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 == 'push' ||
-        github.event_name == 'workflow_dispatch'
-      )
+      (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
+      github.event_name == 'push' ||
+      github.event_name == 'workflow_dispatch'
     outputs:
       image_tag: ${{ steps.set_tag.outputs.image_tag }}
     env:

.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:

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.
-- `lerobot/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
-- `lerobot/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
+- `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
+- `aliberts/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
+- `aliberts/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
 
 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 \
-        lerobot/paris_street \
-        lerobot/kitchen \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        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 \
-        lerobot/paris_street \
-        lerobot/kitchen \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        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 |           |           |           |
-| --------------------------------- | ---------- | ------- | --------- | --------- | --------- |
-|                                   | libx264    |         | libx265   |           | libsvtav1 |
-| repo_id                           | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
-| lerobot/pusht_image               | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
-| lerobot/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
-| lerobot/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
-| lerobot/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
+| video_images_size_ratio            | vcodec     | pix_fmt |           |           |           |
+| ---------------------------------- | ---------- | ------- | --------- | --------- | --------- |
+|                                    | libx264    |         | libx265   |           | libsvtav1 |
+| repo_id                            | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
+| lerobot/pusht_image                | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
+| aliberts/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
+| aliberts/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
+| aliberts/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
 
-| video_images_load_time_ratio      | vcodec  | pix_fmt |          |         |           |
-| --------------------------------- | ------- | ------- | -------- | ------- | --------- |
-|                                   | libx264 |         | libx265  |         | libsvtav1 |
-| repo_id                           | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
-| lerobot/pusht_image               | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
-| lerobot/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
-| lerobot/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
-| lerobot/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
+| video_images_load_time_ratio       | vcodec  | pix_fmt |          |         |           |
+| ---------------------------------- | ------- | ------- | -------- | ------- | --------- |
+|                                    | libx264 |         | libx265  |         | libsvtav1 |
+| repo_id                            | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
+| lerobot/pusht_image                | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
+| aliberts/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
+| aliberts/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
+| aliberts/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
 
-|                                   |          | vcodec   | pix_fmt      |          |           |              |
-| --------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
-|                                   |          | libx264  |              | libx265  |           | libsvtav1    |
-| repo_id                           | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
-| lerobot/pusht_image               | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
-|                                   | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
-|                                   | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
-| lerobot/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
-|                                   | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
-|                                   | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
-| lerobot/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
-|                                   | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
-|                                   | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
-| lerobot/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
-|                                   | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
-|                                   | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |
+|                                    |          | vcodec   | pix_fmt      |          |           |              |
+| ---------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
+|                                    |          | libx264  |              | libx265  |           | libsvtav1    |
+| repo_id                            | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
+| lerobot/pusht_image                | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
+|                                    | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
+|                                    | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
+| aliberts/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
+|                                    | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
+|                                    | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
+| aliberts/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
+|                                    | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
+|                                    | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
+| aliberts/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
+|                                    | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
+|                                    | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |

docs/source/_toctree.yml

@@ -27,6 +27,8 @@
     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
   title: "Datasets"

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

@@ -185,7 +185,7 @@ 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 \

docs/source/hope_jr.mdx

@@ -224,7 +224,7 @@ lerobot-record \
     --teleop.port=/dev/tty.usbmodem1201 \
     --teleop.id=right \
     --teleop.side=right \
-    --dataset.repo_id=<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 \
@@ -241,7 +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
 ```
 
@@ -249,13 +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
@@ -270,7 +270,7 @@ lerobot-record \
   --robot.side=right \
   --robot.cameras='{"main": {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}}' \
   --display_data=false \
-  --dataset.repo_id=<USER>/eval_hopejr \
+  --dataset.repo_id=nepyope/eval_hopejr \
   --dataset.single_task="Evaluate hopejr hand policy" \
   --dataset.num_episodes=10 \
   --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model

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-).
-
-## Step 1: Install [`miniforge`](https://conda-forge.org/download/)
+## 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:
 
@@ -40,7 +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`.
 
-## Step 3: Install LeRobot 🤗
+## Install LeRobot 🤗
 
 ### From Source
 

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/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/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}}' \
@@ -266,7 +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}}' \

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
 ```
 """

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 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --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 \
-        --dataset.repo_id=<USER>/so101_cube_in_cup \
+        --policy.path=lipsop/reuben_pi0 \
+        --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 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --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 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --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 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --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 \

examples/tutorial/rl/hilserl_example.py

@@ -4,6 +4,7 @@ from pathlib import Path
 from queue import Empty, Full
 
 import torch
+import torch.optim as optim
 
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
@@ -11,7 +12,6 @@ from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
 from lerobot.policies.sac.configuration_sac import SACConfig
 from lerobot.policies.sac.modeling_sac import SACPolicy
 from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
-from lerobot.rl.algorithms.sac import SACAlgorithm, SACAlgorithmConfig
 from lerobot.rl.buffer import ReplayBuffer
 from lerobot.rl.gym_manipulator import make_robot_env
 from lerobot.robots.so_follower import SO100FollowerConfig
@@ -40,9 +40,8 @@ def run_learner(
     policy_learner.train()
     policy_learner.to(device)
 
-    algo_config = SACAlgorithmConfig.from_policy_config(policy_learner.config)
-    algorithm = SACAlgorithm(policy=policy_learner, config=algo_config)
-    algorithm.make_optimizers()
+    # Create Adam optimizer from scratch - simple and clean
+    optimizer = optim.Adam(policy_learner.parameters(), lr=lr)
 
     print(f"[LEARNER] Online buffer capacity: {online_buffer.capacity}")
     print(f"[LEARNER] Offline buffer capacity: {offline_buffer.capacity}")
@@ -84,26 +83,24 @@ def run_learner(
                 else:
                     batch[key] = online_batch[key]
 
-            def batch_iter(b=batch):
-                while True:
-                    yield b
+            loss, _ = policy_learner.forward(batch)
 
