao = ["lerobot[can-dep]"]
+damiao = ["python-can>=4.2.0,<5.0.0"]
-robstride = ["lerobot[can-dep]"]
 
 # Robots
 openarms = ["lerobot[damiao]"]
@@ -214,9 +212,6 @@ lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
 lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
 
 # ---------------- Tool Configurations ----------------
-[tool.setuptools.package-data]
-lerobot = ["envs/*.json"]
-
 [tool.setuptools.packages.find]
 where = ["src"]
 
@@ -365,9 +360,9 @@ ignore_errors = false
 module = "lerobot.cameras.*"
 ignore_errors = false
 
-[[tool.mypy.overrides]]
+# [[tool.mypy.overrides]]
-module = "lerobot.motors.*"
+# module = "lerobot.motors.*"
-ignore_errors = false
+# ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.robots.*"

src/lerobot/async_inference/robot_client.py

@@ -49,18 +49,23 @@ import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.robots import (
+from lerobot.robots import (  # noqa: F401
-    RobotConfig,  # noqa: F401
+    Robot,
+    RobotConfig,
+    bi_so_follower,
+    koch_follower,
     make_robot_from_config,
+    omx_follower,
+    so_follower,
 )
 from lerobot.transport import (
     services_pb2,  # type: ignore
     services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import grpc_channel_options, send_bytes_in_chunks
-from lerobot.utils.import_utils import register_third_party_plugins
 
 from .configs import RobotClientConfig
+from .constants import SUPPORTED_ROBOTS
 from .helpers import (
     Action,
     FPSTracker,
@@ -480,9 +485,8 @@ class RobotClient:
 def async_client(cfg: RobotClientConfig):
     logging.info(pformat(asdict(cfg)))
 
-    # TODO: Assert if checking robot support is still needed with the plugin system
+    if cfg.robot.type not in SUPPORTED_ROBOTS:
-    # if cfg.robot.type not in SUPPORTED_ROBOTS:
+        raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
-    #     raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
 
     client = RobotClient(cfg)
 
@@ -508,5 +512,4 @@ def async_client(cfg: RobotClientConfig):
 
 
 if __name__ == "__main__":
-    register_third_party_plugins()
     async_client()  # run the client

src/lerobot/cameras/__init__.py

@@ -13,5 +13,5 @@
 # limitations under the License.
 
 from .camera import Camera
-from .configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
+from .configs import CameraConfig, ColorMode, Cv2Rotation
 from .utils import make_cameras_from_configs

src/lerobot/cameras/camera.py

@@ -150,7 +150,7 @@ class Camera(abc.ABC):
         """
         pass
 
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/cameras/configs.py

@@ -25,10 +25,6 @@ class ColorMode(str, Enum):
     RGB = "rgb"
     BGR = "bgr"
 
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`color_mode` is expected to be in {list(cls)}, but {value} is provided.")
-
 
 class Cv2Rotation(int, Enum):
     NO_ROTATION = 0
@@ -36,25 +32,6 @@ class Cv2Rotation(int, Enum):
     ROTATE_180 = 180
     ROTATE_270 = -90
 
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`rotation` is expected to be in {list(cls)}, but {value} is provided.")
-
-
-# Subset from https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html
-class Cv2Backends(int, Enum):
-    ANY = 0
-    V4L2 = 200
-    DSHOW = 700
-    PVAPI = 800
-    ANDROID = 1000
-    AVFOUNDATION = 1200
-    MSMF = 1400
-
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`backend` is expected to be in {list(cls)}, but {value} is provided.")
-
 
 @dataclass(kw_only=True)
 class CameraConfig(draccus.ChoiceRegistry, abc.ABC):  # type: ignore  # TODO: add type stubs for draccus

src/lerobot/cameras/opencv/camera_opencv.py

@@ -32,11 +32,10 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
     os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
-from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
-from ..utils import get_cv2_rotation
+from ..utils import get_cv2_backend, get_cv2_rotation
 from .configuration_opencv import ColorMode, OpenCVCameraConfig
 
 # NOTE(Steven): The maximum opencv device index depends on your operating system. For instance,
@@ -118,7 +117,7 @@ class OpenCVCamera(Camera):
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
-        self.backend: int = config.backend
+        self.backend: int = get_cv2_backend()
 
         if self.height and self.width:
             self.capture_width, self.capture_height = self.width, self.height
@@ -133,7 +132,6 @@ class OpenCVCamera(Camera):
         """Checks if the camera is currently connected and opened."""
         return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the OpenCV camera specified in the configuration.
@@ -150,6 +148,8 @@ class OpenCVCamera(Camera):
             ConnectionError: If the specified camera index/path is not found or fails to open.
             RuntimeError: If the camera opens but fails to apply requested settings.
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         # Use 1 thread for OpenCV operations to avoid potential conflicts or
         # blocking in multi-threaded applications, especially during data collection.
@@ -178,7 +178,6 @@ class OpenCVCamera(Camera):
 
         logger.info(f"{self} connected.")
 
-    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """
         Applies the specified FOURCC, FPS, width, and height settings to the connected camera.
@@ -198,6 +197,8 @@ class OpenCVCamera(Camera):
                           to the requested value.
             DeviceNotConnectedError: If the camera is not connected.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")
 
         # Set FOURCC first (if specified) as it can affect available FPS/resolution options
         if self.config.fourcc is not None:
@@ -347,7 +348,6 @@ class OpenCVCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
@@ -374,6 +374,9 @@ class OpenCVCamera(Camera):
                 f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
             )
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
 
@@ -487,7 +490,6 @@ class OpenCVCamera(Camera):
             self.latest_timestamp = None
             self.new_frame_event.clear()
 
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
@@ -510,6 +512,8 @@ class OpenCVCamera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
@@ -529,8 +533,7 @@ class OpenCVCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the
@@ -545,6 +548,8 @@ class OpenCVCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If the camera is connected but has not captured any frames yet.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")

src/lerobot/cameras/opencv/configuration_opencv.py

@@ -15,9 +15,9 @@
 from dataclasses import dataclass
 from pathlib import Path
 
-from ..configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
+from ..configs import CameraConfig, ColorMode, Cv2Rotation
 
-__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation", "Cv2Backends"]
+__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation"]
 
 
 @CameraConfig.register_subclass("opencv")
@@ -50,7 +50,6 @@ class OpenCVCameraConfig(CameraConfig):
         rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
         warmup_s: Time reading frames before returning from connect (in seconds)
         fourcc: FOURCC code for video format (e.g., "MJPG", "YUYV", "I420"). Defaults to None (auto-detect).
-        backend: OpenCV backend identifier (https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html). Defaults to ANY.
 
     Note:
         - Only 3-channel color output (RGB/BGR) is currently supported.
@@ -63,12 +62,22 @@ class OpenCVCameraConfig(CameraConfig):
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
     fourcc: str | None = None
-    backend: Cv2Backends = Cv2Backends.ANY
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
-        self.rotation = Cv2Rotation(self.rotation)
+            raise ValueError(
-        self.backend = Cv2Backends(self.backend)
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
+
+        if self.rotation not in (
+            Cv2Rotation.NO_ROTATION,
+            Cv2Rotation.ROTATE_90,
+            Cv2Rotation.ROTATE_180,
+            Cv2Rotation.ROTATE_270,
+        ):
+            raise ValueError(
+                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
+            )
 
         if self.fourcc is not None and (not isinstance(self.fourcc, str) or len(self.fourcc) != 4):
             raise ValueError(

src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py

@@ -74,4 +74,7 @@ class Reachy2CameraConfig(CameraConfig):
                 f"`image_type` is expected to be 'left' or 'right' for teleop camera, and 'rgb' or 'depth' for depth camera, but {self.image_type} is provided."
             )
 
-        self.color_mode = ColorMode(self.color_mode)
+        if self.color_mode not in ["rgb", "bgr"]:
+            raise ValueError(
+                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
+            )

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -32,7 +32,6 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy
 
-from lerobot.utils.decorators import check_if_not_connected
 from lerobot.utils.import_utils import _reachy2_sdk_available
 
 if TYPE_CHECKING or _reachy2_sdk_available:
@@ -124,7 +123,6 @@ class Reachy2Camera(Camera):
         """
         raise NotImplementedError("Camera detection is not implemented for Reachy2 cameras.")
 
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
@@ -138,6 +136,9 @@ class Reachy2Camera(Camera):
         """
         start_time = time.perf_counter()
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         if self.cam_manager is None:
             raise DeviceNotConnectedError(f"{self} is not connected.")
 
@@ -183,7 +184,6 @@ class Reachy2Camera(Camera):
 
         return frame
 
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Same as read()
@@ -197,11 +197,12 @@ class Reachy2Camera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         return self.read()
 
-    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the
@@ -218,6 +219,8 @@ class Reachy2Camera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If the camera is connected but has not captured any frames yet.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.latest_frame is None or self.latest_timestamp is None:
             raise RuntimeError(f"{self} has not captured any frames yet.")
@@ -230,7 +233,6 @@ class Reachy2Camera(Camera):
 
         return self.latest_frame
 
-    @check_if_not_connected
     def disconnect(self) -> None:
         """
         Stops the background read thread (if running).
@@ -238,6 +240,8 @@ class Reachy2Camera(Camera):
         Raises:
             DeviceNotConnectedError: If the camera is already disconnected.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} not connected.")
 
         if self.cam_manager is not None:
             self.cam_manager.disconnect()

src/lerobot/cameras/realsense/camera_realsense.py

@@ -30,8 +30,7 @@ try:
 except Exception as e:
     logging.info(f"Could not import realsense: {e}")
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
-from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -153,7 +152,6 @@ class RealSenseCamera(Camera):
         """Checks if the camera pipeline is started and streams are active."""
         return self.rs_pipeline is not None and self.rs_profile is not None
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the RealSense camera specified in the configuration.
@@ -171,6 +169,8 @@ class RealSenseCamera(Camera):
             ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
             RuntimeError: If the pipeline starts but fails to apply requested settings.
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         self.rs_pipeline = rs.pipeline()
         rs_config = rs.config()
@@ -290,7 +290,6 @@ class RealSenseCamera(Camera):
             if self.use_depth:
                 rs_config.enable_stream(rs.stream.depth)
 
-    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """Sets fps, width, and height from device stream if not already configured.
 
@@ -300,6 +299,8 @@ class RealSenseCamera(Camera):
         Raises:
             DeviceNotConnectedError: If device is not connected.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")
 
         if self.rs_profile is None:
             raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
@@ -319,7 +320,6 @@ class RealSenseCamera(Camera):
                 self.width, self.height = actual_width, actual_height
                 self.capture_width, self.capture_height = actual_width, actual_height
 
-    @check_if_not_connected
     def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (depth) synchronously from the camera.
@@ -345,6 +345,9 @@ class RealSenseCamera(Camera):
                 f"Failed to capture depth frame '.read_depth()'. Depth stream is not enabled for {self}."
             )
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
 
@@ -371,7 +374,6 @@ class RealSenseCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 0) -> NDArray[Any]:
         """
         Reads a single frame (color) synchronously from the camera.
@@ -401,6 +403,9 @@ class RealSenseCamera(Camera):
                 f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
             )
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
 
@@ -529,7 +534,6 @@ class RealSenseCamera(Camera):
             self.new_frame_event.clear()
 
     # NOTE(Steven): Missing implementation for depth for now
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame data (color) asynchronously.
@@ -552,6 +556,8 @@ class RealSenseCamera(Camera):
             TimeoutError: If no frame data becomes available within the specified timeout.
             RuntimeError: If the background thread died unexpectedly or another error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
@@ -572,8 +578,7 @@ class RealSenseCamera(Camera):
         return frame
 
     # NOTE(Steven): Missing implementation for depth for now
-    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
         """Return the most recent (color) frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the
@@ -588,6 +593,8 @@ class RealSenseCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If the camera is connected but has not captured any frames yet.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")

src/lerobot/cameras/realsense/configuration_realsense.py

@@ -60,8 +60,20 @@ class RealSenseCameraConfig(CameraConfig):
     warmup_s: int = 1
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
-        self.rotation = Cv2Rotation(self.rotation)
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
+
+        if self.rotation not in (
+            Cv2Rotation.NO_ROTATION,
+            Cv2Rotation.ROTATE_90,
+            Cv2Rotation.ROTATE_180,
+            Cv2Rotation.ROTATE_270,
+        ):
+            raise ValueError(
+                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
+            )
 
         values = (self.fps, self.width, self.height)
         if any(v is not None for v in values) and any(v is None for v in values):

src/lerobot/cameras/utils.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import platform
 from typing import cast
 
 from lerobot.utils.import_utils import make_device_from_device_class
@@ -67,3 +68,14 @@ def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
         return int(cv2.ROTATE_90_COUNTERCLOCKWISE)
     else:
         return None
+
+
+def get_cv2_backend() -> int:
+    import cv2
+
+    if platform.system() == "Windows":
+        return int(cv2.CAP_MSMF)  # Use MSMF for Windows instead of AVFOUNDATION
+    # elif platform.system() == "Darwin":  # macOS
+    #     return cv2.CAP_AVFOUNDATION
+    else:  # Linux and others
+        return int(cv2.CAP_ANY)

src/lerobot/cameras/zmq/camera_zmq.py

@@ -34,8 +34,7 @@ import cv2
 import numpy as np
 from numpy.typing import NDArray
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
-from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -105,7 +104,6 @@ class ZMQCamera(Camera):
         """Checks if the ZMQ socket is initialized and connected."""
         return self._connected and self.context is not None and self.socket is not None
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """Connect to ZMQ camera server.
 
@@ -113,6 +111,8 @@ class ZMQCamera(Camera):
             warmup (bool): If True, waits for the camera to provide at least one
                            valid frame before returning. Defaults to True.
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         logger.info(f"Connecting to {self}...")
 
@@ -211,7 +211,6 @@ class ZMQCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
@@ -229,6 +228,9 @@ class ZMQCamera(Camera):
                 f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
             )
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
 
@@ -299,7 +301,6 @@ class ZMQCamera(Camera):
             self.latest_timestamp = None
             self.new_frame_event.clear()
 
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
@@ -316,6 +317,8 @@ class ZMQCamera(Camera):
             TimeoutError: If no frame data becomes available within the specified timeout.
             RuntimeError: If the background thread is not running.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
@@ -332,7 +335,6 @@ class ZMQCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
     def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
@@ -348,6 +350,8 @@ class ZMQCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If the camera is connected but has not captured any frames yet.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")

src/lerobot/cameras/zmq/configuration_zmq.py

@@ -32,7 +32,10 @@ class ZMQCameraConfig(CameraConfig):
     warmup_s: int = 1
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
 
         if self.timeout_ms <= 0:
             raise ValueError(f"`timeout_ms` must be positive, but {self.timeout_ms} is provided.")

src/lerobot/configs/default.py

@@ -27,7 +27,7 @@ class DatasetConfig:
     # "dataset_index" into the returned item. The index mapping is made according to the order in which the
     # datasets are provided.
     repo_id: str
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | None = None
     episodes: list[int] | None = None
     image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)

src/lerobot/data_processing/annotations/annotate_libero.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+
+# Example script to run synthetic data generation with Qwen VLM
+# This generates user prompts and robot utterances for hierarchical policy training
+
+# Configuration
+REPO_ID="lerobot/libero_10"
+MODEL="Qwen/Qwen3-VL-30B-A3B-Instruct"
+# or: MODEL="Qwen/Qwen2-VL-7B-Instruct"
+
+
+OUTPUT_DIR="/fsx/jade_choghari/outputs/libero-10-annotate-high"
+
+BATCH_SIZE=16
+TEMPERATURE=0.9
+SAMPLE_INTERVAL=5.0  # generate dialogue every 1 second (all episodes processed)
+
+# Run subtask annotation
+# python /admin/home/jade_choghari/lerobot/src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+#     --repo-id "$REPO_ID" \
+#     --video-key observation.images.image \
+#     --output-dir "$OUTPUT_DIR" \
+#     --skip-existing \
+#     --output-repo-id "jadechoghari/libero10-annotate" \
+#     --batch-size "$BATCH_SIZE" \
+# run synthetic data generation (all episodes processed)
+# python examples/dataset/annotate_pgen.py \
+#     --repo-id "$REPO_ID" \
+#     --model "$MODEL" \
+#     --output-dir "$OUTPUT_DIR" \
+#     --temperature "$TEMPERATURE" \
+#     --batch-size "$BATCH_SIZE" \
+#     --sample-interval "$SAMPLE_INTERVAL" \
+#     --image-key observation.images.base \
+#     --num-image-views-per-sample 1
+
+# for faster testing, increase sample interval:
+# --sample-interval 5.0  # Samples every 5 seconds (much faster)
+
+# to push to hub after generation:
+# add --push-to-hub flag
+
+# efficient batch processing: 4 episodes at once
+python src/lerobot/data_processing/annotations/high_level_annotate.py \
+    --data-dir "/fsx/jade_choghari/outputs/libero-10-annotate" \
+    --output-dir "$OUTPUT_DIR" \
+    --video-mode \
+    --video-key observation.images.image \
+    --video-batch-size "$BATCH_SIZE" \
+    --sample-interval 5.0

src/lerobot/data_processing/annotations/load_lerobot_high.py

@@ -0,0 +1,52 @@
+import torch
+from huggingface_hub import HfApi
+
+import lerobot
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.configs.policies import PreTrainedConfig
+
+# /fsx/jade_choghari/data/libero_10_subtasks_kw_converted
+dataset = LeRobotDataset(repo_id="lerobot/libero_10_image_subtask")
+
+dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=0,
+        batch_size=2,
+        shuffle=True,
+)
+
+cfg = PreTrainedConfig.from_pretrained(
+    pretrained_name_or_path="/fsx/jade_choghari/models/pi05-base",
+)
+cfg.dtype = "bfloat16"
+
+pre_processor, post_processor = make_pre_post_processors(
+    policy_cfg=cfg,
+    pretrained_path="/fsx/jade_choghari/models/pi05-base",
+)
+batch = next(iter(dataloader))
+breakpoint()
+batch1 = pre_processor(batch)
+breakpoint()
+print(batch.keys())
+# print(batch['task_index_high_level'].shape)
+# print(batch['task_index_high_level'])
+# print(batch['user_prompt'][0])
+# print(batch['robot_utterance'][0])
+# print(batch['task'][0])
+
+valid_episode_list = []
+for episode_idx in range(len(dataset.meta.episodes)):
+        subtask_index = dataset[episode_idx]["subtask_index"]
+        valid_episode_list.append(episode_idx)
+
+print(len(valid_episode_list))
+
+# read this parquet /fsx/jade_choghari/outputs/pgen_annotations1/meta/tasks.parquett
+# import pandas as pd
+# tasks_df = pd.read_parquet('/fsx/jade_choghari/outputs/pgen_annotations1/meta/tasks.parquet')
+
+# # print all
+# print(tasks_df.columns)
+# breakpoint()

src/lerobot/data_processing/annotations/run_pgen.sh

@@ -0,0 +1,74 @@
+#!/bin/bash
+
+# Example script to run synthetic data generation with Qwen VLM
+# This generates user prompts and robot utterances for hierarchical policy training
+
+# Configuration
+REPO_ID="jadechoghari/piper-demo-20260205_103303"
+# MODEL="Qwen/Qwen3-VL-30B-A3B-Thinking"
+MODEL="Qwen/Qwen3.5-27B"
+# or: MODEL="Qwen/Qwen2-VL-7B-Instruct"
+
+
+OUTPUT_DIR="/fsx/jade_choghari/outputs/collect-data-pgen_new"
+
+BATCH_SIZE=2
+TEMPERATURE=0.9
+SAMPLE_INTERVAL=5.0  # generate dialogue every 1 second (all episodes processed)
+
+# Run subtask annotation.
+# To use closed-vocabulary labels, add a line: --subtask-labels "label1" "label2" ...
+# Example (add backslash after "$MODEL" and uncomment the next line):
+#   --model "$MODEL" \
+#   --subtask-labels "pick_up_yellow_nut_bar" "pick_up_cake" "pick_up_biscuit_pack" "pick_up_soda_can"
+python /home/lerobot/src/lerobot/data_processing/annotations/subtask_annotate.py \
+    --repo-id "$REPO_ID" \
+    --video-key observation.images.top \
+    --output-dir "$OUTPUT_DIR" \
+    --output-repo-id "jadechoghari/piper-demo-annotated1" \
+    --push-to-hub \
+    --no-timer-overlay \
+    --model "$MODEL" \
+    --subtask-labels "pick_up_yellow_nut_bar" "pick_up_cake" "pick_up_biscuit_pack" "pick_up_soda_can" \
+    --batch-size 2
+
+# Run subtask annotation (image-window: frames as images for better accuracy)
+# python /admin/home/jade_choghari/lerobot/src/lerobot/data_processing/annotations/subtask_annotate_image.py \
+#     --repo-id "$REPO_ID" \
+#     --camera-key observation.images.wrist \
+#     --output-dir "$OUTPUT_DIR" \
+#     --output-repo-id "jadechoghari/piper-demo-annotated1-image" \
+#     --push-to-hub \
+#     --model "$MODEL" \
+#     --window-size 184 \
+#     --max-frames-per-window 16 \
+#     --subtask-labels "pick_up_yellow_nut_bar" "pick_up_cake" "pick_up_biscuit_pack" "pick_up_soda_can" \
+#     --batch-size 2
+
+    
+# run synthetic data generation (all episodes processed)
+# python examples/dataset/annotate_pgen.py \
+#     --repo-id "$REPO_ID" \
+#     --model "$MODEL" \
+#     --output-dir "$OUTPUT_DIR" \
+#     --temperature "$TEMPERATURE" \
+#     --batch-size "$BATCH_SIZE" \
+#     --sample-interval "$SAMPLE_INTERVAL" \
+#     --image-key observation.images.base \
+#     --num-image-views-per-sample 1
+
+# for faster testing, increase sample interval:
+# --sample-interval 5.0  # Samples every 5 seconds (much faster)
+
+# to push to hub after generation:
+# add --push-to-hub flag
+
+# efficient batch processing: 4 episodes at once
+# python examples/dataset/annotate_pgen.py \
+#     --repo-id "$REPO_ID" \
+#     --model "$MODEL" \
+#     --output-dir "$OUTPUT_DIR" \
+#     --video-mode \
+#     --video-key observation.images.up \
+#     --video-batch-size "$BATCH_SIZE" \
+#     --sample-interval 1.0