-            stats = algorithm.update(batch_iter())
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
             training_step += 1
 
             if training_step % LOG_EVERY == 0:
-                log_dict = stats.to_log_dict()
                 print(
-                    f"[LEARNER] Training step {training_step}, "
-                    f"critic_loss: {log_dict.get('critic', 'N/A'):.4f}, "
+                    f"[LEARNER] Training step {training_step}, Loss: {loss.item():.4f}, "
                     f"Buffers: Online={len(online_buffer)}, Offline={len(offline_buffer)}"
                 )
 
             # Send updated parameters to actor every 10 training steps
             if training_step % SEND_EVERY == 0:
                 try:
-                    weights = algorithm.get_weights()
-                    parameters_queue.put_nowait(weights)
+                    state_dict = {k: v.cpu() for k, v in policy_learner.state_dict().items()}
+                    parameters_queue.put_nowait(state_dict)
                     print("[LEARNER] Sent updated parameters to actor")
                 except Full:
                     # Missing write due to queue not being consumed (should happen rarely)
@@ -147,15 +144,15 @@ def run_actor(
 
             while step < MAX_STEPS_PER_EPISODE and not shutdown_event.is_set():
                 try:
-                    new_weights = parameters_queue.get_nowait()
-                    policy_actor.load_state_dict(new_weights)
+                    new_params = parameters_queue.get_nowait()
+                    policy_actor.load_state_dict(new_params)
                     print("[ACTOR] Updated policy parameters from learner")
                 except Empty:  # No new updated parameters available from learner, waiting
                     pass
 
-                # Get action from policy (returns full action: continuous + discrete)
+                # Get action from policy
                 policy_obs = make_policy_obs(obs, device=device)
-                action_tensor = policy_actor.select_action(policy_obs)
+                action_tensor = policy_actor.select_action(policy_obs)  # predicts a single action
                 action = action_tensor.squeeze(0).cpu().numpy()
 
                 # Step environment

pyproject.toml

@@ -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
-    "torchcodec>=0.2.1,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bump dependency
-    "torchvision>=0.21.0,<0.26.0", # TODO: Bump dependency
+    "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
+    "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
+    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
 
     "draccus==0.10.0", # TODO: Remove ==
     "gymnasium>=1.1.1,<2.0.0",
@@ -360,9 +360,9 @@ ignore_errors = false
 module = "lerobot.cameras.*"
 ignore_errors = false
 
-[[tool.mypy.overrides]]
-module = "lerobot.motors.*"
-ignore_errors = false
+# [[tool.mypy.overrides]]
+# module = "lerobot.motors.*"
+# ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.robots.*"

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 DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, 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 = 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
@@ -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)
-        self.rotation = Cv2Rotation(self.rotation)
-        self.backend = Cv2Backends(self.backend)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
+
+        if self.rotation not in (
+            Cv2Rotation.NO_ROTATION,
+            Cv2Rotation.ROTATE_90,
+            Cv2Rotation.ROTATE_180,
+            Cv2Rotation.ROTATE_270,
+        ):
+            raise ValueError(
+                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
+            )
 
         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 = 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
@@ -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 DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, 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 = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> 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)
-        self.rotation = Cv2Rotation(self.rotation)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
+
+        if self.rotation not in (
+            Cv2Rotation.NO_ROTATION,
+            Cv2Rotation.ROTATE_90,
+            Cv2Rotation.ROTATE_180,
+            Cv2Rotation.ROTATE_270,
+        ):
+            raise ValueError(
+                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
+            )
 
         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 DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, 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/train.py

@@ -211,15 +211,3 @@ class TrainRLServerPipelineConfig(TrainPipelineConfig):
     # NOTE: In RL, we don't need an offline dataset
     # TODO: Make `TrainPipelineConfig.dataset` optional
     dataset: DatasetConfig | None = None  # type: ignore[assignment] # because the parent class has made it's type non-optional
-
-    # Algorithm name registered in RLAlgorithmConfig registry
-    algorithm: str = "sac"
-
-    # Data mixer strategy name. Currently supports "online_offline"
-    mixer: str = "online_offline"
-    # Fraction sampled from online replay when using OnlineOfflineMixer
-    online_ratio: float = 0.5
-
-    # RL trainer iterator
-    async_prefetch: bool = True
-    queue_size: int = 2

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/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,
@@ -163,6 +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)
@@ -520,6 +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
@@ -656,7 +659,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             repo_id (str): This is the repo id that will be used to fetch the dataset. Locally, the dataset
                 will be stored under root/repo_id.
             root (Path | None, optional): Local directory to use for downloading/writing files. You can also
-                set the HF_LEROBOT_HOME environment variable to point to a different location. Defaults to
+                set the LEROBOT_HOME environment variable to point to a different location. Defaults to
                 '~/.cache/huggingface/lerobot'.
             episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                 their episode_index in this list. Defaults to None.
@@ -1067,7 +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:
@@ -1078,6 +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:

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
-        filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
-        return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)
+        # When no filtering needed, Dataset uses memory-mapped loading for efficiency
+        # PyArrow loads the entire dataset into memory
+        if episodes is None:
+            return Dataset.from_parquet([str(path) for path in paths], features=features)
+
+        arrow_dataset = pa_ds.dataset(paths, format="parquet")
+        filter_expr = pa_ds.field("episode_index").isin(episodes)
+        table = arrow_dataset.to_table(filter=filter_expr)
+
+        if features is not None:
+            table = table.cast(features.arrow_schema)
+
+        return Dataset(table)
 
 
 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/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)])
-            self.init_state_id += self._reset_stride  # Change init_state_id when reset
+            raw_obs = self._env.set_init_state(self._init_states[self._init_state_id])
 
         # 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)
-            for motors in self.groups.values()
-            for m in motors
+            m: bus.read("Present_Position", m, normalize=False) for motors in 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: dict[str, Value]) -> None:
+    def sync_write(self, data_name: str, values: Value | 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]),
-                homing_offset=int(offsets[motor]),
-                range_min=int(mins[motor]),
-                range_max=int(maxes[motor]),
+                drive_mode=drive_modes[motor],
+                homing_offset=offsets[motor],
+                range_min=mins[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]),
-                range_min=int(mins[motor]),
-                range_max=int(maxes[motor]),
+                homing_offset=offsets[motor],
+                range_min=mins[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
-    baudrate: int
-    packet_start_time: float
-    packet_timeout: float
-    tx_time_per_byte: float
-    is_using: bool
-    port_name: str
-    ser: serial.Serial
-
-    def __init__(self, port_name: str) -> None: ...
+    def __init__(self, port_name):
+        self.is_open: bool
+        self.baudrate: int
+        self.packet_start_time: float
+        self.packet_timeout: float
+        self.tx_time_per_byte: float
+        self.is_using: bool
+        self.port_name: str
+        self.ser: serial.Serial
 