src/lerobot/data_processing/annotations/subtask_annotate_image.py

@@ -0,0 +1,561 @@
+#!/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.
+
+"""
+Image-window subtask annotation for LeRobot datasets using Qwen VLMs.
+
+This script assigns a subtask to each window of consecutive frames by sending
+those frames as images to the VLM (instead of a video) for better accuracy.
+Supports Qwen2-VL and Qwen3-VL (same models as subtask_annotate.py).
+
+Pipeline:
+1. Load a LeRobot dataset (local or Hub).
+2. For each episode, slide a window over frame indices.
+3. For each window, load the corresponding images (from image_key or decoded video_key).
+4. Send the window of images to Qwen2-VL with the same skill prompt; get one subtask name.
+5. Assign that subtask to all frames in the window.
+6. Write subtasks.parquet and add subtask_index via add_features (same as subtask_annotate).
+
+Usage:
+  python -m lerobot.data_processing.annotations.subtask_annotate_image \\
+    --data-dir /path/to/dataset --camera-key observation.images.base \\
+    --window-size 8 --stride 8 --output-dir ./output
+"""
+
+from __future__ import annotations
+
+import argparse
+import random
+import textwrap
+from pathlib import Path
+
+import numpy as np
+import PIL.Image
+import torch
+from rich.console import Console
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Reuse data structures and save/load from the video-based annotator
+from lerobot.data_processing.annotations.subtask_annotate import (
+    EpisodeSkills,
+    Skill,
+    load_skill_annotations,
+    save_skill_annotations,
+)
+
+
+def create_window_skill_prompt(
+    coarse_goal: str | None = None,
+    subtask_labels: list[str] | None = None,
+) -> str:
+    """Prompt for labeling a single window of frames with one atomic skill.
+    If subtask_labels are provided, the model must choose exactly one from that list.
+    """
+    goal_context = f'The overall goal is: "{coarse_goal}".\n\n' if coarse_goal else ""
+    if subtask_labels:
+        labels_list = ", ".join(f'"{l}"' for l in subtask_labels)
+        label_instruction = (
+            f"You must choose exactly ONE skill from this list: [{labels_list}]. "
+            "Do not create new labels. Reply with only that label.\n\n"
+        )
+    else:
+        label_instruction = ""
+    return textwrap.dedent(f"""\
+        # Role
+        You are a Robotics Vision System that labels short clips from robot manipulation demonstrations.
+
+        # Task
+        {goal_context}{label_instruction}The following images are consecutive frames from a single short clip of a robot demonstration.
+        What single atomic manipulation skill is being performed in this clip?
+
+        # Requirements
+        - Reply with ONLY one short skill name (e.g. "pick up object", "move arm left", "release gripper").
+        - No explanation, no timestamps, no JSON. Just the skill name.
+        """).strip()
+
+
+def _run_image_segmenter(
+    self,
+    images: list[PIL.Image.Image],
+    coarse_goal: str | None,
+    subtask_labels: list[str] | None = None,
+) -> str:
+    """Shared inference for Qwen2-VL and Qwen3-VL image window labeling."""
+    prompt = create_window_skill_prompt(coarse_goal, subtask_labels)
+    content = []
+    for img in images:
+        content.append({"type": "image", "image": img})
+    content.append({"type": "text", "text": "What single atomic skill is shown in these frames? Reply with only the skill name."})
+
+    messages = [
+        {"role": "system", "content": [{"type": "text", "text": prompt}]},
+        {"role": "user", "content": content},
+    ]
+    text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    image_inputs, video_inputs = self.process_vision_info(messages)
+    inputs = self.processor(
+        text=[text],
+        images=image_inputs,
+        videos=video_inputs,
+        padding=True,
+        return_tensors="pt",
+    ).to(self.device)
+
+    with torch.no_grad():
+        generated_ids = self.model.generate(**inputs, max_new_tokens=128, do_sample=False)
+
+    response = self.processor.batch_decode(
+        [out[len(inp) :] for inp, out in zip(inputs.input_ids, generated_ids)],
+        skip_special_tokens=True,
+    )[0].strip()
+    skill_name = response.split("\n")[0].strip().strip('."')
+    return skill_name if skill_name else "unknown"
+
+
+def _run_image_segmenter_batch(
+    self,
+    batch_images: list[list[PIL.Image.Image]],
+    coarse_goal: str | None,
+    subtask_labels: list[str] | None = None,
+) -> list[str]:
+    """Run VLM on multiple windows at once; returns one skill name per window."""
+    if not batch_images:
+        return []
+    prompt = create_window_skill_prompt(coarse_goal, subtask_labels)
+    all_texts = []
+    all_image_inputs = []
+    all_video_inputs = []
+    for images in batch_images:
+        content = []
+        for img in images:
+            content.append({"type": "image", "image": img})
+        content.append({"type": "text", "text": "What single atomic skill is shown in these frames? Reply with only the skill name."})
+        messages = [
+            {"role": "system", "content": [{"type": "text", "text": prompt}]},
+            {"role": "user", "content": content},
+        ]
+        text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        image_inputs, video_inputs = self.process_vision_info(messages)
+        all_texts.append(text)
+        if image_inputs is not None:
+            all_image_inputs.extend(image_inputs if isinstance(image_inputs, list) else [image_inputs])
+        if video_inputs is not None:
+            all_video_inputs.extend(video_inputs if isinstance(video_inputs, list) else [video_inputs])
+    inputs = self.processor(
+        text=all_texts,
+        images=all_image_inputs if all_image_inputs else None,
+        videos=all_video_inputs if all_video_inputs else None,
+        padding=True,
+        return_tensors="pt",
+    ).to(self.device)
+    with torch.no_grad():
+        generated_ids = self.model.generate(**inputs, max_new_tokens=128, do_sample=False)
+    responses = self.processor.batch_decode(
+        [out[len(inp) :] for inp, out in zip(inputs.input_ids, generated_ids)],
+        skip_special_tokens=True,
+    )
+    return [
+        (r.split("\n")[0].strip().strip('."') or "unknown")
+        for r in responses
+    ]
+
+
+class Qwen2VLImageSegmenter:
+    """Uses Qwen2-VL to assign one skill name to a window of images (same model as subtask_annotate)."""
+
+    def __init__(self, model_name: str, device: str = "cuda", torch_dtype: torch.dtype = torch.bfloat16):
+        from qwen_vl_utils import process_vision_info
+        from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
+
+        self.console = Console()
+        self.device = device
+        self.process_vision_info = process_vision_info
+        self.console.print(f"[cyan]Loading Qwen2-VL for image-window labeling: {model_name}...[/cyan]")
+        self.model = Qwen2VLForConditionalGeneration.from_pretrained(
+            model_name, torch_dtype=torch_dtype, device_map=device, trust_remote_code=True
+        )
+        self.processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
+        self.console.print(f"[green]✓ Model loaded on {device}[/green]")
+
+    def segment_skill_from_images(
+        self,
+        images: list[PIL.Image.Image],
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> str:
+        """Return a single skill name for the given window of images."""
+        return _run_image_segmenter(self, images, coarse_goal, subtask_labels)
+
+    def segment_skill_from_images_batch(
+        self,
+        batch_images: list[list[PIL.Image.Image]],
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> list[str]:
+        """Return one skill name per window; processes multiple windows in one forward pass."""
+        return _run_image_segmenter_batch(self, batch_images, coarse_goal, subtask_labels)
+
+
+class Qwen3VLImageSegmenter:
+    """Uses Qwen3-VL (MoE) to assign one skill name to a window of images."""
+
+    def __init__(self, model_name: str, device: str = "cuda", torch_dtype: torch.dtype = torch.bfloat16):
+        from qwen_vl_utils import process_vision_info
+        from transformers import AutoProcessor, Qwen3VLMoeForConditionalGeneration
+
+        self.console = Console()
+        self.device = device
+        self.process_vision_info = process_vision_info
+        self.console.print(f"[cyan]Loading Qwen3-VL for image-window labeling: {model_name}...[/cyan]")
+        self.model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
+            model_name, torch_dtype=torch_dtype, device_map=device, trust_remote_code=True
+        )
+        self.processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
+        self.console.print(f"[green]✓ Model loaded on {device}[/green]")
+
+    def segment_skill_from_images(
+        self,
+        images: list[PIL.Image.Image],
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> str:
+        """Return a single skill name for the given window of images."""
+        return _run_image_segmenter(self, images, coarse_goal, subtask_labels)
+
+    def segment_skill_from_images_batch(
+        self,
+        batch_images: list[list[PIL.Image.Image]],
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> list[str]:
+        """Return one skill name per window; processes multiple windows in one forward pass."""
+        return _run_image_segmenter_batch(self, batch_images, coarse_goal, subtask_labels)
+
+
+def get_image_segmenter(
+    model_name: str,
+    device: str = "cuda",
+    torch_dtype: torch.dtype = torch.bfloat16,
+):
+    """Return the appropriate image-window segmenter for the model (Qwen2-VL or Qwen3-VL)."""
+    model_lower = model_name.lower()
+    if "qwen3" in model_lower:
+        return Qwen3VLImageSegmenter(model_name, device, torch_dtype)
+    return Qwen2VLImageSegmenter(model_name, device, torch_dtype)
+
+
+def frame_to_pil(frame_value) -> PIL.Image.Image:
+    """Convert a single frame from dataset (tensor or PIL or path) to PIL.Image."""
+    if isinstance(frame_value, PIL.Image.Image):
+        return frame_value
+    if isinstance(frame_value, (str, Path)):
+        return PIL.Image.open(frame_value).convert("RGB")
+    if hasattr(frame_value, "numpy"):
+        arr = frame_value.numpy()
+    else:
+        arr = np.asarray(frame_value)
+    if arr.ndim == 3 and arr.shape[0] in (1, 3, 4):
+        arr = np.transpose(arr, (1, 2, 0))
+    if arr.dtype == np.float32 or arr.dtype == np.float64:
+        arr = (np.clip(arr, 0, 1) * 255).astype(np.uint8)
+    elif arr.dtype != np.uint8:
+        arr = np.clip(arr, 0, 255).astype(np.uint8)
+    if arr.shape[-1] == 1:
+        arr = np.repeat(arr, 3, axis=-1)
+    return PIL.Image.fromarray(arr)
+
+
+def _sample_window_indices(window_length: int, max_frames: int) -> list[int]:
+    """Return indices into a window of length window_length, at most max_frames, in order.
+    If window_length <= max_frames, returns range(window_length).
+    Otherwise returns sorted random sample of max_frames indices (temporal order preserved).
+    """
+    if max_frames <= 0 or window_length <= max_frames:
+        return list(range(window_length))
+    return sorted(random.sample(range(window_length), max_frames))
+
+
+class SkillAnnotatorImage:
+    """Annotates episodes by sliding a window over frames and labeling each window with the VLM."""
+
+    def __init__(
+        self,
+        segmenter: Qwen2VLImageSegmenter | Qwen3VLImageSegmenter,
+        window_size: int = 8,
+        stride: int | None = None,
+        batch_size: int = 1,
+        max_frames_per_window: int | None = None,
+        console: Console | None = None,
+    ):
+        self.segmenter = segmenter
+        self.window_size = window_size
+        self.stride = stride if stride is not None else window_size
+        self.batch_size = max(1, batch_size)
+        self.max_frames_per_window = max_frames_per_window
+        self.console = console or Console()
+
+    def annotate_dataset(
+        self,
+        dataset: LeRobotDataset,
+        camera_key: str,
+        episodes: list[int] | None = None,
+        skip_existing: bool = False,
+        subtask_labels: list[str] | None = None,
+    ) -> dict[int, EpisodeSkills]:
+        """Annotate episodes using image windows. camera_key can be an image_key or video_key."""
+        episode_indices = episodes or list(range(dataset.meta.total_episodes))
+        coarse_goal = self._get_coarse_goal(dataset)
+        annotations: dict[int, EpisodeSkills] = {}
+
+        if skip_existing:
+            existing = load_skill_annotations(dataset.root)
+            if existing and existing.get("episodes"):
+                existing_eps = {int(k) for k in existing["episodes"] if existing["episodes"][k].get("skills")}
+                episode_indices = [i for i in episode_indices if i not in existing_eps]
+
+        for ep_idx in episode_indices:
+            try:
+                skills = self._annotate_episode(
+                    dataset, ep_idx, camera_key, coarse_goal, subtask_labels
+                )
+                if skills:
+                    annotations[ep_idx] = EpisodeSkills(
+                        episode_index=ep_idx,
+                        description=coarse_goal,
+                        skills=skills,
+                    )
+                    self.console.print(f"[green]✓ Episode {ep_idx}: {len(skills)} window skills[/green]")
+                else:
+                    self.console.print(f"[yellow]⚠ Episode {ep_idx}: no skills[/yellow]")
+            except Exception as e:
+                self.console.print(f"[red]Episode {ep_idx} failed: {e}[/red]")
+
+        return annotations
+
+    def _get_coarse_goal(self, dataset: LeRobotDataset) -> str:
+        if dataset.meta.tasks is not None and len(dataset.meta.tasks) > 0:
+            return str(dataset.meta.tasks.index[0])
+        return "Perform the demonstrated manipulation task."
+
+    def _annotate_episode(
+        self,
+        dataset: LeRobotDataset,
+        episode_index: int,
+        camera_key: str,
+        coarse_goal: str,
+        subtask_labels: list[str] | None = None,
+    ) -> list[Skill]:
+        ep = dataset.meta.episodes[episode_index]
+        ep_from = int(ep["dataset_from_index"])
+        ep_to = int(ep["dataset_to_index"])
+        length = ep_to - ep_from
+        fps = dataset.meta.fps
+        if length == 0:
+            return []
+
+        # Collect full windows: (images, t_start, t_end) using frame timestamps.
+        # If max_frames_per_window is set and window is larger, sample that many frames (order preserved).
+        window_specs: list[tuple[list[PIL.Image.Image], float, float]] = []
+        start = 0
+        while start + self.window_size <= length:
+            offsets = _sample_window_indices(
+                self.window_size,
+                self.max_frames_per_window or self.window_size,
+            )
+            frame_indices = [ep_from + start + i for i in offsets]
+            images = []
+            t_start = float(dataset[frame_indices[0]]["timestamp"].item())
+            for idx in frame_indices:
+                item = dataset[idx]
+                images.append(frame_to_pil(item[camera_key]))
+            t_end = t_start + self.window_size / fps
+            window_specs.append((images, t_start, t_end))
+            start += self.stride
+
+        # Last partial window
+        if start < length:
+            partial_len = ep_to - (ep_from + start)
+            offsets = _sample_window_indices(
+                partial_len,
+                self.max_frames_per_window or partial_len,
+            )
+            frame_indices = [ep_from + start + i for i in offsets]
+            images = []
+            t_start = float(dataset[frame_indices[0]]["timestamp"].item())
+            for idx in frame_indices:
+                item = dataset[idx]
+                images.append(frame_to_pil(item[camera_key]))
+            t_end = float(dataset[frame_indices[-1]]["timestamp"].item()) + 1.0 / fps
+            window_specs.append((images, t_start, t_end))
+
+        # Run in batches
+        skills: list[Skill] = []
+        for i in range(0, len(window_specs), self.batch_size):
+            chunk = window_specs[i : i + self.batch_size]
+            batch_images = [spec[0] for spec in chunk]
+            if len(batch_images) > 1:
+                skill_names = self.segmenter.segment_skill_from_images_batch(
+                    batch_images, coarse_goal, subtask_labels
+                )
+            else:
+                skill_names = [
+                    self.segmenter.segment_skill_from_images(
+                        batch_images[0], coarse_goal, subtask_labels
+                    )
+                ]
+            for (_, t_start, t_end), name in zip(chunk, skill_names, strict=True):
+                skills.append(Skill(name=name, start=t_start, end=t_end))
+
+        return skills
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Image-window subtask annotation using Qwen VLM (frames as images for better accuracy)",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog=textwrap.dedent("""\
+            Examples:
+              python -m lerobot.data_processing.annotations.subtask_annotate_image \\
+                --data-dir /path/to/dataset --camera-key observation.images.base \\
+                --window-size 8 --output-dir ./output
+
+              python -m lerobot.data_processing.annotations.subtask_annotate_image \\
+                --repo-id user/dataset --camera-key observation.images.base \\
+                --window-size 6 --stride 3 --model Qwen/Qwen2-VL-7B-Instruct
+
+              # Use Qwen3-VL (MoE)
+              python -m lerobot.data_processing.annotations.subtask_annotate_image \\
+                --data-dir /path/to/dataset --camera-key observation.images.base \\
+                --model Qwen/Qwen3-VL-30B-A3B-Instruct
+        """),
+    )
+    data_group = parser.add_mutually_exclusive_group(required=True)
+    data_group.add_argument("--data-dir", type=str, help="Path to local LeRobot dataset")
+    data_group.add_argument("--repo-id", type=str, help="HuggingFace Hub dataset repository ID")
+
+    parser.add_argument(
+        "--camera-key",
+        type=str,
+        required=True,
+        help="Image or video observation key (e.g. observation.images.base)",
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="Qwen/Qwen2-VL-7B-Instruct",
+        help="VLM model: Qwen2-VL or Qwen3-VL (default: Qwen/Qwen2-VL-7B-Instruct)",
+    )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda",
+    )
+    parser.add_argument(
+        "--window-size",
+        type=int,
+        default=8,
+        help="Number of frames per window (default: 8)",
+    )
+    parser.add_argument(
+        "--stride",
+        type=int,
+        default=None,
+        help="Stride for sliding window (default: window_size = non-overlapping)",
+    )
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=1,
+        help="Number of windows to process in one VLM call (default: 1; increase for speed)",
+    )
+    parser.add_argument(
+        "--max-frames-per-window",
+        type=int,
+        default=None,
+        metavar="N",
+        help="If window has more than N frames, randomly sample N frames (order kept) to avoid OOM (e.g. 16)",
+    )
+    parser.add_argument("--episodes", type=int, nargs="+", help="Episode indices to annotate (default: all)")
+    parser.add_argument("--skip-existing", action="store_true", help="Skip episodes that already have annotations")
+    parser.add_argument(
+        "--subtask-labels",
+        type=str,
+        nargs="*",
+        default=None,
+        help="Closed vocabulary: model must choose only from these labels",
+    )
+    parser.add_argument("--output-dir", type=str, help="Output directory for dataset with subtask_index")
+    parser.add_argument("--output-repo-id", type=str, help="Output repo id (default: <repo_id>_with_subtasks)")
+    parser.add_argument("--push-to-hub", action="store_true")
+
+    args = parser.parse_args()
+    console = Console()
+
+    # Load dataset
+    console.print("[cyan]Loading dataset...[/cyan]")
+    if args.data_dir:
+        dataset = LeRobotDataset(repo_id="local/dataset", root=args.data_dir, download_videos=False)
+    else:
+        dataset = LeRobotDataset(repo_id=args.repo_id, download_videos=True)
+    camera_keys = dataset.meta.camera_keys
+    if args.camera_key not in camera_keys:
+        console.print(f"[red]Error: camera key '{args.camera_key}' not in {camera_keys}[/red]")
+        return
+    console.print(f"[green]✓ Loaded dataset, {dataset.meta.total_episodes} episodes[/green]")
+
+    # Same Qwen VLM as subtask_annotate (Qwen2-VL or Qwen3-VL), image windows instead of video
+    segmenter = get_image_segmenter(args.model, args.device, torch.bfloat16)
+
+    annotator = SkillAnnotatorImage(
+        segmenter=segmenter,
+        window_size=args.window_size,
+        stride=args.stride,
+        batch_size=args.batch_size,
+        max_frames_per_window=args.max_frames_per_window,
+        console=console,
+    )
+    annotations = annotator.annotate_dataset(
+        dataset=dataset,
+        camera_key=args.camera_key,
+        episodes=args.episodes,
+        skip_existing=args.skip_existing,
+        subtask_labels=args.subtask_labels,
+    )
+
+    if not annotations:
+        console.print("[yellow]No annotations to save.[/yellow]")
+        return
+
+    output_dir = Path(args.output_dir) if args.output_dir else None
+    output_repo_id = args.output_repo_id
+    new_dataset = save_skill_annotations(dataset, annotations, output_dir, output_repo_id)
+
+    total_skills = sum(len(a.skills) for a in annotations.values())
+    console.print(f"[bold green]✓ Done.[/bold green] Episodes: {len(annotations)}, total window skills: {total_skills}")
+    console.print(f"  Dataset with subtask_index: {new_dataset.root}")
+
+    if args.push_to_hub and not args.data_dir:
+        console.print("[cyan]Pushing to Hub...[/cyan]")
+        try:
+            new_dataset.push_to_hub(push_videos=False)
+            console.print("[green]✓ Pushed.[/green]")
+        except Exception as e:
+            console.print(f"[red]Push failed: {e}[/red]")
+
+
+if __name__ == "__main__":
+    main()

src/lerobot/datasets/card_template.md

@@ -7,13 +7,6 @@
 
 This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
 
-{% if repo_id is defined and repo_id %}
-<a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path={{ repo_id }}">
-<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
-<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
-</a>
-{% endif %}
-
 ## Dataset Description
 
 {{ dataset_description | default("", true) }}

src/lerobot/datasets/dataset_tools.py

@@ -567,22 +567,20 @@ def _copy_and_reindex_data(
 def _keep_episodes_from_video_with_av(
     input_path: Path,
     output_path: Path,
-    episodes_to_keep: list[tuple[int, int]],
+    episodes_to_keep: list[tuple[float, float]],
     fps: float,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
 ) -> None:
     """Keep only specified episodes from a video file using PyAV.
 
-    This function decodes frames from specified frame ranges and re-encodes them with
+    This function decodes frames from specified time ranges and re-encodes them with
     properly reset timestamps to ensure monotonic progression.
 
     Args:
         input_path: Source video file path.
         output_path: Destination video file path.
-        episodes_to_keep: List of (start_frame, end_frame) tuples for episodes to keep.
+        episodes_to_keep: List of (start_time, end_time) tuples for episodes to keep.
-            Ranges are half-open intervals: [start_frame, end_frame), where start_frame
-            is inclusive and end_frame is exclusive.
         fps: Frame rate of the video.
         vcodec: Video codec to use for encoding.
         pix_fmt: Pixel format for output video.
@@ -624,10 +622,9 @@ def _keep_episodes_from_video_with_av(
 
     # Create set of (start, end) ranges for fast lookup.
     # Convert to a sorted list for efficient checking.
-    frame_ranges = sorted(episodes_to_keep)
+    time_ranges = sorted(episodes_to_keep)
 
     # Track frame index for setting PTS and current range being processed.
-    src_frame_count = 0
     frame_count = 0
     range_idx = 0
 
@@ -637,20 +634,21 @@ def _keep_episodes_from_video_with_av(
             if frame is None:
                 continue
 
-            # Check if frame is in any of our desired frame ranges.
+            # Get frame timestamp.
+            frame_time = float(frame.pts * frame.time_base) if frame.pts is not None else 0.0
+
+            # Check if frame is in any of our desired time ranges.
             # Skip ranges that have already passed.
-            while range_idx < len(frame_ranges) and src_frame_count >= frame_ranges[range_idx][1]:
+            while range_idx < len(time_ranges) and frame_time >= time_ranges[range_idx][1]:
                 range_idx += 1
 
             # If we've passed all ranges, stop processing.
-            if range_idx >= len(frame_ranges):
+            if range_idx >= len(time_ranges):
                 break
 
             # Check if frame is in current range.
-            start_frame = frame_ranges[range_idx][0]
+            start_ts, end_ts = time_ranges[range_idx]
-
+            if frame_time < start_ts:
-            if src_frame_count < start_frame:
-                src_frame_count += 1
                 continue
 
             # Frame is in range - create a new frame with reset timestamps.
@@ -663,7 +661,6 @@ def _keep_episodes_from_video_with_av(
             for pkt in v_out.encode(new_frame):
                 out.mux(pkt)
 
-            src_frame_count += 1
             frame_count += 1
 
     # Flush encoder.
@@ -752,17 +749,15 @@ def _copy_and_reindex_videos(
                         f"videos/{video_key}/to_timestamp"
                     ]
             else:
-                # Build list of frame ranges to keep, in sorted order.
+                # Build list of time ranges to keep, in sorted order.
                 sorted_keep_episodes = sorted(episodes_in_file, key=lambda x: episode_mapping[x])
-                episodes_to_keep_ranges: list[tuple[int, int]] = []
+                episodes_to_keep_ranges: list[tuple[float, float]] = []
+
                 for old_idx in sorted_keep_episodes:
                     src_ep = src_dataset.meta.episodes[old_idx]
-                    from_frame = round(src_ep[f"videos/{video_key}/from_timestamp"] * src_dataset.meta.fps)
+                    from_ts = src_ep[f"videos/{video_key}/from_timestamp"]
-                    to_frame = round(src_ep[f"videos/{video_key}/to_timestamp"] * src_dataset.meta.fps)
+                    to_ts = src_ep[f"videos/{video_key}/to_timestamp"]
-                    assert src_ep["length"] == to_frame - from_frame, (
+                    episodes_to_keep_ranges.append((from_ts, to_ts))
-                        f"Episode length mismatch: {src_ep['length']} vs {to_frame - from_frame}"
-                    )
-                    episodes_to_keep_ranges.append((from_frame, to_frame))
 
                 # Use PyAV filters to efficiently re-encode only the desired segments.
                 assert src_dataset.meta.video_path is not None

src/lerobot/datasets/lerobot_dataset.py

@@ -59,6 +59,7 @@ from lerobot.datasets.utils import (
     load_stats,
     load_subtasks,
     load_tasks,
+    load_tasks_high_level,
     update_chunk_file_indices,
     validate_episode_buffer,
     validate_frame,
@@ -68,7 +69,6 @@ from lerobot.datasets.utils import (
     write_tasks,
 )
 from lerobot.datasets.video_utils import (
-    StreamingVideoEncoder,
     VideoFrame,
     concatenate_video_files,
     decode_video_frames,
@@ -76,11 +76,11 @@ from lerobot.datasets.video_utils import (
     get_safe_default_codec,
     get_video_duration_in_s,
     get_video_info,
-    resolve_vcodec,
 )
 from lerobot.utils.constants import HF_LEROBOT_HOME
 
 CODEBASE_VERSION = "v3.0"
+VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1"}
 
 
 class LeRobotDatasetMetadata:
@@ -164,6 +164,7 @@ class LeRobotDatasetMetadata:
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
         self.tasks = load_tasks(self.root)
+        self.tasks_high_level = load_tasks_high_level(self.root)
         self.subtasks = load_subtasks(self.root)
         self.episodes = load_episodes(self.root)
         self.stats = load_stats(self.root)
@@ -521,6 +522,7 @@ class LeRobotDatasetMetadata:
         _validate_feature_names(features)
 
         obj.tasks = None
+        obj.tasks_high_level = None
         obj.subtasks = None
         obj.episodes = None
         obj.stats = None
@@ -546,19 +548,12 @@ class LeRobotDatasetMetadata:
 
 
 def _encode_video_worker(
-    video_key: str,
+    video_key: str, episode_index: int, root: Path, fps: int, vcodec: str = "libsvtav1"
-    episode_index: int,
-    root: Path,
-    fps: int,
-    vcodec: str = "libsvtav1",
-    encoder_threads: int | None = None,
 ) -> Path:
     temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
     fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
     img_dir = (root / fpath).parent
-    encode_video_frames(
+    encode_video_frames(img_dir, temp_path, fps, vcodec=vcodec, overwrite=True)
-        img_dir, temp_path, fps, vcodec=vcodec, overwrite=True, encoder_threads=encoder_threads
-    )
     shutil.rmtree(img_dir)
     return temp_path
 
@@ -578,9 +573,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
-        streaming_encoding: bool = False,
-        encoder_queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
     ):
         """
         2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -664,11 +656,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         for the README).
 
         Args:
-            repo_id (str): This is the repo id that will be used to fetch the dataset.
+            repo_id (str): This is the repo id that will be used to fetch the dataset. Locally, the dataset
-            root (Path | None, optional): Local directory where the dataset will be downloaded and
+                will be stored under root/repo_id.
-                stored. If set, all dataset files will be stored directly under this path. If not set, the
+            root (Path | None, optional): Local directory to use for downloading/writing files. You can also
-                dataset files will be stored under $HF_LEROBOT_HOME/repo_id (configurable via the
+                set the LEROBOT_HOME environment variable to point to a different location. Defaults to
-                HF_LEROBOT_HOME environment variable).
+                '~/.cache/huggingface/lerobot'.
             episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                 their episode_index in this list. Defaults to None.
             image_transforms (Callable | None, optional): You can pass standard v2 image transforms from
@@ -694,17 +686,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
             batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                 Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
             vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
-                'libsvtav1', 'auto', or hardware-specific codecs like 'h264_videotoolbox', 'h264_nvenc'.
+                'libsvtav1'. Defaults to 'libsvtav1'. Use 'h264' for faster encoding on systems where AV1
-                Defaults to 'libsvtav1'. Use 'auto' to auto-detect the best available hardware encoder.
+                encoding is CPU-heavy.
-            streaming_encoding (bool, optional): If True, encode video frames in real-time during capture
-                instead of writing PNG images first. This makes save_episode() near-instant. Defaults to False.
-            encoder_queue_maxsize (int, optional): Maximum number of frames to buffer per camera when using
-                streaming encoding. Defaults to 30 (~1s at 30fps).
-            encoder_threads (int | None, optional): Number of threads per encoder instance. None lets the
-                codec auto-detect (default). Lower values reduce CPU usage per encoder. Maps to 'lp' (via svtav1-params) for
-                libsvtav1 and 'threads' for h264/hevc.
         """
         super().__init__()
+        if vcodec not in VALID_VIDEO_CODECS:
+            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         self.repo_id = repo_id
         self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
         self.image_transforms = image_transforms
@@ -716,8 +703,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.delta_indices = None
         self.batch_encoding_size = batch_encoding_size
         self.episodes_since_last_encoding = 0
-        self.vcodec = resolve_vcodec(vcodec)
+        self.vcodec = vcodec
-        self._encoder_threads = encoder_threads
 
         # Unused attributes
         self.image_writer = None
@@ -725,7 +711,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.writer = None
         self.latest_episode = None
         self._current_file_start_frame = None  # Track the starting frame index of the current parquet file
-        self._streaming_encoder = None
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -747,7 +732,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             # Check if cached dataset contains all requested episodes
             if not self._check_cached_episodes_sufficient():
                 raise FileNotFoundError("Cached dataset doesn't contain all requested episodes")
-        except (FileNotFoundError, NotADirectoryError):
+        except (AssertionError, FileNotFoundError, NotADirectoryError):
             if is_valid_version(self.revision):
                 self.revision = get_safe_version(self.repo_id, self.revision)
             self.download(download_videos)
@@ -767,19 +752,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
             self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
 
-        # Initialize streaming encoder for resumed recording
-        if streaming_encoding and len(self.meta.video_keys) > 0:
-            self._streaming_encoder = StreamingVideoEncoder(
-                fps=self.meta.fps,
-                vcodec=self.vcodec,
-                pix_fmt="yuv420p",
-                g=2,
-                crf=30,
-                preset=None,
-                queue_maxsize=encoder_queue_maxsize,
-                encoder_threads=encoder_threads,
-            )
-
     def _close_writer(self) -> None:
         """Close and cleanup the parquet writer if it exists."""
         writer = getattr(self, "writer", None)
@@ -839,7 +811,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             hub_api.upload_folder(**upload_kwargs)
 
         card = create_lerobot_dataset_card(
-            tags=tags, dataset_info=self.meta.info, license=license, repo_id=self.repo_id, **card_kwargs
+            tags=tags, dataset_info=self.meta.info, license=license, **card_kwargs
         )
         card.push_to_hub(repo_id=self.repo_id, repo_type="dataset", revision=branch)
 
@@ -1098,7 +1070,17 @@ class LeRobotDataset(torch.utils.data.Dataset):
         if len(self.meta.video_keys) > 0:
             current_ts = item["timestamp"].item()
             query_timestamps = self._get_query_timestamps(current_ts, query_indices)
-            video_frames = self._query_videos(query_timestamps, ep_idx)
+            try:
+                video_frames = self._query_videos(query_timestamps, ep_idx)
+            except Exception as e:
+                print("\n" + "=" * 120)
+                print("[VIDEO DECODE FAILURE]")
+                print(f"item={item}")
+                print(f"query_indices={query_indices}")
+                print(f"query_timestamps={query_timestamps}")
+                print(f"ep_idx={ep_idx}")
+                print("=" * 120 + "\n")
+                raise
             item = {**video_frames, **item}
 
         if self.image_transforms is not None:
@@ -1109,6 +1091,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
         # Add task as a string
         task_idx = item["task_index"].item()
         item["task"] = self.meta.tasks.iloc[task_idx].name
+        
+        # optionally add high level task index
+        if "task_index_high_level" in self.features:
+            high_level_task_idx = item["task_index_high_level"].item()
+            item["robot_utterance"] = self.meta.tasks_high_level.iloc[high_level_task_idx]["robot_utterance"]
+            item["user_prompt"] = self.meta.tasks_high_level.iloc[high_level_task_idx]["user_prompt"]
+        
+        
 
         # add subtask information if available
         if "subtask_index" in self.features and self.meta.subtasks is not None:
@@ -1135,8 +1125,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         self._close_writer()
         self.meta._close_writer()
-        if self._streaming_encoder is not None:
-            self._streaming_encoder.close()
 
     def create_episode_buffer(self, episode_index: int | None = None) -> dict:
         current_ep_idx = self.meta.total_episodes if episode_index is None else episode_index
@@ -1191,13 +1179,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.episode_buffer["timestamp"].append(timestamp)
         self.episode_buffer["task"].append(frame.pop("task"))  # Remove task from frame after processing
 
-        # Start streaming encoder on first frame of episode (once, before iterating keys)
-        if frame_index == 0 and self._streaming_encoder is not None:
-            self._streaming_encoder.start_episode(
-                video_keys=list(self.meta.video_keys),
-                temp_dir=self.root,
-            )
-
         # Add frame features to episode_buffer
         for key in frame:
             if key not in self.features:
@@ -1205,10 +1186,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                     f"An element of the frame is not in the features. '{key}' not in '{self.features.keys()}'."
                 )
 
-            if self.features[key]["dtype"] == "video" and self._streaming_encoder is not None:
+            if self.features[key]["dtype"] in ["image", "video"]:
-                self._streaming_encoder.feed_frame(key, frame[key])
-                self.episode_buffer[key].append(None)  # Placeholder (video keys are skipped in parquet)
-            elif self.features[key]["dtype"] in ["image", "video"]:
                 img_path = self._get_image_file_path(
                     episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
                 )
@@ -1269,38 +1247,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Wait for image writer to end, so that episode stats over images can be computed
         self._wait_image_writer()
-
+        ep_stats = compute_episode_stats(episode_buffer, self.features)
-        has_video_keys = len(self.meta.video_keys) > 0
-        use_streaming = self._streaming_encoder is not None and has_video_keys
-        use_batched_encoding = self.batch_encoding_size > 1
-
-        if use_streaming:
-            # Compute stats for non-video features only (video stats come from encoder)
-            non_video_buffer = {
-                k: v
-                for k, v in episode_buffer.items()
-                if self.features.get(k, {}).get("dtype") not in ("video",)
-            }
-            non_video_features = {k: v for k, v in self.features.items() if v["dtype"] != "video"}
-            ep_stats = compute_episode_stats(non_video_buffer, non_video_features)
-        else:
-            ep_stats = compute_episode_stats(episode_buffer, self.features)
 
         ep_metadata = self._save_episode_data(episode_buffer)
+        has_video_keys = len(self.meta.video_keys) > 0
+        use_batched_encoding = self.batch_encoding_size > 1
 
-        if use_streaming:
+        if has_video_keys and not use_batched_encoding:
-            # Finish streaming encoding and collect results
-            streaming_results = self._streaming_encoder.finish_episode()
-            for video_key in self.meta.video_keys:
-                temp_path, video_stats = streaming_results[video_key]
-                if video_stats is not None:
-                    # Format stats same as compute_episode_stats: normalize to [0,1], reshape to (C,1,1)
-                    ep_stats[video_key] = {
-                        k: v if k == "count" else np.squeeze(v.reshape(1, -1, 1, 1) / 255.0, axis=0)
-                        for k, v in video_stats.items()
-                    }
-                ep_metadata.update(self._save_episode_video(video_key, episode_index, temp_path=temp_path))
-        elif has_video_keys and not use_batched_encoding:
             num_cameras = len(self.meta.video_keys)
             if parallel_encoding and num_cameras > 1:
                 # TODO(Steven): Ideally we would like to control the number of threads per encoding such that:
@@ -1314,7 +1267,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
                             self.root,
                             self.fps,
                             self.vcodec,
-                            self._encoder_threads,
                         ): video_key
                         for video_key in self.meta.video_keys
                     }
@@ -1583,10 +1535,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return metadata
 
     def clear_episode_buffer(self, delete_images: bool = True) -> None:
-        # Cancel streaming encoder if active
-        if self._streaming_encoder is not None:
-            self._streaming_encoder.cancel_episode()
-
         # Clean up image files for the current episode buffer
         if delete_images:
             # Wait for the async image writer to finish
@@ -1634,9 +1582,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        return _encode_video_worker(
+        return _encode_video_worker(video_key, episode_index, self.root, self.fps, self.vcodec)
-            video_key, episode_index, self.root, self.fps, self.vcodec, self._encoder_threads
-        )
 
     @classmethod
     def create(
@@ -1653,13 +1599,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
-        metadata_buffer_size: int = 10,
-        streaming_encoding: bool = False,
-        encoder_queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
     ) -> "LeRobotDataset":
         """Create a LeRobot Dataset from scratch in order to record data."""
-        vcodec = resolve_vcodec(vcodec)
+        if vcodec not in VALID_VIDEO_CODECS:
+            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         obj = cls.__new__(cls)
         obj.meta = LeRobotDatasetMetadata.create(
             repo_id=repo_id,
@@ -1668,7 +1611,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
             features=features,
             root=root,
             use_videos=use_videos,
-            metadata_buffer_size=metadata_buffer_size,
         )
         obj.repo_id = obj.meta.repo_id
         obj.root = obj.meta.root
@@ -1678,7 +1620,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.batch_encoding_size = batch_encoding_size
         obj.episodes_since_last_encoding = 0
         obj.vcodec = vcodec
-        obj._encoder_threads = encoder_threads
 
         if image_writer_processes or image_writer_threads:
             obj.start_image_writer(image_writer_processes, image_writer_threads)
@@ -1700,22 +1641,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj._lazy_loading = False
         obj._recorded_frames = 0
         obj._writer_closed_for_reading = False
-
-        # Initialize streaming encoder
-        if streaming_encoding and len(obj.meta.video_keys) > 0:
-            obj._streaming_encoder = StreamingVideoEncoder(
-                fps=fps,
-                vcodec=vcodec,
-                pix_fmt="yuv420p",
-                g=2,
-                crf=30,
-                preset=None,
-                queue_maxsize=encoder_queue_maxsize,
-                encoder_threads=encoder_threads,
-            )
-        else:
-            obj._streaming_encoder = None
-
         return obj
 
 
@@ -1771,12 +1696,11 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
             )
         for repo_id, ds in zip(self.repo_ids, self._datasets, strict=True):
             extra_keys = set(ds.features).difference(intersection_features)
-            if extra_keys:
+            logging.warning(
-                logging.warning(
+                f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
-                    f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
+                "other datasets."
-                    "other datasets."
+            )
-                )
+            self.disabled_features.update(extra_keys)
-                self.disabled_features.update(extra_keys)
 
         self.image_transforms = image_transforms
         self.delta_timestamps = delta_timestamps

src/lerobot/datasets/utils.py

@@ -62,6 +62,8 @@ CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
 DEFAULT_TASKS_PATH = "meta/tasks.parquet"
 DEFAULT_SUBTASKS_PATH = "meta/subtasks.parquet"
 DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
+DEFAULT_TASKS_HIGH_LEVEL_PATH = "meta/tasks_high_level.parquet"
+DEFAULT_SUBTASKS_PATH = "meta/subtasks.parquet"
 DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"
 DEFAULT_IMAGE_PATH = "images/{image_key}/episode-{episode_index:06d}/frame-{frame_index:06d}.png"
@@ -122,9 +124,19 @@ def load_nested_dataset(
         raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
 
     with SuppressProgressBars():
-        # We use .from_parquet() memory-mapped loading for efficiency
+        # When no filtering needed, Dataset uses memory-mapped loading for efficiency
-        filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
+        # PyArrow loads the entire dataset into memory
-        return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)
+        if episodes is None:
+            return Dataset.from_parquet([str(path) for path in paths], features=features)
+
+        arrow_dataset = pa_ds.dataset(paths, format="parquet")
+        filter_expr = pa_ds.field("episode_index").isin(episodes)
+        table = arrow_dataset.to_table(filter=filter_expr)
+
+        if features is not None:
+            table = table.cast(features.arrow_schema)
+
+        return Dataset(table)
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:
@@ -343,6 +355,20 @@ def load_tasks(local_dir: Path) -> pandas.DataFrame:
     tasks = pd.read_parquet(local_dir / DEFAULT_TASKS_PATH)
     return tasks
 
+def load_tasks_high_level(local_dir: Path) -> pandas.DataFrame | None:
+    """Load high-level tasks from tasks_high_level.parquet if it exists."""
+    tasks_high_level_path = local_dir / DEFAULT_TASKS_HIGH_LEVEL_PATH
+    if tasks_high_level_path.exists():
+        return pd.read_parquet(tasks_high_level_path)
+    return None
+
+
+def load_subtasks(local_dir: Path) -> pandas.DataFrame | None:
+    """Load subtasks from subtasks.parquet if it exists."""
+    subtasks_path = local_dir / DEFAULT_SUBTASKS_PATH
+    if subtasks_path.exists():
+        return pd.read_parquet(subtasks_path)
+    return None
 
 def load_subtasks(local_dir: Path) -> pandas.DataFrame | None:
     """Load subtasks from subtasks.parquet if it exists."""