     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
-    ph: PortHandler
-    start_address: int
-    data_length: int
-    last_result: bool
-    is_param_changed: bool
-    param: list
-    data_dict: dict
+    def __init__(self, port, ph, start_address, data_length):
+        self.port: str
+        self.ph: PortHandler
+        self.start_address: int
+        self.data_length: int
+        self.last_result: bool
+        self.is_param_changed: bool
+        self.param: list
+        self.data_dict: dict
 
-    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
-    ph: PortHandler
-    start_address: int
-    data_length: int
-    is_param_changed: bool
-    param: list
-    data_dict: dict
+    def __init__(self, port, ph, start_address, data_length):
+        self.port: str
+        self.ph: PortHandler
+        self.start_address: int
+        self.data_length: int
+        self.is_param_changed: bool
+        self.param: list
+        self.data_dict: dict
 
-    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):
-            return [self._id_to_name(motors)]
-        elif isinstance(motors, Sequence):
-            return [m if isinstance(m, str) else self._id_to_name(m) for m in motors]
+        elif isinstance(motors, list):
+            return motors.copy()
         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`.
-                Defaults to `None`.
+            motor (int): Same semantics as :pymeth:`disable_torque`. Defaults to `None`.
             num_retry (int, optional): Number of additional retry attempts on communication failure.
                 Defaults to 0.
         """
         pass
 
     @contextmanager
-    def torque_disabled(self, motors: 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(
-        self, motors: NameOrID | Sequence[NameOrID] | None = None
-    ) -> dict[NameOrID, Value]:
+    def set_half_turn_homings(self, motors: NameOrID | list[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)
-        actual_positions = self.sync_read("Present_Position", motor_names, normalize=False)
+        self.reset_calibration(motors)
+        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
-    ) -> tuple[dict[str, Value], dict[str, Value]]:
+        self, motors: NameOrID | list[NameOrID] | None = None, display_values: bool = True
+    ) -> 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)
-            return normalized[id_]
+            id_value = self._normalize(id_value)
 
-        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."
-        self._write(addr, length, id_, int_value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
+        err_msg = f"Failed to write '{data_name}' on {id_=} with '{value}' after {num_retry + 1} tries."
+        self._write(addr, length, id_, value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
 
     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)
-            return {self._id_to_name(id_): value for id_, value in normalized.items()}
+            ids_values = self._normalize(ids_values)
 
-        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)
-        models = [self._id_to_model(id_) for id_ in raw_ids_values]
+        ids_values = self._get_ids_values_dict(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."
-        self._sync_write(
-            addr, length, int_ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
-        )
+        err_msg = f"Failed to sync write '{data_name}' with {ids_values=} after {num_retry + 1} tries."
+        self._sync_write(addr, length, ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
 
     def _sync_write(
         self,

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

@@ -1,4 +1,6 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#!/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.
@@ -12,7 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.policies.rlt.configuration_rlt import RLTConfig
-from lerobot.policies.rlt.modeling_rlt import RLTPolicy
+from .configuration_pi05 import PI05FullConfig
+from .modeling_pi05 import PI05FullPolicy
+from .processor_pi05 import make_pi05_full_pre_post_processors
 
-__all__ = ["RLTConfig", "RLTPolicy"]
+__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/rlt/configuration_rlt.py

@@ -1,156 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""RLT (RL Token) policy configuration.
-
-Reference: "RL Token: Bootstrapping Online RL with Vision-Language-Action Models"
-(Xu et al., Physical Intelligence, 2026)
-"""
-
-from __future__ import annotations
-
-from dataclasses import dataclass, field
-
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import NormalizationMode
-from lerobot.policies.sac.configuration_sac import ActorLearnerConfig, ConcurrencyConfig
-from lerobot.utils.constants import ACTION, OBS_IMAGE, OBS_STATE
-
-
-@dataclass
-class RLTokenConfig:
-    """Configuration for the RL-token encoder/decoder transformer."""
-
-    input_dim: int = 2048
-    rl_token_dim: int = 2048
-    num_encoder_layers: int = 2
-    num_decoder_layers: int = 2
-    num_heads: int = 8
-    ff_dim: int = 2048
-    dropout: float = 0.0
-
-
-@dataclass
-class RLTActorConfig:
-    """Configuration for the lightweight RL actor MLP."""
-
-    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
-    std: float = 0.1
-
-
-@dataclass
-class RLTCriticConfig:
-    """Configuration for the RLT critic MLP."""
-
-    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
-
-
-@PreTrainedConfig.register_subclass("rlt")
-@dataclass
-class RLTConfig(PreTrainedConfig):
-    """Configuration for the RLT (RL Token) policy.
-
-    RLT adds an RL-token encoder/decoder to a frozen VLA backbone, then trains
-    a lightweight actor-critic head using the RL token as state representation.
-    The frozen VLA also provides reference action chunks that the actor refines.
-    """
-
-    normalization_mapping: dict[str, NormalizationMode] = field(
-        default_factory=lambda: {
-            "VISUAL": NormalizationMode.MEAN_STD,
-            "STATE": NormalizationMode.MIN_MAX,
-            "ACTION": NormalizationMode.MIN_MAX,
-        }
-    )
-
-    dataset_stats: dict[str, dict[str, list[float]]] | None = field(
-        default_factory=lambda: {
-            OBS_IMAGE: {
-                "mean": [0.485, 0.456, 0.406],
-                "std": [0.229, 0.224, 0.225],
-            },
-            OBS_STATE: {"min": [0.0], "max": [1.0]},
-            ACTION: {"min": [0.0], "max": [1.0]},
-        }
-    )
-
-    # ── Device ──
-    device: str = "cuda"
-    storage_device: str = "cpu"
-
-    # ── VLA backbone ──
-    vla_checkpoint: str | None = None
-
-    # ── RL-token ──
-    rl_token: RLTokenConfig = field(default_factory=RLTokenConfig)
-
-    # ── Actor / Critic heads ──
-    actor: RLTActorConfig = field(default_factory=RLTActorConfig)
-    critic: RLTCriticConfig = field(default_factory=RLTCriticConfig)
-
-    # ── Action chunks ──
-    chunk_size: int = 10
-    vla_chunk_size: int = 50
-
-    # ── Training parameters ──
-    online_steps: int = 50000
-    offline_steps: int = 5000
-    online_buffer_capacity: int = 100000
-    offline_buffer_capacity: int = 100000
-    online_step_before_learning: int = 500
-    warmup_steps: int = 500
-    async_prefetch: bool = False
-
-    # ── Algorithm hyperparameters ──
-    utd_ratio: int = 5
-    policy_update_freq: int = 2
-    discount: float = 0.99
-    critic_lr: float = 3e-4
-    actor_lr: float = 3e-4
-    rl_token_lr: float = 1e-4
-    tau: float = 0.005
-    clip_grad_norm: float = 10.0
-    num_critics: int = 2
-    bc_reg_coeff: float = 0.1
-    ref_dropout: float = 0.5
-    chunk_stride: int = 2
-    vla_finetune_weight: float = 0.0
-
-    # ── Distributed ──
-    actor_learner_config: ActorLearnerConfig = field(default_factory=ActorLearnerConfig)
-    concurrency: ConcurrencyConfig = field(default_factory=ConcurrencyConfig)
-
-    def __post_init__(self):
-        super().__post_init__()
-
-    def get_optimizer_preset(self):
-        return None
-
-    def get_scheduler_preset(self):
-        return None
-
-    def validate_features(self) -> None:
-        if ACTION not in self.output_features:
-            raise ValueError("You must provide 'action' in the output features")
-
-    @property
-    def observation_delta_indices(self) -> list | None:
-        return None
-
-    @property
-    def action_delta_indices(self) -> list | None:
-        return None
-
-    @property
-    def reward_delta_indices(self) -> None:
-        return None