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -529,7 +529,7 @@ if __name__ == "__main__":
         type=str,
         required=True,
         help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset "
-        "(e.g. `lerobot/pusht`, `<USER>/aloha_sim_insertion_human`).",
+        "(e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
     )
     parser.add_argument(
         "--branch",

src/lerobot/datasets/video_utils.py

@@ -13,106 +13,25 @@
 # 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 contextlib
 import glob
 import importlib
 import logging
-import queue
 import shutil
 import tempfile
-import threading
 import warnings
 from dataclasses import dataclass, field
-from fractions import Fraction
 from pathlib import Path
 from threading import Lock
 from typing import Any, ClassVar
 
 import av
 import fsspec
-import numpy as np
 import pyarrow as pa
 import torch
 import torchvision
 from datasets.features.features import register_feature
 from PIL import Image
 
-# List of hardware encoders to probe for auto-selection. Availability depends on the platform and FFmpeg build.
-# Determines the order of preference for auto-selection when vcodec="auto" is used.
-HW_ENCODERS = [
-    "h264_videotoolbox",  # macOS
-    "hevc_videotoolbox",  # macOS
-    "h264_nvenc",  # NVIDIA GPU
-    "hevc_nvenc",  # NVIDIA GPU
-    "h264_vaapi",  # Linux Intel/AMD
-    "h264_qsv",  # Intel Quick Sync
-]
-
-VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1", "auto"} | set(HW_ENCODERS)
-
-
-def _get_codec_options(
-    vcodec: str,
-    g: int | None = 2,
-    crf: int | None = 30,
-    preset: int | None = None,
-) -> dict:
-    """Build codec-specific options dict for video encoding."""
-    options = {}
-
-    # GOP size (keyframe interval) - supported by VideoToolbox and software encoders
-    if g is not None and (vcodec in ("h264_videotoolbox", "hevc_videotoolbox") or vcodec not in HW_ENCODERS):
-        options["g"] = str(g)
-
-    # Quality control (codec-specific parameter names)
-    if crf is not None:
-        if vcodec in ("h264", "hevc", "libsvtav1"):
-            options["crf"] = str(crf)
-        elif vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
-            quality = max(1, min(100, int(100 - crf * 2)))
-            options["q:v"] = str(quality)
-        elif vcodec in ("h264_nvenc", "hevc_nvenc"):
-            options["rc"] = "constqp"
-            options["qp"] = str(crf)
-        elif vcodec in ("h264_vaapi",):
-            options["qp"] = str(crf)
-        elif vcodec in ("h264_qsv",):
-            options["global_quality"] = str(crf)
-
-    # Preset (only for libsvtav1)
-    if vcodec == "libsvtav1":
-        options["preset"] = str(preset) if preset is not None else "12"
-
-    return options
-
-
-def detect_available_hw_encoders() -> list[str]:
-    """Probe PyAV/FFmpeg for available hardware video encoders."""
-    available = []
-    for codec_name in HW_ENCODERS:
-        try:
-            av.codec.Codec(codec_name, "w")
-            available.append(codec_name)
-        except Exception:  # nosec B110
-            pass  # nosec B110
-    return available
-
-
-def resolve_vcodec(vcodec: str) -> str:
-    """Validate vcodec and resolve 'auto' to best available HW encoder, fallback to libsvtav1."""
-    if vcodec not in VALID_VIDEO_CODECS:
-        raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
-    if vcodec != "auto":
-        logging.info(f"Using video codec: {vcodec}")
-        return vcodec
-    available = detect_available_hw_encoders()
-    for encoder in HW_ENCODERS:
-        if encoder in available:
-            logging.info(f"Auto-selected video codec: {encoder}")
-            return encoder
-    logging.info("No hardware encoder available, falling back to software encoder 'libsvtav1'")
-    return "libsvtav1"
-
 
 def get_safe_default_codec():
     if importlib.util.find_spec("torchcodec"):
@@ -227,17 +146,16 @@ def decode_video_frames_torchvision(
     min_, argmin_ = dist.min(1)
 
     is_within_tol = min_ < tolerance_s
-    if not is_within_tol.all():
+    assert is_within_tol.all(), (
-        raise FrameTimestampError(
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
-            f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
-            " It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
-            " This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
-            " To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
-            f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
-            f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
-            f"\nvideo: {video_path}"
+        f"\nbackend: {backend}"
-            f"\nbackend: {backend}"
+    )
-        )
 
     # get closest frames to the query timestamps
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -249,11 +167,7 @@ def decode_video_frames_torchvision(
     # convert to the pytorch format which is float32 in [0,1] range (and channel first)
     closest_frames = closest_frames.type(torch.float32) / 255
 
-    if len(timestamps) != len(closest_frames):
+    assert len(timestamps) == len(closest_frames)
-        raise FrameTimestampError(
-            f"Number of retrieved frames ({len(closest_frames)}) does not match "
-            f"number of queried timestamps ({len(timestamps)})"
-        )
     return closest_frames
 
 
@@ -358,16 +272,15 @@ def decode_video_frames_torchcodec(
     min_, argmin_ = dist.min(1)
 
     is_within_tol = min_ < tolerance_s
-    if not is_within_tol.all():
+    assert is_within_tol.all(), (
-        raise FrameTimestampError(
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
-            f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
-            " It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
-            " This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
-            " To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
-            f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
-            f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
-            f"\nvideo: {video_path}"
+    )
-        )
 
     # get closest frames to the query timestamps
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -396,13 +309,14 @@ def encode_video_frames(
     g: int | None = 2,
     crf: int | None = 30,
     fast_decode: int = 0,
-    log_level: int | None = av.logging.WARNING,
+    log_level: int | None = av.logging.ERROR,
     overwrite: bool = False,
     preset: int | None = None,
-    encoder_threads: int | None = None,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
-    vcodec = resolve_vcodec(vcodec)
+    # Check encoder availability
+    if vcodec not in ["h264", "hevc", "libsvtav1"]:
+        raise ValueError(f"Unsupported video codec: {vcodec}. Supported codecs are: h264, hevc, libsvtav1.")
 
     video_path = Path(video_path)
     imgs_dir = Path(imgs_dir)
@@ -433,22 +347,21 @@ def encode_video_frames(
         width, height = dummy_image.size
 
     # Define video codec options
-    video_options = _get_codec_options(vcodec, g, crf, preset)
+    video_options = {}
+
+    if g is not None:
+        video_options["g"] = str(g)
+
+    if crf is not None:
+        video_options["crf"] = str(crf)
 
     if fast_decode:
         key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
         value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
         video_options[key] = value
 
-    if encoder_threads is not None:
+    if vcodec == "libsvtav1":
-        if vcodec == "libsvtav1":
+        video_options["preset"] = str(preset) if preset is not None else "12"
-            lp_param = f"lp={encoder_threads}"
-            if "svtav1-params" in video_options:
-                video_options["svtav1-params"] += f":{lp_param}"
-            else:
-                video_options["svtav1-params"] = lp_param
-        else:
-            video_options["threads"] = str(encoder_threads)
 
     # Set logging level
     if log_level is not None:
@@ -567,348 +480,6 @@ def concatenate_video_files(
     Path(tmp_concatenate_path).unlink()
 
 
-class _CameraEncoderThread(threading.Thread):
-    """A thread that encodes video frames streamed via a queue into an MP4 file.
-
-    One instance is created per camera per episode. Frames are received as numpy arrays
-    from the main thread, encoded in real-time using PyAV (which releases the GIL during
-    encoding), and written to disk. Stats are computed incrementally using
-    RunningQuantileStats and returned via result_queue.
-    """
-
-    def __init__(
-        self,
-        video_path: Path,
-        fps: int,
-        vcodec: str,
-        pix_fmt: str,
-        g: int | None,
-        crf: int | None,
-        preset: int | None,
-        frame_queue: queue.Queue,
-        result_queue: queue.Queue,
-        stop_event: threading.Event,
-        encoder_threads: int | None = None,
-    ):
-        super().__init__(daemon=True)
-        self.video_path = video_path
-        self.fps = fps
-        self.vcodec = vcodec
-        self.pix_fmt = pix_fmt
-        self.g = g
-        self.crf = crf
-        self.preset = preset
-        self.frame_queue = frame_queue
-        self.result_queue = result_queue
-        self.stop_event = stop_event
-        self.encoder_threads = encoder_threads
-
-    def run(self) -> None:
-        from lerobot.datasets.compute_stats import RunningQuantileStats, auto_downsample_height_width
-
-        container = None
-        output_stream = None
-        stats_tracker = RunningQuantileStats()
-        frame_count = 0
-
-        try:
-            logging.getLogger("libav").setLevel(av.logging.WARNING)
-
-            while True:
-                try:
-                    frame_data = self.frame_queue.get(timeout=1)
-                except queue.Empty:
-                    if self.stop_event.is_set():
-                        break
-                    continue
-
-                if frame_data is None:
-                    # Sentinel: flush and close
-                    break
-
-                # Ensure HWC uint8 numpy array
-                if isinstance(frame_data, np.ndarray):
-                    if frame_data.ndim == 3 and frame_data.shape[0] == 3:
-                        # CHW -> HWC
-                        frame_data = frame_data.transpose(1, 2, 0)
-                    if frame_data.dtype != np.uint8:
-                        frame_data = (frame_data * 255).astype(np.uint8)
-
-                # Open container on first frame (to get width/height)
-                if container is None:
-                    height, width = frame_data.shape[:2]
-                    video_options = _get_codec_options(self.vcodec, self.g, self.crf, self.preset)
-                    if self.encoder_threads is not None:
-                        if self.vcodec == "libsvtav1":
-                            lp_param = f"lp={self.encoder_threads}"
-                            if "svtav1-params" in video_options:
-                                video_options["svtav1-params"] += f":{lp_param}"
-                            else:
-                                video_options["svtav1-params"] = lp_param
-                        else:
-                            video_options["threads"] = str(self.encoder_threads)
-                    Path(self.video_path).parent.mkdir(parents=True, exist_ok=True)
-                    container = av.open(str(self.video_path), "w")
-                    output_stream = container.add_stream(self.vcodec, self.fps, options=video_options)
-                    output_stream.pix_fmt = self.pix_fmt
-                    output_stream.width = width
-                    output_stream.height = height
-                    output_stream.time_base = Fraction(1, self.fps)
-
-                # Encode frame with explicit timestamps
-                pil_img = Image.fromarray(frame_data)
-                video_frame = av.VideoFrame.from_image(pil_img)
-                video_frame.pts = frame_count
-                video_frame.time_base = Fraction(1, self.fps)
-                packet = output_stream.encode(video_frame)
-                if packet:
-                    container.mux(packet)
-
-                # Update stats with downsampled frame (per-channel stats like compute_episode_stats)
-                img_chw = frame_data.transpose(2, 0, 1)  # HWC -> CHW
-                img_downsampled = auto_downsample_height_width(img_chw)
-                # Reshape CHW to (H*W, C) for per-channel stats
-                channels = img_downsampled.shape[0]
-                img_for_stats = img_downsampled.transpose(1, 2, 0).reshape(-1, channels)
-                stats_tracker.update(img_for_stats)
-
-                frame_count += 1
-
-            # Flush encoder
-            if output_stream is not None:
-                packet = output_stream.encode()
-                if packet:
-                    container.mux(packet)
-
-            if container is not None:
-                container.close()
-
-            av.logging.restore_default_callback()
-
-            # Get stats and put on result queue
-            if frame_count >= 2:
-                stats = stats_tracker.get_statistics()
-                self.result_queue.put(("ok", stats))
-            else:
-                self.result_queue.put(("ok", None))
-
-        except Exception as e:
-            logging.error(f"Encoder thread error: {e}")
-            if container is not None:
-                with contextlib.suppress(Exception):
-                    container.close()
-            self.result_queue.put(("error", str(e)))
-
-
-class StreamingVideoEncoder:
-    """Manages per-camera encoder threads for real-time video encoding during recording.
-
-    Instead of writing frames as PNG images and then encoding to MP4 at episode end,
-    this class streams frames directly to encoder threads, eliminating the
-    PNG round-trip and making save_episode() near-instant.
-
-    Uses threading instead of multiprocessing to avoid the overhead of pickling large
-    numpy arrays through multiprocessing.Queue. PyAV's encode() releases the GIL,
-    so encoding runs in parallel with the main recording loop.
-    """
-
-    def __init__(
-        self,
-        fps: int,
-        vcodec: str = "libsvtav1",
-        pix_fmt: str = "yuv420p",
-        g: int | None = 2,
-        crf: int | None = 30,
-        preset: int | None = None,
-        queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
-    ):
-        self.fps = fps
-        self.vcodec = resolve_vcodec(vcodec)
-        self.pix_fmt = pix_fmt
-        self.g = g
-        self.crf = crf
-        self.preset = preset
-        self.queue_maxsize = queue_maxsize
-        self.encoder_threads = encoder_threads
-
-        self._frame_queues: dict[str, queue.Queue] = {}
-        self._result_queues: dict[str, queue.Queue] = {}
-        self._threads: dict[str, _CameraEncoderThread] = {}
-        self._stop_events: dict[str, threading.Event] = {}
-        self._video_paths: dict[str, Path] = {}
-        self._dropped_frames: dict[str, int] = {}
-        self._episode_active = False
-
-    def start_episode(self, video_keys: list[str], temp_dir: Path) -> None:
-        """Start encoder threads for a new episode.
-
-        Args:
-            video_keys: List of video feature keys (e.g. ["observation.images.laptop"])
-            temp_dir: Base directory for temporary MP4 files
-        """
-        if self._episode_active:
-            self.cancel_episode()
-
-        self._dropped_frames.clear()
-
-        for video_key in video_keys:
-            frame_queue: queue.Queue = queue.Queue(maxsize=self.queue_maxsize)
-            result_queue: queue.Queue = queue.Queue(maxsize=1)
-            stop_event = threading.Event()
-
-            temp_video_dir = Path(tempfile.mkdtemp(dir=temp_dir))
-            video_path = temp_video_dir / f"{video_key.replace('/', '_')}_streaming.mp4"
-
-            encoder_thread = _CameraEncoderThread(
-                video_path=video_path,
-                fps=self.fps,
-                vcodec=self.vcodec,
-                pix_fmt=self.pix_fmt,
-                g=self.g,
-                crf=self.crf,
-                preset=self.preset,
-                frame_queue=frame_queue,
-                result_queue=result_queue,
-                stop_event=stop_event,
-                encoder_threads=self.encoder_threads,
-            )
-            encoder_thread.start()
-
-            self._frame_queues[video_key] = frame_queue
-            self._result_queues[video_key] = result_queue
-            self._threads[video_key] = encoder_thread
-            self._stop_events[video_key] = stop_event
-            self._video_paths[video_key] = video_path
-
-        self._episode_active = True
-
-    def feed_frame(self, video_key: str, image: np.ndarray) -> None:
-        """Feed a frame to the encoder for a specific camera.
-
-        A copy of the image is made before enqueueing to prevent race conditions
-        with camera drivers that may reuse buffers. If the encoder queue is full
-        (encoder can't keep up), the frame is dropped with a warning instead of
-        crashing the recording session.
-
-        Args:
-            video_key: The video feature key
-            image: numpy array in (H,W,C) or (C,H,W) format, uint8 or float
-
-        Raises:
-            RuntimeError: If the encoder thread has crashed
-        """
-        if not self._episode_active:
-            raise RuntimeError("No active episode. Call start_episode() first.")
-
-        thread = self._threads[video_key]
-        if not thread.is_alive():
-            # Check for error
-            try:
-                status, msg = self._result_queues[video_key].get_nowait()
-                if status == "error":
-                    raise RuntimeError(f"Encoder thread for {video_key} crashed: {msg}")
-            except queue.Empty:
-                pass
-            raise RuntimeError(f"Encoder thread for {video_key} is not alive")
-
-        try:
-            self._frame_queues[video_key].put(image.copy(), timeout=0.1)
-        except queue.Full:
-            self._dropped_frames[video_key] = self._dropped_frames.get(video_key, 0) + 1
-            count = self._dropped_frames[video_key]
-            # Log periodically to avoid spam (1st, then every 10th)
-            if count == 1 or count % 10 == 0:
-                logging.warning(
-                    f"Encoder queue full for {video_key}, dropped {count} frame(s). "
-                    f"Consider using vcodec='auto' for hardware encoding or increasing encoder_queue_maxsize."
-                )
-
-    def finish_episode(self) -> dict[str, tuple[Path, dict | None]]:
-        """Finish encoding the current episode.
-
-        Sends sentinel values, waits for encoder threads to complete,
-        and collects results.
-
-        Returns:
-            Dict mapping video_key to (mp4_path, stats_dict_or_None)
-        """
-        if not self._episode_active:
-            raise RuntimeError("No active episode to finish.")
-
-        results = {}
-
-        # Report dropped frames
-        for video_key, count in self._dropped_frames.items():
-            if count > 0:
-                logging.warning(f"Episode finished with {count} dropped frame(s) for {video_key}.")
-
-        # Send sentinel to all queues
-        for video_key in self._frame_queues:
-            self._frame_queues[video_key].put(None)
-
-        # Wait for all threads and collect results
-        for video_key in self._threads:
-            self._threads[video_key].join(timeout=120)
-            if self._threads[video_key].is_alive():
-                logging.error(f"Encoder thread for {video_key} did not finish in time")
-                self._stop_events[video_key].set()
-                self._threads[video_key].join(timeout=5)
-                results[video_key] = (self._video_paths[video_key], None)
-                continue
-
-            try:
-                status, data = self._result_queues[video_key].get(timeout=5)
-                if status == "error":
-                    raise RuntimeError(f"Encoder thread for {video_key} failed: {data}")
-                results[video_key] = (self._video_paths[video_key], data)
-            except queue.Empty:
-                logging.error(f"No result from encoder thread for {video_key}")
-                results[video_key] = (self._video_paths[video_key], None)
-
-        self._cleanup()
-        self._episode_active = False
-        return results
-
-    def cancel_episode(self) -> None:
-        """Cancel the current episode, stopping encoder threads and cleaning up."""
-        if not self._episode_active:
-            return
-
-        # Signal all threads to stop
-        for video_key in self._stop_events:
-            self._stop_events[video_key].set()
-
-        # Wait for threads to finish
-        for video_key in self._threads:
-            self._threads[video_key].join(timeout=5)
-
-            # Clean up temp MP4 files
-            video_path = self._video_paths.get(video_key)
-            if video_path is not None and video_path.exists():
-                shutil.rmtree(str(video_path.parent), ignore_errors=True)
-
-        self._cleanup()
-        self._episode_active = False
-
-    def close(self) -> None:
-        """Close the encoder, canceling any in-progress episode."""
-        if self._episode_active:
-            self.cancel_episode()
-
-    def _cleanup(self) -> None:
-        """Clean up queues and thread tracking dicts."""
-        for q in self._frame_queues.values():
-            with contextlib.suppress(Exception):
-                while not q.empty():
-                    q.get_nowait()
-        self._frame_queues.clear()
-        self._result_queues.clear()
-        self._threads.clear()
-        self._stop_events.clear()
-        self._video_paths.clear()
-
-
 @dataclass
 class VideoFrame:
     # TODO(rcadene, lhoestq): move to Hugging Face `datasets` repo
@@ -943,7 +514,7 @@ with warnings.catch_warnings():
 
 def get_audio_info(video_path: Path | str) -> dict:
     # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.WARNING)
+    logging.getLogger("libav").setLevel(av.logging.ERROR)
 
     # Getting audio stream information
     audio_info = {}
@@ -975,7 +546,7 @@ def get_audio_info(video_path: Path | str) -> dict:
 
 def get_video_info(video_path: Path | str) -> dict:
     # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.WARNING)
+    logging.getLogger("libav").setLevel(av.logging.ERROR)
 
     # Getting video stream information
     video_info = {}
@@ -1061,15 +632,8 @@ class VideoEncodingManager:
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
-        streaming_encoder = getattr(self.dataset, "_streaming_encoder", None)
+        # Handle any remaining episodes that haven't been batch encoded
-
+        if self.dataset.episodes_since_last_encoding > 0:
-        if streaming_encoder is not None:
-            # Handle streaming encoder cleanup
-            if exc_type is not None:
-                streaming_encoder.cancel_episode()
-            streaming_encoder.close()
-        elif self.dataset.episodes_since_last_encoding > 0:
-            # Handle any remaining episodes that haven't been batch encoded
             if exc_type is not None:
                 logging.info("Exception occurred. Encoding remaining episodes before exit...")
             else:
@@ -1086,8 +650,8 @@ class VideoEncodingManager:
         # Finalize the dataset to properly close all writers
         self.dataset.finalize()
 
-        # Clean up episode images if recording was interrupted (only for non-streaming mode)
+        # Clean up episode images if recording was interrupted
-        if exc_type is not None and streaming_encoder is None:
+        if exc_type is not None:
             interrupted_episode_index = self.dataset.num_episodes
             for key in self.dataset.meta.video_keys:
                 img_dir = self.dataset._get_image_file_path(
@@ -1101,12 +665,14 @@ class VideoEncodingManager:
 
         # Clean up any remaining images directory if it's empty
         img_dir = self.dataset.root / "images"
-        if img_dir.exists():
+        # Check for any remaining PNG files
-            png_files = list(img_dir.rglob("*.png"))
+        png_files = list(img_dir.rglob("*.png"))
-            if len(png_files) == 0:
+        if len(png_files) == 0:
+            # Only remove the images directory if no PNG files remain
+            if img_dir.exists():
                 shutil.rmtree(img_dir)
                 logging.debug("Cleaned up empty images directory")
-            else:
+        else:
-                logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
+            logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
 
         return False  # Don't suppress the original exception

src/lerobot/envs/libero.py

@@ -112,7 +112,6 @@ class LiberoEnv(gym.Env):
         visualization_height: int = 480,
         init_states: bool = True,
         episode_index: int = 0,
-        n_envs: int = 1,
         camera_name_mapping: dict[str, str] | None = None,
         num_steps_wait: int = 10,
         control_mode: str = "relative",
@@ -146,9 +145,7 @@ class LiberoEnv(gym.Env):
         self.episode_length = episode_length
         # Load once and keep
         self._init_states = get_task_init_states(task_suite, self.task_id) if self.init_states else None
-        self._reset_stride = n_envs  # when performing a reset, append `_reset_stride` to `init_state_id`.
+        self._init_state_id = self.episode_index  # tie each sub-env to a fixed init state
-
-        self.init_state_id = self.episode_index  # tie each sub-env to a fixed init state
 
         self._env = self._make_envs_task(task_suite, self.task_id)
         default_steps = 500
@@ -298,8 +295,7 @@ class LiberoEnv(gym.Env):
         self._env.seed(seed)
         raw_obs = self._env.reset()
         if self.init_states and self._init_states is not None:
-            raw_obs = self._env.set_init_state(self._init_states[self.init_state_id % len(self._init_states)])
+            raw_obs = self._env.set_init_state(self._init_states[self._init_state_id])
-            self.init_state_id += self._reset_stride  # Change init_state_id when reset
 
         # After reset, objects may be unstable (slightly floating, intersecting, etc.).
         # Step the simulator with a no-op action for a few frames so everything settles.
@@ -377,7 +373,6 @@ def _make_env_fns(
             init_states=init_states,
             episode_length=episode_length,
             episode_index=episode_index,
-            n_envs=n_envs,
             control_mode=control_mode,
             **local_kwargs,
         )

src/lerobot/motors/calibration_gui.py

@@ -221,7 +221,7 @@ class RangeFinderGUI:
 
         self.bus = bus
         self.groups = groups if groups is not None else {"all": list(bus.motors)}
-        self.group_names = list(self.groups)
+        self.group_names = list(groups)
         self.current_group = self.group_names[0]
 
         if not bus.is_connected:
@@ -230,20 +230,18 @@ class RangeFinderGUI:
         self.calibration = bus.read_calibration()
         self.res_table = bus.model_resolution_table
         self.present_cache = {
-            m: bus.read("Present_Position", m, normalize=False)
+            m: bus.read("Present_Position", m, normalize=False) for motors in groups.values() for m in motors
-            for motors in self.groups.values()
-            for m in motors
         }
 
         pygame.init()
         self.font = pygame.font.Font(None, FONT_SIZE)
 
-        label_pad = max(self.font.size(m)[0] for ms in self.groups.values() for m in ms)
+        label_pad = max(self.font.size(m)[0] for ms in groups.values() for m in ms)
         self.label_pad = label_pad
         width = 40 + label_pad + BAR_LEN + 6 + BTN_W + 10 + SAVE_W + 10
         self.controls_bottom = 10 + SAVE_H
         self.base_y = self.controls_bottom + TOP_GAP
-        height = self.base_y + PADDING_Y * len(self.groups[self.current_group]) + 40
+        height = self.base_y + PADDING_Y * len(groups[self.current_group]) + 40
 
         self.screen = pygame.display.set_mode((width, height))
         pygame.display.set_caption("Motors range finder")

src/lerobot/motors/damiao/damiao.py

@@ -23,7 +23,6 @@ from copy import deepcopy
 from functools import cached_property
 from typing import TYPE_CHECKING, Any, TypedDict
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.import_utils import _can_available
 
 if TYPE_CHECKING or _can_available:
@@ -37,6 +36,7 @@ else:
 
 import numpy as np
 
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import enter_pressed, move_cursor_up
 
@@ -155,7 +155,6 @@ class DamiaoMotorsBus(MotorsBusBase):
         """Check if the CAN bus is connected."""
         return self._is_connected and self.canbus is not None
 
-    @check_if_already_connected
     def connect(self, handshake: bool = True) -> None:
         """
         Open the CAN bus and initialize communication.
@@ -163,6 +162,10 @@ class DamiaoMotorsBus(MotorsBusBase):
         Args:
             handshake: If True, ping all motors to verify they're present
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(
+                f"{self.__class__.__name__}('{self.port}') is already connected."
+            )
 
         try:
             # Auto-detect interface type based on port name
@@ -208,9 +211,6 @@ class DamiaoMotorsBus(MotorsBusBase):
         logger.info("Starting handshake with motors...")
 