src/lerobot/policies/rlt/modeling_rlt.py

@@ -1,318 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""RLT (RL Token) policy networks.
-
-Reference: "RL Token: Bootstrapping Online RL with Vision-Language-Action Models"
-(Xu et al., Physical Intelligence, 2026)
-
-Architecture:
-  - RLTokenEncoder: compresses VLA token embeddings into a single compact RL token
-  - RLTokenDecoder: reconstructs VLA embeddings from the RL token (Stage 1 training only)
-  - RLTActor: refines VLA reference action chunks conditioned on (z_rl, proprioception, ref_action)
-  - RLTCritic: Q(x, action_chunk) where x = (z_rl, proprioception)
-  - RLTPolicy: bundles RL-token modules + actor into a PreTrainedPolicy for inference
-"""
-
-from __future__ import annotations
-
-import math
-
-import torch
-import torch.nn as nn
-from torch import Tensor
-
-from lerobot.policies.pretrained import PreTrainedPolicy
-from lerobot.policies.rlt.configuration_rlt import RLTConfig
-
-# ── Building blocks ──────────────────────────────────────────────────
-
-
-class MLP(nn.Module):
-    """Simple feedforward network with ReLU activations."""
-
-    def __init__(self, input_dim: int, hidden_dims: list[int], output_dim: int):
-        super().__init__()
-        layers: list[nn.Module] = []
-        prev = input_dim
-        for h in hidden_dims:
-            layers.append(nn.Linear(prev, h))
-            layers.append(nn.ReLU())
-            prev = h
-        layers.append(nn.Linear(prev, output_dim))
-        self.net = nn.Sequential(*layers)
-
-    def forward(self, x: Tensor) -> Tensor:
-        return self.net(x)
-
-
-# ── RL Token Encoder ─────────────────────────────────────────────────
-
-
-class RLTokenEncoder(nn.Module):
-    """Compress VLA token embeddings into a single RL token via a small transformer.
-
-    Appends a learnable ``e_rl`` embedding to the VLA token sequence, processes
-    through transformer encoder layers, and returns the output at the ``e_rl``
-    position as the RL token ``z_rl``.
-
-    Paper Eq. 1: z_rl = g_phi([z_{1:M}, e_rl])_{M+1}
-    """
-
-    def __init__(
-        self,
-        input_dim: int,
-        rl_token_dim: int,
-        num_layers: int,
-        num_heads: int,
-        ff_dim: int,
-        dropout: float = 0.0,
-    ):
-        super().__init__()
-        self.rl_token_dim = rl_token_dim
-
-        self.e_rl = nn.Parameter(torch.randn(1, 1, input_dim) * 0.02)
-
-        if input_dim != rl_token_dim:
-            self.input_proj = nn.Linear(input_dim, rl_token_dim)
-        else:
-            self.input_proj = nn.Identity()
-
-        encoder_layer = nn.TransformerEncoderLayer(
-            d_model=rl_token_dim,
-            nhead=num_heads,
-            dim_feedforward=ff_dim,
-            dropout=dropout,
-            batch_first=True,
-        )
-        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
-
-    def forward(self, z_vla: Tensor) -> Tensor:
-        """
-        Args:
-            z_vla: VLA token embeddings, shape ``(B, M, D)``.
-
-        Returns:
-            RL token ``z_rl``, shape ``(B, rl_token_dim)``.
-        """
-        batch_size = z_vla.shape[0]
-        e_rl = self.e_rl.expand(batch_size, -1, -1)
-        seq = torch.cat([z_vla, e_rl], dim=1)  # (B, M+1, D)
-        seq = self.input_proj(seq)
-        out = self.transformer(seq)
-        z_rl = out[:, -1, :]  # output at e_rl position
-        return z_rl
-
-
-# ── RL Token Decoder ─────────────────────────────────────────────────
-
-
-class RLTokenDecoder(nn.Module):
-    """Autoregressively reconstruct VLA embeddings from z_rl.
-
-    Used only during Stage 1 (offline RL-token training).
-
-    Paper Eq. 2: L_ro = E[sum_i || h(d([z_rl, z_bar_{1:i-1}]))_i - z_bar_i ||^2]
-    """
-
-    def __init__(
-        self,
-        rl_token_dim: int,
-        output_dim: int,
-        num_layers: int,
-        num_heads: int,
-        ff_dim: int,
-        dropout: float = 0.0,
-    ):
-        super().__init__()
-        self.output_dim = output_dim
-
-        if rl_token_dim != output_dim:
-            self.rl_proj = nn.Linear(rl_token_dim, output_dim)
-        else:
-            self.rl_proj = nn.Identity()
-
-        decoder_layer = nn.TransformerDecoderLayer(
-            d_model=output_dim,
-            nhead=num_heads,
-            dim_feedforward=ff_dim,
-            dropout=dropout,
-            batch_first=True,
-        )
-        self.transformer = nn.TransformerDecoder(decoder_layer, num_layers=num_layers)
-        self.output_head = nn.Linear(output_dim, output_dim)
-
-    def forward(self, z_rl: Tensor, z_vla_stopped: Tensor) -> Tensor:
-        """
-        Args:
-            z_rl: RL token, shape ``(B, D_rl)``.
-            z_vla_stopped: Stop-gradient VLA embeddings, shape ``(B, M, D)``.
-
-        Returns:
-            Reconstructed embeddings, shape ``(B, M, D)``.
-        """
-        seq_len = z_vla_stopped.shape[1]
-        z_rl_proj = self.rl_proj(z_rl).unsqueeze(1)
-
-        target = torch.cat([z_rl_proj, z_vla_stopped[:, :-1, :]], dim=1)
-
-        causal_mask = nn.Transformer.generate_square_subsequent_mask(seq_len, device=z_rl.device)
-
-        decoded = self.transformer(
-            tgt=target,
-            memory=z_rl_proj,
-            tgt_mask=causal_mask,
-        )
-        return self.output_head(decoded)  # (B, M, D)
-
-