         # Drain any pending messages
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
         while self.canbus.recv(timeout=0.01):
             pass
 
@@ -246,7 +246,6 @@ class DamiaoMotorsBus(MotorsBusBase):
             )
         logger.info("Handshake successful. All motors ready.")
 
-    @check_if_not_connected
     def disconnect(self, disable_torque: bool = True) -> None:
         """
         Close the CAN bus connection.
@@ -254,6 +253,8 @@ class DamiaoMotorsBus(MotorsBusBase):
         Args:
             disable_torque: If True, disable torque on all motors before disconnecting
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self.__class__.__name__}('{self.port}') is not connected.")
 
         if disable_torque:
             try:
@@ -282,10 +283,6 @@ class DamiaoMotorsBus(MotorsBusBase):
         recv_id = self._get_motor_recv_id(motor)
         data = [0xFF] * 7 + [command_byte]
         msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
-
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
         self.canbus.send(msg)
         if msg := self._recv_motor_response(expected_recv_id=recv_id):
             self._process_response(motor_name, msg)
@@ -344,10 +341,6 @@ class DamiaoMotorsBus(MotorsBusBase):
         recv_id = self._get_motor_recv_id(motor)
         data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
         msg = can.Message(arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False, is_fd=self.use_can_fd)
-
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
         self.canbus.send(msg)
         return self._recv_motor_response(expected_recv_id=recv_id)
 
@@ -363,10 +356,6 @@ class DamiaoMotorsBus(MotorsBusBase):
         Returns:
             CAN message if received, None otherwise
         """
-
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
         try:
             start_time = time.time()
             messages_seen = []
@@ -405,13 +394,10 @@ class DamiaoMotorsBus(MotorsBusBase):
         Returns:
             Dictionary mapping recv_id to CAN message
         """
-        responses: dict[int, can.Message] = {}
+        responses = {}
         expected_set = set(expected_recv_ids)
         start_time = time.time()
 
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
         try:
             while len(responses) < len(expected_recv_ids) and (time.time() - start_time) < timeout:
                 # 100us poll timeout
@@ -475,9 +461,6 @@ class DamiaoMotorsBus(MotorsBusBase):
         motor_name = self._get_motor_name(motor)
         motor_type = self._motor_types[motor_name]
 
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
         data = self._encode_mit_packet(motor_type, kp, kd, position_degrees, velocity_deg_per_sec, torque)
         msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
         self.canbus.send(msg)
@@ -505,9 +488,6 @@ class DamiaoMotorsBus(MotorsBusBase):
 
         recv_id_to_motor: dict[int, str] = {}
 
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
         # Step 1: Send all MIT control commands
         for motor, (kp, kd, position_degrees, velocity_deg_per_sec, torque) in commands.items():
             motor_id = self._get_motor_id(motor)
@@ -582,9 +562,10 @@ class DamiaoMotorsBus(MotorsBusBase):
         except Exception as e:
             logger.warning(f"Failed to decode response from {motor}: {e}")
 
-    @check_if_not_connected
     def read(self, data_name: str, motor: str) -> Value:
         """Read a value from a single motor. Positions are always in degrees."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         # Refresh motor to get latest state
         msg = self._refresh_motor(motor)
@@ -614,7 +595,6 @@ class DamiaoMotorsBus(MotorsBusBase):
             raise ValueError(f"Unknown data_name: {data_name}")
         return mapping[data_name]
 
-    @check_if_not_connected
     def write(
         self,
         data_name: str,
@@ -625,6 +605,8 @@ class DamiaoMotorsBus(MotorsBusBase):
         Write a value to a single motor. Positions are always in degrees.
         Can write 'Goal_Position', 'Kp', or 'Kd'.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if data_name in ("Kp", "Kd"):
             self._gains[motor][data_name.lower()] = float(value)
@@ -674,10 +656,6 @@ class DamiaoMotorsBus(MotorsBusBase):
 
     def _batch_refresh(self, motors: list[str]) -> None:
         """Internal helper to refresh a list of motors and update cache."""
-
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
         # Send refresh commands
         for motor in motors:
             motor_id = self._get_motor_id(motor)
@@ -700,12 +678,10 @@ class DamiaoMotorsBus(MotorsBusBase):
             else:
                 logger.warning(f"Packet drop: {motor} (ID: 0x{recv_id:02X}). Using last known state.")
 
-    @check_if_not_connected
+    def sync_write(self, data_name: str, values: Value | dict[str, Value]) -> None:
-    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
         """
         Write values to multiple motors simultaneously. Positions are always in degrees.
         """
-
         if data_name in ("Kp", "Kd"):
             key = data_name.lower()
             for motor, val in values.items():
@@ -714,8 +690,6 @@ class DamiaoMotorsBus(MotorsBusBase):
         elif data_name == "Goal_Position":
             # Step 1: Send all MIT control commands
             recv_id_to_motor: dict[int, str] = {}
-            if self.canbus is None:
-                raise RuntimeError("CAN bus is not initialized.")
             for motor, value_degrees in values.items():
                 motor_id = self._get_motor_id(motor)
                 motor_name = self._get_motor_name(motor)
@@ -758,9 +732,9 @@ class DamiaoMotorsBus(MotorsBusBase):
 
     def record_ranges_of_motion(
         self,
-        motors: str | list[str] | None = None,
+        motors: NameOrID | list[NameOrID] | None = None,
         display_values: bool = True,
-    ) -> tuple[dict[str, Value], dict[str, Value]]:
+    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
         """
         Interactively record the min/max values of each motor in degrees.
 

src/lerobot/motors/dynamixel/dynamixel.py

@@ -181,10 +181,10 @@ class DynamixelMotorsBus(SerialMotorsBus):
         for motor, m in self.motors.items():
             calibration[motor] = MotorCalibration(
                 id=m.id,
-                drive_mode=int(drive_modes[motor]),
+                drive_mode=drive_modes[motor],
-                homing_offset=int(offsets[motor]),
+                homing_offset=offsets[motor],
-                range_min=int(mins[motor]),
+                range_min=mins[motor],
-                range_max=int(maxes[motor]),
+                range_max=maxes[motor],
             )
 
         return calibration
@@ -198,7 +198,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
         if cache:
             self.calibration = calibration_dict
 
-    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
 
@@ -206,7 +206,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
         addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
         self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)
 
-    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
 
@@ -235,7 +235,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
         On Dynamixel Motors:
         Present_Position = Actual_Position + Homing_Offset
         """
-        half_turn_homings: dict[NameOrID, Value] = {}
+        half_turn_homings = {}
         for motor, pos in positions.items():
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
@@ -258,6 +258,6 @@ class DynamixelMotorsBus(SerialMotorsBus):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
 
-            return None
+            return
 
         return {id_: data[0] for id_, data in data_list.items()}

src/lerobot/motors/feetech/feetech.py

@@ -126,7 +126,7 @@ class FeetechMotorsBus(SerialMotorsBus):
 
         self.port_handler = scs.PortHandler(self.port)
         # HACK: monkeypatch
-        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(  # type: ignore[method-assign]
+        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(
             self.port_handler, scs.PortHandler
         )
         self.packet_handler = scs.PacketHandler(protocol_version)
@@ -262,9 +262,9 @@ class FeetechMotorsBus(SerialMotorsBus):
             calibration[motor] = MotorCalibration(
                 id=m.id,
                 drive_mode=0,
-                homing_offset=int(offsets[motor]),
+                homing_offset=offsets[motor],
-                range_min=int(mins[motor]),
+                range_min=mins[motor],
-                range_max=int(maxes[motor]),
+                range_max=maxes[motor],
             )
 
         return calibration
@@ -284,7 +284,7 @@ class FeetechMotorsBus(SerialMotorsBus):
         On Feetech Motors:
         Present_Position = Actual_Position - Homing_Offset
         """
-        half_turn_homings: dict[NameOrID, Value] = {}
+        half_turn_homings = {}
         for motor, pos in positions.items():
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
@@ -292,7 +292,7 @@ class FeetechMotorsBus(SerialMotorsBus):
 
         return half_turn_homings
 
-    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
             self.write("Lock", motor, 0, num_retry=num_retry)
@@ -303,7 +303,7 @@ class FeetechMotorsBus(SerialMotorsBus):
         addr, length = get_address(self.model_ctrl_table, model, "Lock")
         self._write(addr, length, motor, 0, num_retry=num_retry)
 
-    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
             self.write("Lock", motor, 1, num_retry=num_retry)
@@ -334,7 +334,7 @@ class FeetechMotorsBus(SerialMotorsBus):
     def _broadcast_ping(self) -> tuple[dict[int, int], int]:
         import scservo_sdk as scs
 
-        data_list: dict[int, int] = {}
+        data_list = {}
 
         status_length = 6
 
@@ -414,7 +414,7 @@ class FeetechMotorsBus(SerialMotorsBus):
         if not self._is_comm_success(comm):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
-            return None
+            return
 
         ids_errors = {id_: status for id_, status in ids_status.items() if self._is_error(status)}
         if ids_errors:

src/lerobot/motors/motors_bus.py

@@ -23,7 +23,6 @@ from __future__ import annotations
 
 import abc
 import logging
-from collections.abc import Sequence
 from contextlib import contextmanager
 from dataclasses import dataclass
 from enum import Enum
@@ -94,7 +93,7 @@ class MotorsBusBase(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
+    def sync_write(self, data_name: str, values: Value | dict[str, Value]) -> None:
         """Write values to multiple motors."""
         pass
 
@@ -180,16 +179,15 @@ class Motor:
 
 
 class PortHandler(Protocol):
-    is_open: bool
+    def __init__(self, port_name):
-    baudrate: int
+        self.is_open: bool
-    packet_start_time: float
+        self.baudrate: int
-    packet_timeout: float
+        self.packet_start_time: float
-    tx_time_per_byte: float
+        self.packet_timeout: float
-    is_using: bool
+        self.tx_time_per_byte: float
-    port_name: str
+        self.is_using: bool
-    ser: serial.Serial
+        self.port_name: str
-
+        self.ser: serial.Serial
-    def __init__(self, port_name: str) -> None: ...
 
     def openPort(self): ...
     def closePort(self): ...
@@ -242,22 +240,19 @@ class PacketHandler(Protocol):
     def regWriteTxRx(self, port, id, address, length, data): ...
     def syncReadTx(self, port, start_address, data_length, param, param_length): ...
     def syncWriteTxOnly(self, port, start_address, data_length, param, param_length): ...
-    def broadcastPing(self, port): ...
 
 
 class GroupSyncRead(Protocol):
-    port: str
+    def __init__(self, port, ph, start_address, data_length):
-    ph: PortHandler
+        self.port: str
-    start_address: int
+        self.ph: PortHandler
-    data_length: int
+        self.start_address: int
-    last_result: bool
+        self.data_length: int
-    is_param_changed: bool
+        self.last_result: bool
-    param: list
+        self.is_param_changed: bool
-    data_dict: dict
+        self.param: list
+        self.data_dict: dict
 
-    def __init__(
-        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
-    ) -> None: ...
     def makeParam(self): ...
     def addParam(self, id): ...
     def removeParam(self, id): ...
@@ -270,17 +265,15 @@ class GroupSyncRead(Protocol):
 
 
 class GroupSyncWrite(Protocol):
-    port: str
+    def __init__(self, port, ph, start_address, data_length):
-    ph: PortHandler
+        self.port: str
-    start_address: int
+        self.ph: PortHandler
-    data_length: int
+        self.start_address: int
-    is_param_changed: bool
+        self.data_length: int
-    param: list
+        self.is_param_changed: bool
-    data_dict: dict
+        self.param: list
+        self.data_dict: dict
 
-    def __init__(
-        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
-    ) -> None: ...
     def makeParam(self): ...
     def addParam(self, id, data): ...
     def removeParam(self, id): ...
@@ -407,7 +400,7 @@ class SerialMotorsBus(MotorsBusBase):
         else:
             raise TypeError(f"'{motor}' should be int, str.")
 
-    def _get_motor_model(self, motor: NameOrID) -> str:
+    def _get_motor_model(self, motor: NameOrID) -> int:
         if isinstance(motor, str):
             return self.motors[motor].model
         elif isinstance(motor, int):
@@ -415,19 +408,17 @@ class SerialMotorsBus(MotorsBusBase):
         else:
             raise TypeError(f"'{motor}' should be int, str.")
 
-    def _get_motors_list(self, motors: NameOrID | Sequence[NameOrID] | None) -> list[str]:
+    def _get_motors_list(self, motors: str | list[str] | None) -> list[str]:
         if motors is None:
             return list(self.motors)
         elif isinstance(motors, str):
             return [motors]
-        elif isinstance(motors, int):
+        elif isinstance(motors, list):
-            return [self._id_to_name(motors)]
+            return motors.copy()
-        elif isinstance(motors, Sequence):
-            return [m if isinstance(m, str) else self._id_to_name(m) for m in motors]
         else:
             raise TypeError(motors)
 
-    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> dict[int, Value]:
+    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> list[str]:
         if isinstance(values, (int | float)):
             return dict.fromkeys(self.ids, values)
         elif isinstance(values, dict):
@@ -649,19 +640,18 @@ class SerialMotorsBus(MotorsBusBase):
         pass
 
     @abc.abstractmethod
-    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         """Enable torque on selected motors.
 
         Args:
-            motors (int | str | list[str] | None, optional): Same semantics as :pymeth:`disable_torque`.
+            motor (int): Same semantics as :pymeth:`disable_torque`. Defaults to `None`.
-                Defaults to `None`.
             num_retry (int, optional): Number of additional retry attempts on communication failure.
                 Defaults to 0.
         """
         pass
 
     @contextmanager
-    def torque_disabled(self, motors: str | list[str] | None = None):
+    def torque_disabled(self, motors: int | str | list[str] | None = None):
         """Context-manager that guarantees torque is re-enabled.
 
         This helper is useful to temporarily disable torque when configuring motors.
@@ -738,19 +728,24 @@ class SerialMotorsBus(MotorsBusBase):
         """
         pass
 
-    def reset_calibration(self, motors: NameOrID | Sequence[NameOrID] | None = None) -> None:
+    def reset_calibration(self, motors: NameOrID | list[NameOrID] | None = None) -> None:
         """Restore factory calibration for the selected motors.
 
         Homing offset is set to ``0`` and min/max position limits are set to the full usable range.
         The in-memory :pyattr:`calibration` is cleared.
 
         Args:
-            motors (NameOrID | Sequence[NameOrID] | None, optional): Selection of motors. `None` (default)
+            motors (NameOrID | list[NameOrID] | None, optional): Selection of motors. `None` (default)
                 resets every motor.
         """
-        motor_names = self._get_motors_list(motors)
+        if motors is None:
+            motors = list(self.motors)
+        elif isinstance(motors, (str | int)):
+            motors = [motors]
+        elif not isinstance(motors, list):
+            raise TypeError(motors)
 
-        for motor in motor_names:
+        for motor in motors:
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
             self.write("Homing_Offset", motor, 0, normalize=False)
@@ -759,9 +754,7 @@ class SerialMotorsBus(MotorsBusBase):
 
         self.calibration = {}
 
-    def set_half_turn_homings(
+    def set_half_turn_homings(self, motors: NameOrID | list[NameOrID] | None = None) -> dict[NameOrID, Value]:
-        self, motors: NameOrID | Sequence[NameOrID] | None = None
-    ) -> dict[NameOrID, Value]:
         """Centre each motor range around its current position.
 
         The function computes and writes a homing offset such that the present position becomes exactly one
@@ -771,12 +764,17 @@ class SerialMotorsBus(MotorsBusBase):
             motors (NameOrID | list[NameOrID] | None, optional): Motors to adjust. Defaults to all motors (`None`).
 
         Returns:
-            dict[str, Value]: Mapping *motor name → written homing offset*.
+            dict[NameOrID, Value]: Mapping *motor → written homing offset*.
         """
-        motor_names = self._get_motors_list(motors)
+        if motors is None:
+            motors = list(self.motors)
+        elif isinstance(motors, (str | int)):
+            motors = [motors]
+        elif not isinstance(motors, list):
+            raise TypeError(motors)
 
-        self.reset_calibration(motor_names)
+        self.reset_calibration(motors)
-        actual_positions = self.sync_read("Present_Position", motor_names, normalize=False)
+        actual_positions = self.sync_read("Present_Position", motors, normalize=False)
         homing_offsets = self._get_half_turn_homings(actual_positions)
         for motor, offset in homing_offsets.items():
             self.write("Homing_Offset", motor, offset)
@@ -788,8 +786,8 @@ class SerialMotorsBus(MotorsBusBase):
         pass
 
     def record_ranges_of_motion(
-        self, motors: NameOrID | Sequence[NameOrID] | None = None, display_values: bool = True
+        self, motors: NameOrID | list[NameOrID] | None = None, display_values: bool = True
-    ) -> tuple[dict[str, Value], dict[str, Value]]:
+    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
         """Interactively record the min/max encoder values of each motor.
 
         Move the joints by hand (with torque disabled) while the method streams live positions. Press
@@ -801,25 +799,30 @@ class SerialMotorsBus(MotorsBusBase):
             display_values (bool, optional): When `True` (default) a live table is printed to the console.
 