-# ── Actor ────────────────────────────────────────────────────────────
-
-
-class RLTActor(nn.Module):
-    """Lightweight actor that refines VLA reference action chunks.
-
-    Paper Eq. 4: pi_theta(a_{1:C} | x, a_tilde_{1:C}) = N(mu_theta(x, a_tilde), sigma^2 I)
-
-    The actor is conditioned on both the RL state and the VLA's proposed action
-    chunk, acting as a "VLA-guided action editor".
-    """
-
-    def __init__(self, state_dim: int, action_chunk_dim: int, hidden_dims: list[int], std: float = 0.1):
-        super().__init__()
-        input_dim = state_dim + action_chunk_dim
-        self.net = MLP(input_dim, hidden_dims, action_chunk_dim)
-        self.log_std = math.log(std)
-
-    def forward(self, state: Tensor, ref_action_chunk: Tensor) -> Tensor:
-        """Return the mean action chunk.
-
-        Args:
-            state: RL state ``x = (z_rl, proprioception)``, shape ``(B, state_dim)``.
-            ref_action_chunk: Flattened VLA reference chunk, shape ``(B, C*d)``.
-
-        Returns:
-            Refined action chunk (mean), shape ``(B, C*d)``.
-        """
-        x = torch.cat([state, ref_action_chunk], dim=-1)
-        return self.net(x)
-
-    def sample(self, state: Tensor, ref_action_chunk: Tensor) -> tuple[Tensor, Tensor]:
-        """Sample an action and return (action, log_prob)."""
-        mean = self.forward(state, ref_action_chunk)
-        std = math.exp(self.log_std)
-        noise = torch.randn_like(mean) * std
-        action = mean + noise
-        log_prob = -0.5 * (noise / std).pow(2).sum(dim=-1) - mean.shape[-1] * math.log(
-            std * math.sqrt(2 * math.pi)
-        )
-        return action, log_prob
-
-
-# ── Policy (inference bundle) ────────────────────────────────────────
-
-
-class RLTPolicy(PreTrainedPolicy):
-    """RLT policy — bundles the RL-token encoder and actor for inference.
-
-    The frozen VLA backbone is **not** part of this module; it is loaded
-    separately and its embeddings / reference actions are passed in via the
-    observation dict (populated by the actor process or a preprocessor).
-
-    During training, the :class:`RLTAlgorithm` holds the critic, target networks,
-    and optimizers. This class only contains what is needed for ``select_action``.
-    """
-
-    name = "rlt"
-    config_class = RLTConfig
-
-    def __init__(self, config: RLTConfig, dataset_stats=None):
-        super().__init__(config, dataset_stats)
-        action_dim = config.output_features["action"].shape[0]
-        action_chunk_dim = config.chunk_size * action_dim
-        prop_feature = config.input_features.get("observation.state", None)
-        proprioception_dim = prop_feature.shape[0] if prop_feature is not None else 0
-
-        state_dim = config.rl_token.rl_token_dim + proprioception_dim
-
-        # RL-token encoder (frozen after Stage 1)
-        self.rl_token_encoder = RLTokenEncoder(
-            input_dim=config.rl_token.input_dim,
-            rl_token_dim=config.rl_token.rl_token_dim,
-            num_layers=config.rl_token.num_encoder_layers,
-            num_heads=config.rl_token.num_heads,
-            ff_dim=config.rl_token.ff_dim,
-            dropout=config.rl_token.dropout,
-        )
-
-        # RL-token decoder (used only during Stage 1 training)
-        self.rl_token_decoder = RLTokenDecoder(
-            rl_token_dim=config.rl_token.rl_token_dim,
-            output_dim=config.rl_token.input_dim,
-            num_layers=config.rl_token.num_decoder_layers,
-            num_heads=config.rl_token.num_heads,
-            ff_dim=config.rl_token.ff_dim,
-            dropout=config.rl_token.dropout,
-        )
-
-        # Actor MLP
-        self.actor = RLTActor(
-            state_dim=state_dim,
-            action_chunk_dim=action_chunk_dim,
-            hidden_dims=config.actor.hidden_dims,
-            std=config.actor.std,
-        )
-
-        self._action_dim = action_dim
-        self._action_chunk_dim = action_chunk_dim
-        self._state_dim = state_dim
-        self._proprioception_dim = proprioception_dim
-
-    @torch.no_grad()
-    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
-        """Select a refined action chunk given an observation.
-
-        Expects the observation dict to contain:
-          - ``"observation.vla_embeddings"``: VLA internal token embeddings ``(M, D)``
-          - ``"observation.reference_action"``: VLA reference chunk ``(C*d,)``
-          - ``"observation.state"`` (optional): proprioceptive state ``(P,)``
-
-        Returns:
-            Action chunk tensor of shape ``(C*d,)``.
-        """
-        self.eval()
-
-        vla_emb = batch["observation.vla_embeddings"]
-        if vla_emb.dim() == 2:
-            vla_emb = vla_emb.unsqueeze(0)
-
-        z_rl = self.rl_token_encoder(vla_emb)  # (1, D_rl)
-
-        parts = [z_rl]
-        if "observation.state" in batch and self._proprioception_dim > 0:
-            prop = batch["observation.state"]
-            if prop.dim() == 1:
-                prop = prop.unsqueeze(0)
-            parts.append(prop)
-
-        state = torch.cat(parts, dim=-1)
-
-        ref = batch["observation.reference_action"]
-        if ref.dim() == 1:
-            ref = ref.unsqueeze(0)
-
-        action = self.actor(state, ref)
-        return action.squeeze(0)
-
-    def reset(self):
-        pass