         Returns:
-            tuple[dict[str, Value], dict[str, Value]]: Two dictionaries *mins* and *maxes* with the
+            tuple[dict[NameOrID, Value], dict[NameOrID, Value]]: Two dictionaries *mins* and *maxes* with the
                 extreme values observed for each motor.
         """
-        motor_names = self._get_motors_list(motors)
+        if motors is None:
+            motors = list(self.motors)
+        elif isinstance(motors, (str | int)):
+            motors = [motors]
+        elif not isinstance(motors, list):
+            raise TypeError(motors)
 
-        start_positions = self.sync_read("Present_Position", motor_names, normalize=False)
+        start_positions = self.sync_read("Present_Position", motors, normalize=False)
         mins = start_positions.copy()
         maxes = start_positions.copy()
 
         user_pressed_enter = False
         while not user_pressed_enter:
-            positions = self.sync_read("Present_Position", motor_names, normalize=False)
+            positions = self.sync_read("Present_Position", motors, normalize=False)
             mins = {motor: min(positions[motor], min_) for motor, min_ in mins.items()}
             maxes = {motor: max(positions[motor], max_) for motor, max_ in maxes.items()}
 
             if display_values:
                 print("\n-------------------------------------------")
                 print(f"{'NAME':<15} | {'MIN':>6} | {'POS':>6} | {'MAX':>6}")
-                for motor in motor_names:
+                for motor in motors:
                     print(f"{motor:<15} | {mins[motor]:>6} | {positions[motor]:>6} | {maxes[motor]:>6}")
 
             if enter_pressed():
@@ -827,9 +830,9 @@ class SerialMotorsBus(MotorsBusBase):
 
             if display_values and not user_pressed_enter:
                 # Move cursor up to overwrite the previous output
-                move_cursor_up(len(motor_names) + 3)
+                move_cursor_up(len(motors) + 3)
 
-        same_min_max = [motor for motor in motor_names if mins[motor] == maxes[motor]]
+        same_min_max = [motor for motor in motors if mins[motor] == maxes[motor]]
         if same_min_max:
             raise ValueError(f"Some motors have the same min and max values:\n{pformat(same_min_max)}")
 
@@ -952,12 +955,12 @@ class SerialMotorsBus(MotorsBusBase):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
             else:
-                return None
+                return
         if self._is_error(error):
             if raise_on_error:
                 raise RuntimeError(self.packet_handler.getRxPacketError(error))
             else:
-                return None
+                return
 
         return model_number
 
@@ -1004,13 +1007,12 @@ class SerialMotorsBus(MotorsBusBase):
         err_msg = f"Failed to read '{data_name}' on {id_=} after {num_retry + 1} tries."
         value, _, _ = self._read(addr, length, id_, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
 
-        decoded = self._decode_sign(data_name, {id_: value})
+        id_value = self._decode_sign(data_name, {id_: value})
 
         if normalize and data_name in self.normalized_data:
-            normalized = self._normalize(decoded)
+            id_value = self._normalize(id_value)
-            return normalized[id_]
 
-        return decoded[id_]
+        return id_value[id_]
 
     def _read(
         self,
@@ -1021,7 +1023,7 @@ class SerialMotorsBus(MotorsBusBase):
         num_retry: int = 0,
         raise_on_error: bool = True,
         err_msg: str = "",
-    ) -> tuple[int, int, int]:
+    ) -> tuple[int, int]:
         if length == 1:
             read_fn = self.packet_handler.read1ByteTxRx
         elif length == 2:
@@ -1071,14 +1073,13 @@ class SerialMotorsBus(MotorsBusBase):
         model = self.motors[motor].model
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
-        int_value = int(value)
         if normalize and data_name in self.normalized_data:
-            int_value = self._unnormalize({id_: value})[id_]
+            value = self._unnormalize({id_: value})[id_]
 
-        int_value = self._encode_sign(data_name, {id_: int_value})[id_]
+        value = self._encode_sign(data_name, {id_: value})[id_]
 
-        err_msg = f"Failed to write '{data_name}' on {id_=} with '{int_value}' after {num_retry + 1} tries."
+        err_msg = f"Failed to write '{data_name}' on {id_=} with '{value}' after {num_retry + 1} tries."
-        self._write(addr, length, id_, int_value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
+        self._write(addr, length, id_, value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
 
     def _write(
         self,
@@ -1112,7 +1113,7 @@ class SerialMotorsBus(MotorsBusBase):
     def sync_read(
         self,
         data_name: str,
-        motors: NameOrID | Sequence[NameOrID] | None = None,
+        motors: str | list[str] | None = None,
         *,
         normalize: bool = True,
         num_retry: int = 0,
@@ -1121,7 +1122,7 @@ class SerialMotorsBus(MotorsBusBase):
 
         Args:
             data_name (str): Register name.
-            motors (NameOrID | Sequence[NameOrID] | None, optional): Motors to query. `None` (default) reads every motor.
+            motors (str | list[str] | None, optional): Motors to query. `None` (default) reads every motor.
             normalize (bool, optional): Normalisation flag.  Defaults to `True`.
             num_retry (int, optional): Retry attempts.  Defaults to `0`.
 
@@ -1142,17 +1143,16 @@ class SerialMotorsBus(MotorsBusBase):
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
         err_msg = f"Failed to sync read '{data_name}' on {ids=} after {num_retry + 1} tries."
-        raw_ids_values, _ = self._sync_read(
+        ids_values, _ = self._sync_read(
             addr, length, ids, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
         )
 
-        decoded = self._decode_sign(data_name, raw_ids_values)
+        ids_values = self._decode_sign(data_name, ids_values)
 
         if normalize and data_name in self.normalized_data:
-            normalized = self._normalize(decoded)
+            ids_values = self._normalize(ids_values)
-            return {self._id_to_name(id_): value for id_, value in normalized.items()}
 
-        return {self._id_to_name(id_): value for id_, value in decoded.items()}
+        return {self._id_to_name(id_): value for id_, value in ids_values.items()}
 
     def _sync_read(
         self,
@@ -1224,24 +1224,21 @@ class SerialMotorsBus(MotorsBusBase):
             num_retry (int, optional): Retry attempts.  Defaults to `0`.
         """
 
-        raw_ids_values = self._get_ids_values_dict(values)
+        ids_values = self._get_ids_values_dict(values)
-        models = [self._id_to_model(id_) for id_ in raw_ids_values]
+        models = [self._id_to_model(id_) for id_ in ids_values]
         if self._has_different_ctrl_tables:
             assert_same_address(self.model_ctrl_table, models, data_name)
 
         model = next(iter(models))
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
-        int_ids_values = {id_: int(val) for id_, val in raw_ids_values.items()}
         if normalize and data_name in self.normalized_data:
-            int_ids_values = self._unnormalize(raw_ids_values)
+            ids_values = self._unnormalize(ids_values)
 
-        int_ids_values = self._encode_sign(data_name, int_ids_values)
+        ids_values = self._encode_sign(data_name, ids_values)
 
-        err_msg = f"Failed to sync write '{data_name}' with ids_values={int_ids_values} after {num_retry + 1} tries."
+        err_msg = f"Failed to sync write '{data_name}' with {ids_values=} after {num_retry + 1} tries."
-        self._sync_write(
+        self._sync_write(addr, length, ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
-            addr, length, int_ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
-        )
 
     def _sync_write(
         self,

src/lerobot/motors/robstride/__init__.py

@@ -1,18 +0,0 @@
-#!/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.
-
-from .robstride import RobstrideMotorsBus
-from .tables import *

src/lerobot/motors/robstride/tables.py

@@ -1,120 +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.
-
-"""Configuration tables for Damiao motors."""
-
-from enum import IntEnum
-
-
-# Motor type definitions
-class MotorType(IntEnum):
-    O0 = 0
-    O1 = 1
-    O2 = 2
-    O3 = 3
-    O4 = 4
-    O5 = 5
-    ELO5 = 6
-    O6 = 7
-
-
-class CommMode(IntEnum):
-    PrivateProtocole = 0
-    CANopen = 1
-    MIT = 2
-
-
-# Control modes
-class ControlMode(IntEnum):
-    MIT = 0
-    POS_VEL = 1
-    VEL = 2
-
-
-# Motor limit parameters [PMAX, VMAX, TMAX]
-# PMAX: Maximum position (rad)
-# VMAX: Maximum velocity (rad/s)
-# TMAX: Maximum torque (N·m)
-MOTOR_LIMIT_PARAMS: dict[MotorType, tuple[float, float, float]] = {
-    MotorType.O0: (12.57, 33, 14),
-    MotorType.O1: (12.57, 44, 17),
-    MotorType.O2: (12.57, 33, 20),
-    MotorType.O3: (12.57, 33, 60),
-    MotorType.O4: (12.57, 33, 120),
-    MotorType.O5: (12.57, 50, 5.5),
-    MotorType.ELO5: (12.57, 50, 6),
-    MotorType.O6: (112.5, 50, 36),
-}
-
-# Motor model names
-MODEL_NAMES = {
-    MotorType.O0: "O0",
-    MotorType.O1: "O1",
-    MotorType.O2: "O2",
-    MotorType.O3: "O3",
-    MotorType.O4: "O4",
-    MotorType.O5: "O5",
-    MotorType.ELO5: "ELO5",
-    MotorType.O6: "O6",
-}
-
-# Motor resolution table (encoder counts per revolution)
-MODEL_RESOLUTION = {
-    "O0": 65536,
-    "O1": 65536,
-    "O2": 65536,
-    "O3": 65536,
-    "O4": 65536,
-    "O5": 65536,
-    "ELO5": 65536,
-    "O6": 65536,
-}
-
-# CAN baudrates supported by Robstride motors
-AVAILABLE_BAUDRATES = [
-    1000000,  # 4: 1 mbps (default)
-]
-DEFAULT_BAUDRATE = 1000000
-
-# Default timeout in milliseconds
-DEFAULT_TIMEOUT_MS = 0  # disabled by default, otherwise 20000 is 1s
-
-
-# Data that should be normalized
-NORMALIZED_DATA = ["Present_Position", "Goal_Position"]
-
-
-# MIT control parameter ranges
-MIT_KP_RANGE = (0.0, 500.0)
-MIT_KD_RANGE = (0.0, 5.0)
-
-# CAN frame command IDs
-CAN_CMD_ENABLE = 0xFC
-CAN_CMD_DISABLE = 0xFD
-CAN_CMD_SET_ZERO = 0xFE
-CAN_CMD_CLEAR_FAULT = 0xFB
-
-
-CAN_CMD_QUERY_PARAM = 0x33
-CAN_CMD_WRITE_PARAM = 0x55
-CAN_CMD_SAVE_PARAM = 0xAA
-
-# CAN ID for parameter operations
-CAN_PARAM_ID = 0x7FF
-
-
-RUNNING_TIMEOUT = 0.001
-PARAM_TIMEOUT = 0.01
-
-STATE_CACHE_TTL_S = 0.02

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -55,16 +55,10 @@ class DiffusionConfig(PreTrainedConfig):
         normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
             a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
-        resize_shape: (H, W) shape to resize images to as a preprocessing step for the vision
+        crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
-            backbone. If None, no resizing is done and the original image resolution is used.
+            within the image size. If None, no cropping is done.
-        crop_ratio: Ratio in (0, 1] used to derive the crop size from resize_shape
+        crop_is_random: Whether the crop should be random at training time (it's always a center crop in eval
-            (crop_h = int(resize_shape[0] * crop_ratio), likewise for width).
+            mode).
-            Set to 1.0 to disable cropping. Only takes effect when resize_shape is not None.
-        crop_shape: (H, W) shape to crop images to. When resize_shape is set and crop_ratio < 1.0,
-            this is computed automatically. Can also be set directly for legacy configs that use
-            crop-only (without resize). If None and no derivation applies, no cropping is done.
-        crop_is_random: Whether the crop should be random at training time (it's always a center
-            crop in eval mode).
         pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
             `None` means no pretrained weights.
         use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
@@ -120,9 +114,7 @@ class DiffusionConfig(PreTrainedConfig):
     # Architecture / modeling.
     # Vision backbone.
     vision_backbone: str = "resnet18"
-    resize_shape: tuple[int, int] | None = None
+    crop_shape: tuple[int, int] | None = (84, 84)
-    crop_ratio: float = 1.0
-    crop_shape: tuple[int, int] | None = None
     crop_is_random: bool = True
     pretrained_backbone_weights: str | None = None
     use_group_norm: bool = True
@@ -147,10 +139,6 @@ class DiffusionConfig(PreTrainedConfig):
     # Inference
     num_inference_steps: int | None = None
 
-    # Optimization
-    compile_model: bool = False
-    compile_mode: str = "reduce-overhead"
-
     # Loss computation
     do_mask_loss_for_padding: bool = False
 
@@ -183,25 +171,6 @@ class DiffusionConfig(PreTrainedConfig):
                 f"Got {self.noise_scheduler_type}."
             )
 
-        if self.resize_shape is not None and (
-            len(self.resize_shape) != 2 or any(d <= 0 for d in self.resize_shape)
-        ):
-            raise ValueError(f"`resize_shape` must be a pair of positive integers. Got {self.resize_shape}.")
-        if not (0 < self.crop_ratio <= 1.0):
-            raise ValueError(f"`crop_ratio` must be in (0, 1]. Got {self.crop_ratio}.")
-
-        if self.resize_shape is not None:
-            if self.crop_ratio < 1.0:
-                self.crop_shape = (
-                    int(self.resize_shape[0] * self.crop_ratio),
-                    int(self.resize_shape[1] * self.crop_ratio),
-                )
-            else:
-                # Explicitly disable cropping for resize+ratio path when crop_ratio == 1.0.
-                self.crop_shape = None
-        if self.crop_shape is not None and (self.crop_shape[0] <= 0 or self.crop_shape[1] <= 0):
-            raise ValueError(f"`crop_shape` must have positive dimensions. Got {self.crop_shape}.")
-
         # Check that the horizon size and U-Net downsampling is compatible.
         # U-Net downsamples by 2 with each stage.
         downsampling_factor = 2 ** len(self.down_dims)
@@ -229,12 +198,13 @@ class DiffusionConfig(PreTrainedConfig):
         if len(self.image_features) == 0 and self.env_state_feature is None:
             raise ValueError("You must provide at least one image or the environment state among the inputs.")
 
-        if self.resize_shape is None and self.crop_shape is not None:
+        if self.crop_shape is not None:
             for key, image_ft in self.image_features.items():
                 if self.crop_shape[0] > image_ft.shape[1] or self.crop_shape[1] > image_ft.shape[2]:
                     raise ValueError(
-                        f"`crop_shape` should fit within the image shapes. Got {self.crop_shape} "
+                        f"`crop_shape` should fit within the images shapes. Got {self.crop_shape} "
-                        f"for `crop_shape` and {image_ft.shape} for `{key}`."
+                        f"for `crop_shape` and {image_ft.shape} for "
+                        f"`{key}`."
                     )
 
         # Check that all input images have the same shape.

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -142,9 +142,6 @@ class DiffusionPolicy(PreTrainedPolicy):
         """Run the batch through the model and compute the loss for training or validation."""
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            for key in self.config.image_features:
-                if self.config.n_obs_steps == 1 and batch[key].ndim == 4:
-                    batch[key] = batch[key].unsqueeze(1)
             batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
         loss = self.diffusion.compute_loss(batch)
         # no output_dict so returning None
@@ -185,11 +182,6 @@ class DiffusionModel(nn.Module):
 
         self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)
 
-        if config.compile_model:
-            # Compile the U-Net. "reduce-overhead" is preferred for the small-batch repetitive loops
-            # common in diffusion inference.
-            self.unet = torch.compile(self.unet, mode=config.compile_mode)
-
         self.noise_scheduler = _make_noise_scheduler(
             config.noise_scheduler_type,
             num_train_timesteps=config.num_train_timesteps,
@@ -454,18 +446,12 @@ class DiffusionRgbEncoder(nn.Module):
     def __init__(self, config: DiffusionConfig):
         super().__init__()
         # Set up optional preprocessing.
-        if config.resize_shape is not None:
+        if config.crop_shape is not None:
-            self.resize = torchvision.transforms.Resize(config.resize_shape)
-        else:
-            self.resize = None
-
-        crop_shape = config.crop_shape
-        if crop_shape is not None:
             self.do_crop = True
             # Always use center crop for eval
-            self.center_crop = torchvision.transforms.CenterCrop(crop_shape)
+            self.center_crop = torchvision.transforms.CenterCrop(config.crop_shape)
             if config.crop_is_random:
-                self.maybe_random_crop = torchvision.transforms.RandomCrop(crop_shape)
+                self.maybe_random_crop = torchvision.transforms.RandomCrop(config.crop_shape)
             else:
                 self.maybe_random_crop = self.center_crop
         else:
@@ -491,16 +477,13 @@ class DiffusionRgbEncoder(nn.Module):
 
         # Set up pooling and final layers.
         # Use a dry run to get the feature map shape.
-        # The dummy shape mirrors the runtime preprocessing order: resize -> crop.
+        # The dummy input should take the number of image channels from `config.image_features` and it should
+        # use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
+        # height and width from `config.image_features`.
 
         # Note: we have a check in the config class to make sure all images have the same shape.
         images_shape = next(iter(config.image_features.values())).shape
-        if config.crop_shape is not None:
+        dummy_shape_h_w = config.crop_shape if config.crop_shape is not None else images_shape[1:]
-            dummy_shape_h_w = config.crop_shape
-        elif config.resize_shape is not None:
-            dummy_shape_h_w = config.resize_shape
-        else:
-            dummy_shape_h_w = images_shape[1:]
         dummy_shape = (1, images_shape[0], *dummy_shape_h_w)
         feature_map_shape = get_output_shape(self.backbone, dummy_shape)[1:]
 
@@ -516,10 +499,7 @@ class DiffusionRgbEncoder(nn.Module):
         Returns:
             (B, D) image feature.
         """
-        # Preprocess: resize if configured, then crop if configured.
+        # Preprocess: maybe crop (if it was set up in the __init__).
-
-        if self.resize is not None:
-            x = self.resize(x)
         if self.do_crop:
             if self.training:  # noqa: SIM108
                 x = self.maybe_random_crop(x)

src/lerobot/policies/factory.py

@@ -34,6 +34,7 @@ from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.policies.groot.configuration_groot import GrootConfig
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.policies.pi05.configuration_pi05 import PI05Config
+from lerobot.policies.pi05_full.configuration_pi05 import PI05FullConfig
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.sac.configuration_sac import SACConfig
 from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
@@ -390,6 +391,13 @@ def make_pre_post_processors(
             config=policy_cfg,
             dataset_stats=kwargs.get("dataset_stats"),
         )
+    elif isinstance(policy_cfg, PI05FullConfig):
+        from lerobot.policies.pi05_full.processor_pi05 import make_pi05_full_pre_post_processors
+
+        processors = make_pi05_full_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
 
     else:
         try:

src/lerobot/policies/pi05_full/README.md

@@ -0,0 +1,49 @@
+# π₀.₅ (pi05)
+
+This repository contains the Hugging Face port of **π₀.₅**, adapted from [OpenPI](https://github.com/Physical-Intelligence/openpi) by the Physical Intelligence.
+It is designed as a **Vision-Language-Action model with open-world generalization**.
+
+---
+
+## Model Overview
+
+| Feature              | π₀                                                     | π₀.₅                                      |
+| -------------------- | ------------------------------------------------------ | ----------------------------------------- |
+| Time Conditioning    | Concatenates time with actions via `action_time_mlp_*` | Uses `time_mlp_*` for AdaRMS conditioning |
+| AdaRMS               | Not used                                               | Used in action expert                     |
+| Tokenizer Length     | 48 tokens                                              | 200 tokens                                |
+| Discrete State Input | False (Uses `state_proj` layer)                        | True                                      |
+| Parameter Count      | Higher (includes state embedding)                      | Lower (no state embedding)                |
+
+---
+
+## Citation
+
+If you use this work, please cite both **OpenPI** and the π₀.₅ paper:
+
+```bibtex
+@misc{openpi2024,
+  author       = {Physical Intelligence Lab},
+  title        = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
+  year         = {2024},
+  publisher    = {GitHub},
+  howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
+  license      = {Apache-2.0}
+}
+
+@misc{intelligence2025pi05visionlanguageactionmodelopenworld,
+  title        = {π₀.₅: a Vision-Language-Action Model with Open-World Generalization},
+  author       = {Physical Intelligence and Kevin Black and Noah Brown and James Darpinian and Karan Dhabalia and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Manuel Y. Galliker and Dibya Ghosh and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Devin LeBlanc and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Allen Z. Ren and Lucy Xiaoyang Shi and Laura Smith and Jost Tobias Springenberg and Kyle Stachowicz and James Tanner and Quan Vuong and Homer Walke and Anna Walling and Haohuan Wang and Lili Yu and Ury Zhilinsky},
+  year         = {2025},
+  eprint       = {2504.16054},
+  archivePrefix= {arXiv},
+  primaryClass = {cs.LG},
+  url          = {https://arxiv.org/abs/2504.16054},
+}
+```
+
+---
+
+## License
+
+This port follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).

src/lerobot/policies/pi05_full/__init__.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python
 
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -14,7 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .config_openarm_mini import OpenArmMiniConfig
+from .configuration_pi05 import PI05FullConfig
-from .openarm_mini import OpenArmMini
+from .modeling_pi05 import PI05FullPolicy
+from .processor_pi05 import make_pi05_full_pre_post_processors
 
-__all__ = ["OpenArmMini", "OpenArmMiniConfig"]
+__all__ = ["PI05FullConfig", "PI05FullPolicy", "make_pi05_full_pre_post_processors"]

src/lerobot/policies/pi05_full/annotate/annotate_libero.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+
+# Example script to run synthetic data generation with Qwen VLM
+# This generates user prompts and robot utterances for hierarchical policy training
+
+# Configuration
+REPO_ID="lerobot/libero_10"
+MODEL="Qwen/Qwen3-VL-30B-A3B-Instruct"
+# or: MODEL="Qwen/Qwen2-VL-7B-Instruct"
+
+
+OUTPUT_DIR="/fsx/jade_choghari/outputs/libero-10-annotate-high"
+
+BATCH_SIZE=16
+TEMPERATURE=0.9
+SAMPLE_INTERVAL=5.0  # generate dialogue every 1 second (all episodes processed)
+
+# Run subtask annotation
+# python /admin/home/jade_choghari/lerobot/src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+#     --repo-id "$REPO_ID" \
+#     --video-key observation.images.image \
+#     --output-dir "$OUTPUT_DIR" \
+#     --skip-existing \
+#     --output-repo-id "jadechoghari/libero10-annotate" \
+#     --batch-size "$BATCH_SIZE" \
+# run synthetic data generation (all episodes processed)
+# python examples/dataset/annotate_pgen.py \
+#     --repo-id "$REPO_ID" \
+#     --model "$MODEL" \
+#     --output-dir "$OUTPUT_DIR" \
+#     --temperature "$TEMPERATURE" \
+#     --batch-size "$BATCH_SIZE" \
+#     --sample-interval "$SAMPLE_INTERVAL" \
+#     --image-key observation.images.base \
+#     --num-image-views-per-sample 1
+
+# for faster testing, increase sample interval:
+# --sample-interval 5.0  # Samples every 5 seconds (much faster)
+
+# to push to hub after generation:
+# add --push-to-hub flag
+
+# efficient batch processing: 4 episodes at once
+python /admin/home/jade_choghari/lerobot/src/lerobot/policies/pi05_full/annotate/high_level_annotate.py \
+    --data-dir "/fsx/jade_choghari/outputs/libero-10-annotate" \
+    --output-dir "$OUTPUT_DIR" \
+    --video-mode \
+    --video-key observation.images.image \
+    --video-batch-size "$BATCH_SIZE" \
+    --sample-interval 5.0

src/lerobot/policies/pi05_full/annotate/load_lerobot_high.py

@@ -0,0 +1,52 @@
+import torch
+from huggingface_hub import HfApi
+
+import lerobot
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.configs.policies import PreTrainedConfig
+
+# /fsx/jade_choghari/data/libero_10_subtasks_kw_converted
+dataset = LeRobotDataset(repo_id="lerobot/libero_10_image_subtask")
+
+dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=0,
+        batch_size=2,
+        shuffle=True,
+)
+
+cfg = PreTrainedConfig.from_pretrained(
+    pretrained_name_or_path="/fsx/jade_choghari/models/pi05-base",
+)
+cfg.dtype = "bfloat16"
+
+pre_processor, post_processor = make_pre_post_processors(
+    policy_cfg=cfg,
+    pretrained_path="/fsx/jade_choghari/models/pi05-base",
+)
+batch = next(iter(dataloader))
+breakpoint()
+batch1 = pre_processor(batch)
+breakpoint()
+print(batch.keys())
+# print(batch['task_index_high_level'].shape)
+# print(batch['task_index_high_level'])
+# print(batch['user_prompt'][0])
+# print(batch['robot_utterance'][0])
+# print(batch['task'][0])
+
+valid_episode_list = []
+for episode_idx in range(len(dataset.meta.episodes)):
+        subtask_index = dataset[episode_idx]["subtask_index"]
+        valid_episode_list.append(episode_idx)
+
+print(len(valid_episode_list))
+
+# read this parquet /fsx/jade_choghari/outputs/pgen_annotations1/meta/tasks.parquett
+# import pandas as pd
+# tasks_df = pd.read_parquet('/fsx/jade_choghari/outputs/pgen_annotations1/meta/tasks.parquet')
+
+# # print all
+# print(tasks_df.columns)
+# breakpoint()

src/lerobot/policies/pi05_full/annotate/run_pgen.sh

@@ -0,0 +1,49 @@
+#!/bin/bash
+
+# Example script to run synthetic data generation with Qwen VLM
+# This generates user prompts and robot utterances for hierarchical policy training
+
+# Configuration
+REPO_ID="jadechoghari/collect-data"
+MODEL="Qwen/Qwen3-VL-30B-A3B-Instruct"
+# or: MODEL="Qwen/Qwen2-VL-7B-Instruct"
+
+
+OUTPUT_DIR="/fsx/jade_choghari/outputs/collect-data-pgen_new"
+
+BATCH_SIZE=32
+TEMPERATURE=0.9
+SAMPLE_INTERVAL=5.0  # generate dialogue every 1 second (all episodes processed)
+
+# Run subtask annotation
+python /admin/home/jade_choghari/lerobot/src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+    --repo-id "$REPO_ID" \
+    --video-key observation.images.base \
+    --output-dir "$OUTPUT_DIR" \
+    --output-repo-id "jadechoghari/collect-data-with-subtasks"
+# run synthetic data generation (all episodes processed)
+# python examples/dataset/annotate_pgen.py \
+#     --repo-id "$REPO_ID" \
+#     --model "$MODEL" \
+#     --output-dir "$OUTPUT_DIR" \
+#     --temperature "$TEMPERATURE" \
+#     --batch-size "$BATCH_SIZE" \
+#     --sample-interval "$SAMPLE_INTERVAL" \
+#     --image-key observation.images.base \
+#     --num-image-views-per-sample 1
+
+# for faster testing, increase sample interval:
+# --sample-interval 5.0  # Samples every 5 seconds (much faster)
+
+# to push to hub after generation:
+# add --push-to-hub flag
+
+# efficient batch processing: 4 episodes at once
+# python examples/dataset/annotate_pgen.py \
+#     --repo-id "$REPO_ID" \
+#     --model "$MODEL" \
+#     --output-dir "$OUTPUT_DIR" \
+#     --video-mode \
+#     --video-key observation.images.up \
+#     --video-batch-size "$BATCH_SIZE" \
+#     --sample-interval 1.0

src/lerobot/policies/pi05_full/configuration_pi05.py

@@ -0,0 +1,183 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.optim.optimizers import AdamWConfig
+from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
+
+DEFAULT_IMAGE_SIZE = 224
+
+
+@PreTrainedConfig.register_subclass("pi05_full")
+@dataclass
+class PI05FullConfig(PreTrainedConfig):
+    paligemma_variant: str = "gemma_2b"
+    action_expert_variant: str = "gemma_300m"
+    dtype: str = "float32"  # Options: "bfloat16", "float32"
+
+    n_obs_steps: int = 1
+    chunk_size: int = 50  # Number of action steps to predict, in openpi called "action_horizon"
+    n_action_steps: int = 50  # Number of action steps to execute
+
+    # Shorter state and action vectors will be padded to these dimensions
+    max_state_dim: int = 32
+    max_action_dim: int = 32
+
+    # Flow matching parameters: see openpi `PI0Pytorch`
+    num_inference_steps: int = 10
+    time_sampling_beta_alpha: float = 1.5
+    time_sampling_beta_beta: float = 1.0
+    time_sampling_scale: float = 0.999
+    time_sampling_offset: float = 0.001
+    min_period: float = 4e-3
+    max_period: float = 4.0
+
+    # Real-Time Chunking (RTC) configuration
+    rtc_config: RTCConfig | None = None
+
+    image_resolution: tuple[int, int] = (
+        DEFAULT_IMAGE_SIZE,
+        DEFAULT_IMAGE_SIZE,
+    )  # see openpi `preprocessing_pytorch.py`
+
+    # Add empty images. Used to add empty cameras when no image features are present.
+    empty_cameras: int = 0
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.MEAN_STD,  # Pi0.5 uses quantiles for state
+            "ACTION": NormalizationMode.MEAN_STD,  # Pi0.5 uses quantiles for action
+        }
+    )
+
+    action_tokenizer_name: str = "physical-intelligence/fast"
+    text_tokenizer_name: str = "google/paligemma-3b-pt-224"
+    max_action_tokens: int = 256
+    fast_skip_tokens: int = 128
+
+    # subtask stuff
+    max_decoding_steps: int = 200
+    temperature: float = 0.0
+    subtask_regeneration_interval: float = 1.0  # Regenerate subtask tokens every N seconds (0 = every call)
+
+    # Training settings
+    gradient_checkpointing: bool = False  # Enable gradient checkpointing for memory optimization
+    compile_model: bool = False  # Whether to use torch.compile for model optimization
+    compile_mode: str = "max-autotune"  # Torch compile mode
+    device: str | None = None  # Device to use for the model (None = auto-detect)
+
+    # Finetuning settings
+    freeze_vision_encoder: bool = False  # Freeze only the vision encoder
+    train_expert_only: bool = False  # Freeze entire VLM, train only action expert and projections
+    knowledge_insulation: bool = True  # Enable knowledge insulation in attention (blocks gradients from action to VLM K/V)
+
+    # Loss weights (used when knowledge_insulation is enabled)
+    loss_weight_flow: float = 1.0  # Weight for flow matching MSE loss (continuous actions)
+    loss_weight_action_ce: float = 1.0  # Weight for FAST action token cross-entropy loss
+    loss_weight_subtask_ce: float = 1.0  # Weight for subtask token cross-entropy loss
+
+    # Optimizer settings: see openpi `AdamW`
+    optimizer_lr: float = 2.5e-5  # see openpi `CosineDecaySchedule: peak_lr`
+    optimizer_betas: tuple[float, float] = (0.9, 0.95)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 0.01
+    optimizer_grad_clip_norm: float = 1.0
+
+    # Scheduler settings: see openpi `CosineDecaySchedule`
+    # Note: These will auto-scale if --steps < scheduler_decay_steps
+    # For example, --steps=3000 will scale warmup to 100 and decay to 3000
+    scheduler_warmup_steps: int = 1_000
+    scheduler_decay_steps: int = 30_000
+    scheduler_decay_lr: float = 2.5e-6
+
+    tokenizer_max_length: int = 48  # see openpi `__post_init__`
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        # Validate configuration
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"n_action_steps ({self.n_action_steps}) cannot be greater than chunk_size ({self.chunk_size})"
+            )
+
+        if self.paligemma_variant not in ["gemma_300m", "gemma_2b"]:
+            raise ValueError(f"Invalid paligemma_variant: {self.paligemma_variant}")
+
+        if self.action_expert_variant not in ["gemma_300m", "gemma_2b"]:
+            raise ValueError(f"Invalid action_expert_variant: {self.action_expert_variant}")
+
+        if self.dtype not in ["bfloat16", "float32"]:
+            raise ValueError(f"Invalid dtype: {self.dtype}")
+
+    def validate_features(self) -> None:
+        """Validate and set up input/output features."""
+        for i in range(self.empty_cameras):
+            key = OBS_IMAGES + f".empty_camera_{i}"
+            empty_camera = PolicyFeature(
+                type=FeatureType.VISUAL,
+                shape=(3, *self.image_resolution),  # Use configured image resolution
+            )
+            self.input_features[key] = empty_camera
+
+        if OBS_STATE not in self.input_features:
+            state_feature = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(self.max_state_dim,),  # Padded to max_state_dim
+            )
+            self.input_features[OBS_STATE] = state_feature
+
+        if ACTION not in self.output_features:
+            action_feature = PolicyFeature(
+                type=FeatureType.ACTION,
+                shape=(self.max_action_dim,),  # Padded to max_action_dim
+            )
+            self.output_features[ACTION] = action_feature
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+            grad_clip_norm=self.optimizer_grad_clip_norm,
+        )
+
+    def get_scheduler_preset(self):
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.scheduler_decay_lr,
+            num_warmup_steps=self.scheduler_warmup_steps,
+            num_decay_steps=self.scheduler_decay_steps,
+        )
+
+    @property
+    def observation_delta_indices(self) -> None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

src/lerobot/policies/pi05_full/inference.py

@@ -0,0 +1,92 @@
+import torch
+from huggingface_hub import HfApi
+
+import lerobot
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+# import make_pre_post_processors
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies.pi05.configuration_pi05 import PI05Config
+from lerobot.policies.factory import make_policy, make_policy_config
+from lerobot.configs.policies import PreTrainedConfig
+
+cfg = PreTrainedConfig.from_pretrained(
+    pretrained_name_or_path="/fsx/jade_choghari/models/pi05-base",
+)
+cfg.dtype = "bfloat16"
+
+pre_processor, post_processor = make_pre_post_processors(
+    policy_cfg=cfg,
+    pretrained_path="/fsx/jade_choghari/models/pi05-base",
+)
+
+delta_timestamps = {'action': [0.0, 0.03333333333333333, 0.06666666666666667, 0.1, 0.13333333333333333, 0.16666666666666666, 0.2, 0.23333333333333334, 0.26666666666666666, 0.3, 0.3333333333333333, 0.36666666666666664, 0.4, 0.43333333333333335, 0.4666666666666667, 0.5, 0.5333333333333333, 0.5666666666666667, 0.6, 0.6333333333333333, 0.6666666666666666, 0.7, 0.7333333333333333, 0.7666666666666667, 0.8, 0.8333333333333334, 0.8666666666666667, 0.9, 0.9333333333333333, 0.9666666666666667, 1.0, 1.0333333333333334, 1.0666666666666667, 1.1, 1.1333333333333333, 1.1666666666666667, 1.2, 1.2333333333333334, 1.2666666666666666, 1.3, 1.3333333333333333, 1.3666666666666667, 1.4, 1.4333333333333333, 1.4666666666666666, 1.5, 1.5333333333333334, 1.5666666666666667, 1.6, 1.6333333333333333]}
+
+dataset = LeRobotDataset(repo_id="local", root="/fsx/jade_choghari/outputs/pgen_annotations1", delta_timestamps=delta_timestamps)
+
+# rename map --rename_map='{
+#         "observation.images.side": "observation.images.base_0_rgb",
+#         "observation.images.up": "observation.images.left_wrist_0_rgb"
+#         }'
+rename_map = {
+    "observation.images.side": "observation.images.base_0_rgb",
+    "observation.images.up": "observation.images.left_wrist_0_rgb"
+}
+policy = make_policy(
+    cfg=cfg,
+    ds_meta=dataset.meta,
+    rename_map=rename_map,
+)
+
+dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=0,
+        batch_size=4,
+        shuffle=True,
+)
+
+batch = next(iter(dataloader))
+breakpoint()
+batch = pre_processor(batch)
+policy.train()
+# run inference
+# action = policy.select_action(batch)
+loss, loss_dict = policy.forward(batch)
+breakpoint()
+# import requests
+# from PIL import Image
+# from transformers import AutoProcessor
+# model = policy.model.paligemma_with_expert.paligemma
+# model = model.to(device="cuda", dtype=torch.bfloat16)
+# model.eval()
+# prompt = "Describe this image."
+# url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
+# image = Image.open(requests.get(url, stream=True).raw)
+# processor = AutoProcessor.from_pretrained(
+#     "google/paligemma-3b-pt-224",
+# )
+# inputs = processor(image, prompt, return_tensors="pt").to(model.device)
+# print("generating...")
+# output = model.generate(
+#     **inputs,
+#     max_new_tokens=50,
+#     use_cache=True,  # default dynamic cache
+# )
+# print(processor.decode(output[0], skip_special_tokens=True))
+
+
+# # other model
+# from transformers import PaliGemmaForConditionalGeneration
+# model = PaliGemmaForConditionalGeneration.from_pretrained(
+#     "google/paligemma2-3b-pt-224",
+#     torch_dtype=torch.bfloat16,
+#     device_map="auto",
+# )
+# model.eval()
+# print("generating...")
+# output = model.generate(
+#     **inputs,
+#     max_new_tokens=100,
+#     use_cache=True,  # default dynamic cache
+# )
+# print("Model 2 output:")
+# print(processor.decode(output[0], skip_special_tokens=True))

src/lerobot/policies/pi05_full/processor_pi05.py

@@ -0,0 +1,194 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from copy import deepcopy
+from dataclasses import dataclass
+from typing import Any
+
+import numpy as np
+import torch
+
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.policies.pi05_full.configuration_pi05 import PI05FullConfig
+from lerobot.policies.pi05_full.modeling_pi05 import pad_vector
+from lerobot.processor import (
+    ActionTokenizerProcessorStep,
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    RenameObservationsProcessorStep,
+    TokenizerProcessorStep,
+    UnnormalizerProcessorStep,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+from lerobot.processor.core import EnvTransition, TransitionKey
+from lerobot.utils.constants import (
+    OBS_STATE,
+    POLICY_POSTPROCESSOR_DEFAULT_NAME,
+    POLICY_PREPROCESSOR_DEFAULT_NAME,
+)
+
+
+@ProcessorStepRegistry.register(name="pi05_full_prepare_state_tokenizer_processor_step")
+@dataclass
+class Pi05FullPrepareStateTokenizerProcessorStep(ProcessorStep):
+    """
+    Processor step to prepare the state and tokenize the language input.
+    """
+
+    max_state_dim: int = 32
+    task_key: str = "task"
+    subtask_key: str = "subtask"
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        transition = transition.copy()
+
+        state = transition.get(TransitionKey.OBSERVATION, {}).get(OBS_STATE)
+        if state is None:
+            raise ValueError("State is required for PI05")
+        user_prompts = transition.get(TransitionKey.COMPLEMENTARY_DATA, {}).get(self.task_key)
+        if user_prompts is None:
+            raise ValueError("No user prompts found in complementary data")
+        commands = transition.get(TransitionKey.COMPLEMENTARY_DATA, {}).get(self.subtask_key)
+
+        # TODO: check if this necessary
+        state = deepcopy(state)
+
+        # Prepare state (pad to max_state_dim)
+        state = pad_vector(state, self.max_state_dim)
+
+        # State should already be normalized to [-1, 1] by the NormalizerProcessorStep that runs before this step
+        # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
+        state_np = state.cpu().numpy()
+        discretized_states = np.digitize(state_np, bins=np.linspace(-1, 1, 256 + 1)[:-1]) - 1
+
+        full_prompts = []
+        for i, user_prompt in enumerate(user_prompts):
+            cleaned_text = user_prompt.strip().replace("_", " ").replace("\n", " ")
+            cleaned_text = cleaned_text.lower()   # all lowercase # NOTE: added by (jadechoghari)
+            state_str = " ".join(map(str, discretized_states[i]))
+            full_prompt = f"Task: {cleaned_text}, State: {state_str};\n"
+            full_prompts.append(full_prompt)
+        
+        transition[TransitionKey.COMPLEMENTARY_DATA][self.task_key] = full_prompts
+
+        # process commands (optional)
+        if commands is not None:
+            full_commands = []
+            for i, command in enumerate(commands):
+                cleaned_text = command.strip().replace("_", " ").replace("\n", " ")
+                cleaned_text = cleaned_text.lower()   # all lowercase # NOTE: added by (jadechoghari)
+                full_command = f"Subtask: {cleaned_text};\n"
+                full_commands.append(full_command)
+
+            transition[TransitionKey.COMPLEMENTARY_DATA][self.subtask_key] = full_commands
+
+        # note: action tokens will be processed in the ActionTokenizerProcessorStep
+        # Normalize state to [-1, 1] range if needed (assuming it's already normalized by normalizer processor step!!)
+        # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
+        return transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        This step does not alter the feature definitions.
+        """
+        return features
+
+
+def make_pi05_full_pre_post_processors(
+    config: PI05FullConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Constructs pre-processor and post-processor pipelines for the PI0 policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1. Renaming features to match pretrained configurations.
+    2. Normalizing input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Appending a newline character to the task description for tokenizer compatibility.
+    5. Tokenizing the text prompt using the PaliGemma tokenizer.
+    6. Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the PI0 policy.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+
+    # Add remaining processors
+    input_steps: list[ProcessorStep] = [
+        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
+        AddBatchDimensionProcessorStep(),
+        # NOTE: NormalizerProcessorStep MUST come before Pi05PrepareStateTokenizerProcessorStep
+        # because the tokenizer step expects normalized state in [-1, 1] range for discretization
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        Pi05FullPrepareStateTokenizerProcessorStep(max_state_dim=config.max_state_dim),
+        TokenizerProcessorStep(
+            tokenizer_name=config.text_tokenizer_name,
+            max_length=config.tokenizer_max_length,
+            padding_side="right",
+            padding="max_length",
+        ),
+        ActionTokenizerProcessorStep(
+            action_tokenizer_name=config.action_tokenizer_name,
+            max_action_tokens=config.max_action_tokens,
+            fast_skip_tokens=config.fast_skip_tokens,
+            paligemma_tokenizer_name=config.text_tokenizer_name,
+        ),
+        DeviceProcessorStep(device=config.device),
+    ]
+
+    output_steps: list[ProcessorStep] = [
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )

src/lerobot/policies/sarm/compute_rabc_weights.py

@@ -27,18 +27,18 @@ Usage:
     # Full RA-BC computation with visualizations
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4
+        --reward-model-path pepijn223/sarm_single_uni4
 
     # Faster computation with stride (compute every 5 frames, interpolate the rest)
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4 \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
         --stride 5
 
     # Visualize predictions only (no RA-BC computation)
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4 \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
         --visualize-only \\
         --num-visualizations 5
 
@@ -714,12 +714,12 @@ Examples:
     # Full RA-BC computation with visualizations
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4
+        --reward-model-path pepijn223/sarm_single_uni4
 
     # Visualize predictions only (no RA-BC computation)
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4 \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
         --visualize-only \\
         --num-visualizations 10
         """,

src/lerobot/policies/sarm/processor_sarm.py

@@ -277,7 +277,9 @@ class SARMEncodingProcessorStep(ProcessorStep):
 
         # When language is perturbed, targets are zero so perturbed samples don't contribute to progress loss
         if self.dataset_meta is not None:
-            episodes_df = self.dataset_meta.episodes.to_pandas()
+            episodes_df = None
+            if self.sparse_subtask_names != ["task"]:
+                episodes_df = self.dataset_meta.episodes.to_pandas()
 
             # Generate sparse targets
             if self.sparse_temporal_proportions is not None:

src/lerobot/policies/smolvla/configuration_smolvla.py

@@ -85,7 +85,7 @@ class SmolVLAConfig(PreTrainedConfig):
     scheduler_decay_lr: float = 2.5e-6
 
     vlm_model_name: str = "HuggingFaceTB/SmolVLM2-500M-Video-Instruct"  # Select the VLM backbone.
-    load_vlm_weights: bool = False  # Set to False in case of training the expert from scratch. True when init from pretrained SmolVLA weights
+    load_vlm_weights: bool = False  # Set to True in case of training the expert from scratch. True when init from pretrained SmolVLA weights
 
     add_image_special_tokens: bool = False  # Whether to use special image tokens around image features.
 