src/lerobot/policies/sac/modeling_sac.py

@@ -15,11 +15,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import math
 from collections.abc import Callable
 from dataclasses import asdict
+from typing import Literal
 
+import einops
+import numpy as np
 import torch
 import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 from torch.distributions import MultivariateNormal, TanhTransform, Transform, TransformedDistribution
 
@@ -47,13 +52,20 @@ class SACPolicy(
 
         # Determine action dimension and initialize all components
         continuous_action_dim = config.output_features[ACTION].shape[0]
-        self.encoder = SACObservationEncoder(config)
+        self._init_encoders()
+        self._init_critics(continuous_action_dim)
         self._init_actor(continuous_action_dim)
-        self._init_discrete_critic()
+        self._init_temperature()
 
     def get_optim_params(self) -> dict:
         optim_params = {
-            "actor": [self.actor.parameters()],
+            "actor": [
+                p
+                for n, p in self.actor.named_parameters()
+                if not n.startswith("encoder") or not self.shared_encoder
+            ],
+            "critic": self.critic_ensemble.parameters(),
+            "temperature": self.log_alpha,
         }
         if self.config.num_discrete_actions is not None:
             optim_params["discrete_critic"] = self.discrete_critic.parameters()
@@ -71,9 +83,10 @@ class SACPolicy(
     @torch.no_grad()
     def select_action(self, batch: dict[str, Tensor]) -> Tensor:
         """Select action for inference/evaluation"""
+
         observations_features = None
-        if self.encoder.has_images:
-            observations_features = self.encoder.get_cached_image_features(batch)
+        if self.shared_encoder and self.actor.encoder.has_images:
+            observations_features = self.actor.encoder.get_cached_image_features(batch)
 
         actions, _, _ = self.actor(batch, observations_features)
 
@@ -84,35 +97,372 @@ class SACPolicy(
 
         return actions
 
+    def critic_forward(
+        self,
+        observations: dict[str, Tensor],
+        actions: Tensor,
+        use_target: bool = False,
+        observation_features: Tensor | None = None,
+    ) -> Tensor:
+        """Forward pass through a critic network ensemble
+
+        Args:
+            observations: Dictionary of observations
+            actions: Action tensor
+            use_target: If True, use target critics, otherwise use ensemble critics
+
+        Returns:
+            Tensor of Q-values from all critics
+        """
+
+        critics = self.critic_target if use_target else self.critic_ensemble
+        q_values = critics(observations, actions, observation_features)
+        return q_values
+
+    def discrete_critic_forward(
+        self, observations, use_target=False, observation_features=None
+    ) -> torch.Tensor:
+        """Forward pass through a discrete critic network
+
+        Args:
+            observations: Dictionary of observations
+            use_target: If True, use target critics, otherwise use ensemble critics
+            observation_features: Optional pre-computed observation features to avoid recomputing encoder output
+
+        Returns:
+            Tensor of Q-values from the discrete critic network
+        """
+        discrete_critic = self.discrete_critic_target if use_target else self.discrete_critic
+        q_values = discrete_critic(observations, observation_features)
+        return q_values
+
     def forward(
         self,
         batch: dict[str, Tensor | dict[str, Tensor]],
+        model: Literal["actor", "critic", "temperature", "discrete_critic"] = "critic",
     ) -> dict[str, Tensor]:
-        """Actor forward pass."""
-        observations = batch.get("state", batch)
-        observation_features = batch.get("observation_feature") if isinstance(batch, dict) else None
-        actions, log_probs, means = self.actor(observations, observation_features)
-        return {"action": actions, "log_prob": log_probs, "action_mean": means}
+        """Compute the loss for the given model
 