@@ -106,9 +106,6 @@ class SmolVLAConfig(PreTrainedConfig):
     # Real-Time Chunking (RTC) configuration
     rtc_config: RTCConfig | None = None
 
-    compile_model: bool = False  # Whether to use torch.compile for model optimization
-    compile_mode: str = "max-autotune"  # Torch compile mode
-
     def __post_init__(self):
         super().__post_init__()
 

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -30,7 +30,7 @@ Example of finetuning the smolvla pretrained model (`smolvla_base`):
 ```bash
 lerobot-train \
 --policy.path=lerobot/smolvla_base \
---dataset.repo_id=<USER>/svla_so100_task1_v3 \
+--dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```
@@ -40,7 +40,7 @@ and an action expert.
 ```bash
 lerobot-train \
 --policy.type=smolvla \
---dataset.repo_id=<USER>/svla_so100_task1_v3 \
+--dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```
@@ -378,16 +378,16 @@ class SmolVLAPolicy(PreTrainedPolicy):
         actions_is_pad = batch.get("actions_id_pad")
         loss_dict = {}
         losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions, noise, time)
-        loss_dict["losses_after_forward"] = losses.clone().mean().item()
+        loss_dict["losses_after_forward"] = losses.clone()
 
         if actions_is_pad is not None:
             in_episode_bound = ~actions_is_pad
             losses = losses * in_episode_bound.unsqueeze(-1)
-            loss_dict["losses_after_in_ep_bound"] = losses.clone().mean().item()
+            loss_dict["losses_after_in_ep_bound"] = losses.clone()
 
         # Remove padding
         losses = losses[:, :, : self.config.max_action_dim]
-        loss_dict["losses_after_rm_padding"] = losses.clone().mean().item()
+        loss_dict["losses_after_rm_padding"] = losses.clone()
 
         if reduction == "none":
             # Return per-sample losses (B,) by averaging over time and action dims
@@ -593,12 +593,6 @@ class VLAFlowMatching(nn.Module):
         self.prefix_length = self.config.prefix_length
         self.rtc_processor = rtc_processor
 
-        # Compile model if requested
-        if config.compile_model:
-            torch.set_float32_matmul_precision("high")
-            self.sample_actions = torch.compile(self.sample_actions, mode=config.compile_mode)
-            self.forward = torch.compile(self.forward, mode=config.compile_mode)
-
     def _rtc_enabled(self):
         return self.config.rtc_config is not None and self.config.rtc_config.enabled
 

src/lerobot/policies/smolvla/smolvlm_with_expert.py

@@ -77,6 +77,7 @@ class SmolVLMWithExpertModel(nn.Module):
             print(f"Loading  {model_id} weights ...")
             self.vlm = AutoModelForImageTextToText.from_pretrained(
                 model_id,
+                device_map=device,
                 torch_dtype="bfloat16",
                 low_cpu_mem_usage=True,
             )

src/lerobot/processor/__init__.py

@@ -44,7 +44,6 @@ from .hil_processor import (
     AddTeleopActionAsComplimentaryDataStep,
     AddTeleopEventsAsInfoStep,
     GripperPenaltyProcessorStep,
-    GymHILAdapterProcessorStep,
     ImageCropResizeProcessorStep,
     InterventionActionProcessorStep,
     RewardClassifierProcessorStep,
@@ -88,7 +87,6 @@ __all__ = [
     "DoneProcessorStep",
     "EnvAction",
     "EnvTransition",
-    "GymHILAdapterProcessorStep",
     "GripperPenaltyProcessorStep",
     "hotswap_stats",
     "IdentityProcessorStep",

src/lerobot/processor/converters.py

@@ -171,9 +171,11 @@ def _extract_complementary_data(batch: dict[str, Any]) -> dict[str, Any]:
     subtask_key = {"subtask": batch["subtask"]} if "subtask" in batch else {}
     index_key = {"index": batch["index"]} if "index" in batch else {}
     task_index_key = {"task_index": batch["task_index"]} if "task_index" in batch else {}
+    user_prompt_key = {"user_prompt": batch["user_prompt"]} if "user_prompt" in batch else {}
+    subtask_key = {"subtask": batch["subtask"]} if "subtask" in batch else {}
     episode_index_key = {"episode_index": batch["episode_index"]} if "episode_index" in batch else {}
 
-    return {**pad_keys, **task_key, **subtask_key, **index_key, **task_index_key, **episode_index_key}
+    return {**pad_keys, **task_key, **index_key, **task_index_key, **episode_index_key, **user_prompt_key, **subtask_key}
 
 
 def create_transition(

src/lerobot/processor/env_processor.py

@@ -17,7 +17,7 @@ from dataclasses import dataclass
 
 import torch
 
-from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.utils.constants import OBS_IMAGES, OBS_PREFIX, OBS_STATE, OBS_STR
 
 from .pipeline import ObservationProcessorStep, ProcessorStepRegistry
@@ -92,7 +92,7 @@ class LiberoProcessorStep(ObservationProcessorStep):
 
         # copy over non-STATE features
         for ft, feats in features.items():
-            if ft != FeatureType.STATE:
+            if ft != PipelineFeatureType.STATE:
                 new_features[ft] = feats.copy()
 
         # rebuild STATE features
@@ -100,11 +100,13 @@ class LiberoProcessorStep(ObservationProcessorStep):
 
         # add our new flattened state
         state_feats[OBS_STATE] = PolicyFeature(
-            type=FeatureType.STATE,
+            key=OBS_STATE,
             shape=(8,),  # [eef_pos(3), axis_angle(3), gripper(2)]
+            dtype="float32",
+            description=("Concatenated end-effector position (3), axis-angle (3), and gripper qpos (2)."),
         )
 
-        new_features[FeatureType.STATE] = state_feats
+        new_features[PipelineFeatureType.STATE] = state_feats
 
         return new_features
 

src/lerobot/processor/gym_action_processor.py

@@ -20,7 +20,6 @@ from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 
 from .converters import to_tensor
 from .core import EnvAction, EnvTransition, PolicyAction
-from .hil_processor import TELEOP_ACTION_KEY
 from .pipeline import ActionProcessorStep, ProcessorStep, ProcessorStepRegistry
 
 
@@ -90,13 +89,6 @@ class Numpy2TorchActionProcessorStep(ProcessorStep):
             torch_action = to_tensor(action, dtype=None)  # Preserve original dtype
             new_transition[TransitionKey.ACTION] = torch_action
 
-        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
-        if TELEOP_ACTION_KEY in complementary_data:
-            teleop_action = complementary_data[TELEOP_ACTION_KEY]
-            if isinstance(teleop_action, EnvAction):
-                complementary_data[TELEOP_ACTION_KEY] = to_tensor(teleop_action)
-            new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
-
         return new_transition
 
     def transform_features(

src/lerobot/processor/hil_processor.py

@@ -312,37 +312,6 @@ class TimeLimitProcessorStep(TruncatedProcessorStep):
         return features
 
 
-@ProcessorStepRegistry.register("gym_hil_adapter_processor")
-class GymHILAdapterProcessorStep(ProcessorStep):
-    """
-    Adapts the output of the `gym-hil` environment to the format expected by `lerobot` processors.
-
-    This step normalizes the `transition` object by:
-    1. Copying `teleop_action` from `info` to `complementary_data`.
-    2. Copying `is_intervention` from `info` (using the string key) to `info` (using the enum key).
-    """
-
-    def __call__(self, transition: EnvTransition) -> EnvTransition:
-        info = transition.get(TransitionKey.INFO, {})
-        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
-
-        if TELEOP_ACTION_KEY in info:
-            complementary_data[TELEOP_ACTION_KEY] = info[TELEOP_ACTION_KEY]
-
-        if "is_intervention" in info:
-            info[TeleopEvents.IS_INTERVENTION] = info["is_intervention"]
-
-        transition[TransitionKey.INFO] = info
-        transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
-
-        return transition
-
-    def transform_features(
-        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
-    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
-        return features
-
-
 @dataclass
 @ProcessorStepRegistry.register("gripper_penalty_processor")
 class GripperPenaltyProcessorStep(ProcessorStep):

src/lerobot/processor/pipeline.py

@@ -413,7 +413,7 @@ class DataProcessorPipeline(HubMixin, Generic[TInput, TOutput]):
         Args:
             save_directory: The directory where the pipeline will be saved. If None, saves to
                 HF_LEROBOT_HOME/processors/{sanitized_pipeline_name}.
-            repo_id: ID of your repository on the Hub. Used only if `push_to_hub=true`.
+            repo_id: ID of your repository on the Hub. Used only if `push_to_hub=True`.
             push_to_hub: Whether or not to push your object to the Hugging Face Hub after saving it.
             card_kwargs: Additional arguments passed to the card template to customize the card.
             config_filename: The name of the JSON configuration file. If None, a name is

src/lerobot/processor/tokenizer_processor.py

@@ -37,6 +37,9 @@ from lerobot.utils.constants import (
     OBS_LANGUAGE_SUBTASK_ATTENTION_MASK,
     OBS_LANGUAGE_SUBTASK_TOKENS,
     OBS_LANGUAGE_TOKENS,
+    OBS_LANGUAGE_USER_PROMPT,
+    OBS_LANGUAGE_USER_PROMPT_ATTENTION_MASK,
+    OBS_LANGUAGE_USER_PROMPT_TOKENS,
 )
 from lerobot.utils.import_utils import _transformers_available
 
@@ -141,6 +144,32 @@ class TokenizerProcessorStep(ObservationProcessorStep):
 
         return None
 
+    def get_user_prompt(self, transition: EnvTransition) -> list[str] | None:
+        """
+        Extracts the user_prompt from the transition's complementary data.
+
+        Args:
+            transition: The environment transition.
+
+        Returns:
+            A list of user_prompt strings, or None if the user_prompt key is not found or the value is None.
+        """
+        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA)
+        if complementary_data is None:
+            return None
+
+        user_prompt = complementary_data.get("user_prompt")
+        if user_prompt is None:
+            return None
+
+        # Standardize to a list of strings for the tokenizer
+        if isinstance(user_prompt, str):
+            return [user_prompt]
+        elif isinstance(user_prompt, list) and all(isinstance(t, str) for t in user_prompt):
+            return user_prompt
+
+        return None
+    
     def get_subtask(self, transition: EnvTransition) -> list[str] | None:
         """
         Extracts the subtask from the transition's complementary data.
@@ -169,16 +198,16 @@ class TokenizerProcessorStep(ObservationProcessorStep):
 
     def observation(self, observation: RobotObservation) -> RobotObservation:
         """
-        Tokenizes the task description and adds it to the observation dictionary.
+        Tokenizes the task description and user_prompt (if available) and adds them to the observation dictionary.
 
-        This method retrieves the task, tokenizes it, moves the resulting tensors to the
+        This method retrieves the task and user_prompt, tokenizes them, moves the resulting tensors to the
         same device as other data in the transition, and updates the observation.
 
         Args:
             observation: The original observation dictionary.
 
         Returns:
-            The updated observation dictionary including token IDs and an attention mask.
+            The updated observation dictionary including token IDs and attention masks.
         """
         task = self.get_task(self.transition)
         if task is None:
@@ -204,11 +233,45 @@ class TokenizerProcessorStep(ObservationProcessorStep):
         new_observation[OBS_LANGUAGE_TOKENS] = tokenized_prompt["input_ids"]
         new_observation[OBS_LANGUAGE_ATTENTION_MASK] = tokenized_prompt["attention_mask"].to(dtype=torch.bool)
 
+        # Tokenize user_prompt if available
+        user_prompt = self.get_user_prompt(self.transition)
+        if user_prompt is not None:
+            tokenized_user_prompt = self._tokenize_text(user_prompt)
+
+            # Move new tokenized tensors to the detected device
+            if target_device is not None:
+                tokenized_user_prompt = {
+                    k: v.to(target_device) if isinstance(v, torch.Tensor) else v
+                    for k, v in tokenized_user_prompt.items()
+                }
+
+            # Add tokenized user_prompt to the observation
+            new_observation[OBS_LANGUAGE_USER_PROMPT_TOKENS] = tokenized_user_prompt["input_ids"]
+            new_observation[OBS_LANGUAGE_USER_PROMPT_ATTENTION_MASK] = tokenized_user_prompt["attention_mask"].to(dtype=torch.bool)
+
         # Tokenize subtask if available
         subtask = self.get_subtask(self.transition)
         if subtask is not None:
             tokenized_subtask = self._tokenize_text(subtask)
 
+            # Add EOS token at the end of each subtask sequence (before padding)
+            eos_token_id = self.input_tokenizer.eos_token_id
+            input_ids = tokenized_subtask["input_ids"]
+            attention_mask = tokenized_subtask["attention_mask"]
+            for i in range(input_ids.size(0)):
+                # Find the length of actual tokens (sum of attention mask)
+                seq_len = attention_mask[i].sum().item()
+
+                max_len = input_ids.size(1)
+                if seq_len >= max_len:
+                    raise ValueError(
+                        f"No room to append EOS: seq_len={seq_len} equals max_length={max_len}. "
+                        "Increase max_length or tokenize with padding=False then pad after adding EOS."
+                    )
+                # Add EOS token at the end
+                input_ids[i, seq_len] = eos_token_id
+                attention_mask[i, seq_len] = 1
+
             # Move new tokenized tensors to the detected device
             if target_device is not None:
                 tokenized_subtask = {
@@ -320,6 +383,28 @@ class TokenizerProcessorStep(ObservationProcessorStep):
                 type=FeatureType.LANGUAGE, shape=(self.max_length,)
             )
 
+        # Add features for user_prompt tokens and attention mask if they don't already exist
+        if OBS_LANGUAGE_USER_PROMPT_TOKENS not in features[PipelineFeatureType.OBSERVATION]:
+            features[PipelineFeatureType.OBSERVATION][OBS_LANGUAGE_USER_PROMPT_TOKENS] = PolicyFeature(
+                type=FeatureType.LANGUAGE, shape=(self.max_length,)
+            )
+
+        if OBS_LANGUAGE_USER_PROMPT_ATTENTION_MASK not in features[PipelineFeatureType.OBSERVATION]:
+            features[PipelineFeatureType.OBSERVATION][OBS_LANGUAGE_USER_PROMPT_ATTENTION_MASK] = PolicyFeature(
+                type=FeatureType.LANGUAGE, shape=(self.max_length,)
+            )
+
+        # Add features for subtask tokens and attention mask if they don't already exist
+        if OBS_LANGUAGE_SUBTASK_TOKENS not in features[PipelineFeatureType.OBSERVATION]:
+            features[PipelineFeatureType.OBSERVATION][OBS_LANGUAGE_SUBTASK_TOKENS] = PolicyFeature(
+                type=FeatureType.LANGUAGE, shape=(self.max_length,)
+            )
+
+        if OBS_LANGUAGE_SUBTASK_ATTENTION_MASK not in features[PipelineFeatureType.OBSERVATION]:
+            features[PipelineFeatureType.OBSERVATION][OBS_LANGUAGE_SUBTASK_ATTENTION_MASK] = PolicyFeature(
+                type=FeatureType.LANGUAGE, shape=(self.max_length,)
+            )
+
         return features
 
 
@@ -573,4 +658,4 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
         Returns:
             The updated dictionary of policy features.
         """
-        return features
+        return features

src/lerobot/rl/gym_manipulator.py

@@ -36,7 +36,6 @@ from lerobot.processor import (
     DeviceProcessorStep,
     EnvTransition,
     GripperPenaltyProcessorStep,
-    GymHILAdapterProcessorStep,
     ImageCropResizeProcessorStep,
     InterventionActionProcessorStep,
     MapDeltaActionToRobotActionStep,
@@ -380,7 +379,6 @@ def make_processors(
         ]
 
         env_pipeline_steps = [
-            GymHILAdapterProcessorStep(),
             Numpy2TorchActionProcessorStep(),
             VanillaObservationProcessorStep(),
             AddBatchDimensionProcessorStep(),
@@ -610,14 +608,7 @@ def control_loop(
 
     dataset = None
     if cfg.mode == "record":
-        if teleop_device:
+        action_features = teleop_device.action_features
-            action_features = teleop_device.action_features
-        else:
-            action_features = {
-                "dtype": "float32",
-                "shape": (4,),
-                "names": ["delta_x", "delta_y", "delta_z", "gripper"],
-            }
         features = {
             ACTION: action_features,
             REWARD: {"dtype": "float32", "shape": (1,), "names": None},
@@ -665,7 +656,7 @@ def control_loop(
         # Create a neutral action (no movement)
         neutral_action = torch.tensor([0.0, 0.0, 0.0], dtype=torch.float32)
         if use_gripper:
-            neutral_action = torch.cat([neutral_action, torch.tensor([0.0])])  # Gripper stay
+            neutral_action = torch.cat([neutral_action, torch.tensor([1.0])])  # Gripper stay
 
         # Use the new step function
         transition = step_env_and_process_transition(
@@ -734,8 +725,6 @@ def control_loop(
         precise_sleep(max(dt - (time.perf_counter() - step_start_time), 0.0))
 
     if dataset is not None and cfg.dataset.push_to_hub:
-        logging.info("Finalizing dataset before pushing to hub")
-        dataset.finalize()
         logging.info("Pushing dataset to hub")
         dataset.push_to_hub()
 

src/lerobot/robots/bi_openarm_follower/bi_openarm_follower.py

@@ -19,7 +19,6 @@ from functools import cached_property
 
 from lerobot.processor import RobotAction, RobotObservation
 from lerobot.robots.openarm_follower import OpenArmFollower, OpenArmFollowerConfig
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 
 from ..robot import Robot
 from .config_bi_openarm_follower import BiOpenArmFollowerConfig
@@ -113,7 +112,6 @@ class BiOpenArmFollower(Robot):
     def is_connected(self) -> bool:
         return self.left_arm.is_connected and self.right_arm.is_connected
 
-    @check_if_already_connected
     def connect(self, calibrate: bool = True) -> None:
         self.left_arm.connect(calibrate)
         self.right_arm.connect(calibrate)
@@ -135,7 +133,6 @@ class BiOpenArmFollower(Robot):
             "Motor ID configuration is typically done via manufacturer tools for CAN motors."
         )
 
-    @check_if_not_connected
     def get_observation(self) -> RobotObservation:
         obs_dict = {}
 
@@ -149,7 +146,6 @@ class BiOpenArmFollower(Robot):
 
         return obs_dict
 
-    @check_if_not_connected
     def send_action(
         self,
         action: RobotAction,
@@ -174,7 +170,6 @@ class BiOpenArmFollower(Robot):
 
         return {**prefixed_sent_action_left, **prefixed_sent_action_right}
 
-    @check_if_not_connected
     def disconnect(self):
         self.left_arm.disconnect()
         self.right_arm.disconnect()

src/lerobot/robots/bi_so_follower/bi_so_follower.py

@@ -19,7 +19,6 @@ from functools import cached_property
 
 from lerobot.processor import RobotAction, RobotObservation
 from lerobot.robots.so_follower import SOFollower, SOFollowerRobotConfig
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 
 from ..robot import Robot
 from .config_bi_so_follower import BiSOFollowerConfig
@@ -97,7 +96,6 @@ class BiSOFollower(Robot):
     def is_connected(self) -> bool:
         return self.left_arm.is_connected and self.right_arm.is_connected
 
-    @check_if_already_connected
     def connect(self, calibrate: bool = True) -> None:
         self.left_arm.connect(calibrate)
         self.right_arm.connect(calibrate)
@@ -118,7 +116,6 @@ class BiSOFollower(Robot):
         self.left_arm.setup_motors()
         self.right_arm.setup_motors()
 
-    @check_if_not_connected
     def get_observation(self) -> RobotObservation:
         obs_dict = {}
 
@@ -132,7 +129,6 @@ class BiSOFollower(Robot):
 
         return obs_dict
 
-    @check_if_not_connected
     def send_action(self, action: RobotAction) -> RobotAction:
         # Remove "left_" prefix
         left_action = {
@@ -152,7 +148,6 @@ class BiSOFollower(Robot):
 
         return {**prefixed_sent_action_left, **prefixed_sent_action_right}
 
-    @check_if_not_connected
     def disconnect(self):
         self.left_arm.disconnect()
         self.right_arm.disconnect()

src/lerobot/robots/hope_jr/hope_jr_arm.py

@@ -140,7 +140,7 @@ class HopeJrArm(Robot):
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
             start = time.perf_counter()
-            obs_dict[cam_key] = cam.read_latest()
+            obs_dict[cam_key] = cam.async_read()
             dt_ms = (time.perf_counter() - start) * 1e3
             logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
 

src/lerobot/robots/hope_jr/hope_jr_hand.py

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

src/lerobot/robots/koch_follower/koch_follower.py

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

src/lerobot/robots/lekiwi/lekiwi.py

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

src/lerobot/robots/omx_follower/omx_follower.py

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

src/lerobot/robots/openarm_follower/openarm_follower.py

@@ -23,7 +23,7 @@ from lerobot.cameras.utils import make_cameras_from_configs
 from lerobot.motors import Motor, MotorCalibration, MotorNormMode
 from lerobot.motors.damiao import DamiaoMotorsBus
 from lerobot.processor import RobotAction, RobotObservation
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..robot import Robot
 from ..utils import ensure_safe_goal_position
@@ -119,7 +119,6 @@ class OpenArmFollower(Robot):
         """Check if robot is connected."""
         return self.bus.is_connected and all(cam.is_connected for cam in self.cameras.values())
 
-    @check_if_already_connected
     def connect(self, calibrate: bool = True) -> None:
         """
         Connect to the robot and optionally calibrate.
@@ -127,6 +126,8 @@ class OpenArmFollower(Robot):
         We assume that at connection time, the arms are in a safe rest position,
         and torque can be safely disabled to run calibration if needed.
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} already connected")
 
         # Connect to CAN bus
         logger.info(f"Connecting arm on {self.config.port}...")
@@ -218,7 +219,6 @@ class OpenArmFollower(Robot):
             "Motor ID configuration is typically done via manufacturer tools for CAN motors."
         )
 
-    @check_if_not_connected
     def get_observation(self) -> RobotObservation:
         """
         Get current observation from robot including position, velocity, and torque.
@@ -228,6 +228,9 @@ class OpenArmFollower(Robot):
         """
         start = time.perf_counter()
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         obs_dict: dict[str, Any] = {}
 
         states = self.bus.sync_read_all_states()
@@ -241,7 +244,7 @@ class OpenArmFollower(Robot):
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
             start = time.perf_counter()
-            obs_dict[cam_key] = cam.read_latest()
+            obs_dict[cam_key] = cam.async_read()
             dt_ms = (time.perf_counter() - start) * 1e3
             logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
 
@@ -250,7 +253,6 @@ class OpenArmFollower(Robot):
 
         return obs_dict
 
-    @check_if_not_connected
     def send_action(
         self,
         action: RobotAction,
@@ -270,6 +272,8 @@ class OpenArmFollower(Robot):
         Returns:
             The action actually sent (potentially clipped)
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         goal_pos = {key.removesuffix(".pos"): val for key, val in action.items() if key.endswith(".pos")}
 
@@ -329,9 +333,10 @@ class OpenArmFollower(Robot):
 
         return {f"{motor}.pos": val for motor, val in goal_pos.items()}
 
-    @check_if_not_connected
     def disconnect(self):
         """Disconnect from robot."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         # Disconnect CAN bus
         self.bus.disconnect(self.config.disable_torque_on_disconnect)

src/lerobot/robots/reachy2/robot_reachy2.py

@@ -180,7 +180,7 @@ class Reachy2Robot(Robot):
 
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
-            obs_dict[cam_key] = cam.read_latest()
+            obs_dict[cam_key] = cam.async_read()
 
         return obs_dict