-    def _init_actor(self, continuous_action_dim: int) -> None:
-        self.actor = Policy(
-            encoder=self.encoder,
-            network=MLP(input_dim=self.encoder.output_dim, **asdict(self.config.actor_network_kwargs)),
-            action_dim=continuous_action_dim,
-            encoder_is_shared=False,
-            **asdict(self.config.policy_kwargs),
+        Args:
+            batch: Dictionary containing:
+                - action: Action tensor
+                - reward: Reward tensor
+                - state: Observations tensor dict
+                - next_state: Next observations tensor dict
+                - done: Done mask tensor
+                - observation_feature: Optional pre-computed observation features
+                - next_observation_feature: Optional pre-computed next observation features
+            model: Which model to compute the loss for ("actor", "critic", "discrete_critic", or "temperature")
+
+        Returns:
+            The computed loss tensor
+        """
+        # Extract common components from batch
+        actions: Tensor = batch[ACTION]
+        observations: dict[str, Tensor] = batch["state"]
+        observation_features: Tensor = batch.get("observation_feature")
+
+        if model == "critic":
+            # Extract critic-specific components
+            rewards: Tensor = batch["reward"]
+            next_observations: dict[str, Tensor] = batch["next_state"]
+            done: Tensor = batch["done"]
+            next_observation_features: Tensor = batch.get("next_observation_feature")
+
+            loss_critic = self.compute_loss_critic(
+                observations=observations,
+                actions=actions,
+                rewards=rewards,
+                next_observations=next_observations,
+                done=done,
+                observation_features=observation_features,
+                next_observation_features=next_observation_features,
+            )
+
+            return {"loss_critic": loss_critic}
+
+        if model == "discrete_critic" and self.config.num_discrete_actions is not None:
+            # Extract critic-specific components
+            rewards: Tensor = batch["reward"]
+            next_observations: dict[str, Tensor] = batch["next_state"]
+            done: Tensor = batch["done"]
+            next_observation_features: Tensor = batch.get("next_observation_feature")
+            complementary_info = batch.get("complementary_info")
+            loss_discrete_critic = self.compute_loss_discrete_critic(
+                observations=observations,
+                actions=actions,
+                rewards=rewards,
+                next_observations=next_observations,
+                done=done,
+                observation_features=observation_features,
+                next_observation_features=next_observation_features,
+                complementary_info=complementary_info,
+            )
+            return {"loss_discrete_critic": loss_discrete_critic}
+        if model == "actor":
+            return {
+                "loss_actor": self.compute_loss_actor(
+                    observations=observations,
+                    observation_features=observation_features,
+                )
+            }
+
+        if model == "temperature":
+            return {
+                "loss_temperature": self.compute_loss_temperature(
+                    observations=observations,
+                    observation_features=observation_features,
+                )
+            }
+
+        raise ValueError(f"Unknown model type: {model}")
+
+    def update_target_networks(self):
+        """Update target networks with exponential moving average"""
+        for target_param, param in zip(
+            self.critic_target.parameters(),
+            self.critic_ensemble.parameters(),
+            strict=True,
+        ):
+            target_param.data.copy_(
+                param.data * self.config.critic_target_update_weight
+                + target_param.data * (1.0 - self.config.critic_target_update_weight)
+            )
+        if self.config.num_discrete_actions is not None:
+            for target_param, param in zip(
+                self.discrete_critic_target.parameters(),
+                self.discrete_critic.parameters(),
+                strict=True,
+            ):
+                target_param.data.copy_(
+                    param.data * self.config.critic_target_update_weight
+                    + target_param.data * (1.0 - self.config.critic_target_update_weight)
+                )
+
+    @property
+    def temperature(self) -> float:
+        """Return the current temperature value, always in sync with log_alpha."""
+        return self.log_alpha.exp().item()
+
+    def compute_loss_critic(
+        self,
+        observations,
+        actions,
+        rewards,
+        next_observations,
+        done,
+        observation_features: Tensor | None = None,
+        next_observation_features: Tensor | None = None,
+    ) -> Tensor:
+        with torch.no_grad():
+            next_action_preds, next_log_probs, _ = self.actor(next_observations, next_observation_features)
+
+            # 2- compute q targets
+            q_targets = self.critic_forward(
+                observations=next_observations,
+                actions=next_action_preds,
+                use_target=True,
+                observation_features=next_observation_features,
+            )
+
+            # subsample critics to prevent overfitting if use high UTD (update to date)
+            # TODO: Get indices before forward pass to avoid unnecessary computation
+            if self.config.num_subsample_critics is not None:
+                indices = torch.randperm(self.config.num_critics)
+                indices = indices[: self.config.num_subsample_critics]
+                q_targets = q_targets[indices]
+
+            # critics subsample size
+            min_q, _ = q_targets.min(dim=0)  # Get values from min operation
+            if self.config.use_backup_entropy:
+                min_q = min_q - (self.temperature * next_log_probs)
+
+            td_target = rewards + (1 - done) * self.config.discount * min_q
+
+        # 3- compute predicted qs
+        if self.config.num_discrete_actions is not None:
+            # NOTE: We only want to keep the continuous action part
+            # In the buffer we have the full action space (continuous + discrete)
+            # We need to split them before concatenating them in the critic forward
+            actions: Tensor = actions[:, :DISCRETE_DIMENSION_INDEX]
+        q_preds = self.critic_forward(
+            observations=observations,
+            actions=actions,
+            use_target=False,
+            observation_features=observation_features,
         )
 
-    def _init_discrete_critic(self) -> None:
-        if self.config.num_discrete_actions is None:
-            self.discrete_critic = None
-            return
+        # 4- Calculate loss
+        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
+        td_target_duplicate = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
+        # You compute the mean loss of the batch for each critic and then to compute the final loss you sum them up
+        critics_loss = (
+            F.mse_loss(
+                input=q_preds,
+                target=td_target_duplicate,
+                reduction="none",
+            ).mean(dim=1)
+        ).sum()
+        return critics_loss
+
+    def compute_loss_discrete_critic(
+        self,
+        observations,
+        actions,
+        rewards,
+        next_observations,
+        done,
+        observation_features=None,
+        next_observation_features=None,
+        complementary_info=None,
+    ):
+        # NOTE: We only want to keep the discrete action part
+        # In the buffer we have the full action space (continuous + discrete)
+        # We need to split them before concatenating them in the critic forward
+        actions_discrete: Tensor = actions[:, DISCRETE_DIMENSION_INDEX:].clone()
+        actions_discrete = torch.round(actions_discrete)
+        actions_discrete = actions_discrete.long()
+
+        discrete_penalties: Tensor | None = None
+        if complementary_info is not None:
+            discrete_penalties: Tensor | None = complementary_info.get("discrete_penalty")
+
+        with torch.no_grad():
+            # For DQN, select actions using online network, evaluate with target network
+            next_discrete_qs = self.discrete_critic_forward(
+                next_observations, use_target=False, observation_features=next_observation_features
+            )
+            best_next_discrete_action = torch.argmax(next_discrete_qs, dim=-1, keepdim=True)
+
+            # Get target Q-values from target network
+            target_next_discrete_qs = self.discrete_critic_forward(
+                observations=next_observations,
+                use_target=True,
+                observation_features=next_observation_features,
+            )
+
+            # Use gather to select Q-values for best actions
+            target_next_discrete_q = torch.gather(
+                target_next_discrete_qs, dim=1, index=best_next_discrete_action
+            ).squeeze(-1)
+
+            # Compute target Q-value with Bellman equation
+            rewards_discrete = rewards
+            if discrete_penalties is not None:
+                rewards_discrete = rewards + discrete_penalties
+            target_discrete_q = rewards_discrete + (1 - done) * self.config.discount * target_next_discrete_q
+
+        # Get predicted Q-values for current observations
+        predicted_discrete_qs = self.discrete_critic_forward(
+            observations=observations, use_target=False, observation_features=observation_features
+        )
+
+        # Use gather to select Q-values for taken actions
+        predicted_discrete_q = torch.gather(predicted_discrete_qs, dim=1, index=actions_discrete).squeeze(-1)
+
+        # Compute MSE loss between predicted and target Q-values
+        discrete_critic_loss = F.mse_loss(input=predicted_discrete_q, target=target_discrete_q)
+        return discrete_critic_loss
+
+    def compute_loss_temperature(self, observations, observation_features: Tensor | None = None) -> Tensor:
+        """Compute the temperature loss"""
+        # calculate temperature loss
+        with torch.no_grad():
+            _, log_probs, _ = self.actor(observations, observation_features)
+        temperature_loss = (-self.log_alpha.exp() * (log_probs + self.target_entropy)).mean()
+        return temperature_loss
+
+    def compute_loss_actor(
+        self,
+        observations,
+        observation_features: Tensor | None = None,
+    ) -> Tensor:
+        actions_pi, log_probs, _ = self.actor(observations, observation_features)
+
+        q_preds = self.critic_forward(
+            observations=observations,
+            actions=actions_pi,
+            use_target=False,
+            observation_features=observation_features,
+        )
+        min_q_preds = q_preds.min(dim=0)[0]
+
+        actor_loss = ((self.temperature * log_probs) - min_q_preds).mean()
+        return actor_loss
+
+    def _init_encoders(self):
+        """Initialize shared or separate encoders for actor and critic."""
+        self.shared_encoder = self.config.shared_encoder
+        self.encoder_critic = SACObservationEncoder(self.config)
+        self.encoder_actor = (
+            self.encoder_critic if self.shared_encoder else SACObservationEncoder(self.config)
+        )
+
+    def _init_critics(self, continuous_action_dim):
+        """Build critic ensemble, targets, and optional discrete critic."""
+        heads = [
+            CriticHead(
+                input_dim=self.encoder_critic.output_dim + continuous_action_dim,
+                **asdict(self.config.critic_network_kwargs),
+            )
+            for _ in range(self.config.num_critics)
+        ]
+        self.critic_ensemble = CriticEnsemble(encoder=self.encoder_critic, ensemble=heads)
+        target_heads = [
+            CriticHead(
+                input_dim=self.encoder_critic.output_dim + continuous_action_dim,
+                **asdict(self.config.critic_network_kwargs),
+            )
+            for _ in range(self.config.num_critics)
+        ]
+        self.critic_target = CriticEnsemble(encoder=self.encoder_critic, ensemble=target_heads)
+        self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
+
+        if self.config.use_torch_compile:
+            self.critic_ensemble = torch.compile(self.critic_ensemble)
+            self.critic_target = torch.compile(self.critic_target)
+
+        if self.config.num_discrete_actions is not None:
+            self._init_discrete_critics()
+
+    def _init_discrete_critics(self):
+        """Build discrete discrete critic ensemble and target networks."""
         self.discrete_critic = DiscreteCritic(
-            encoder=self.encoder,
-            input_dim=self.encoder.output_dim,
+            encoder=self.encoder_critic,
+            input_dim=self.encoder_critic.output_dim,
             output_dim=self.config.num_discrete_actions,
             **asdict(self.config.discrete_critic_network_kwargs),
         )
+        self.discrete_critic_target = DiscreteCritic(
+            encoder=self.encoder_critic,
+            input_dim=self.encoder_critic.output_dim,
+            output_dim=self.config.num_discrete_actions,
+            **asdict(self.config.discrete_critic_network_kwargs),
+        )
+
+        # TODO: (maractingi, azouitine) Compile the discrete critic
+        self.discrete_critic_target.load_state_dict(self.discrete_critic.state_dict())
+
+    def _init_actor(self, continuous_action_dim):
+        """Initialize policy actor network and default target entropy."""
+        # NOTE: The actor select only the continuous action part
+        self.actor = Policy(
+            encoder=self.encoder_actor,
+            network=MLP(input_dim=self.encoder_actor.output_dim, **asdict(self.config.actor_network_kwargs)),
+            action_dim=continuous_action_dim,
+            encoder_is_shared=self.shared_encoder,
+            **asdict(self.config.policy_kwargs),
+        )
+
+        self.target_entropy = self.config.target_entropy
+        if self.target_entropy is None:
+            dim = continuous_action_dim + (1 if self.config.num_discrete_actions is not None else 0)
+            self.target_entropy = -np.prod(dim) / 2
+
+    def _init_temperature(self) -> None:
+        """Set up temperature parameter (log_alpha)."""
+        temp_init = self.config.temperature_init
+        self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
 
 
 class SACObservationEncoder(nn.Module):

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/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

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/normalize_processor.py

@@ -131,15 +131,6 @@ class _NormalizationMixin:
         if self.dtype is None:
             self.dtype = torch.float32
         self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)
-        self._reshape_visual_stats()
-
-    def _reshape_visual_stats(self) -> None:
-        """Reshape visual stats from ``[C]`` to ``[C, 1, 1]`` for image broadcasting."""
-        for key, feature in self.features.items():
-            if feature.type == FeatureType.VISUAL and key in self._tensor_stats:
-                for stat_name, stat_tensor in self._tensor_stats[key].items():
-                    if isinstance(stat_tensor, Tensor) and stat_tensor.ndim == 1:
-                        self._tensor_stats[key][stat_name] = stat_tensor.reshape(-1, 1, 1)
 
     def to(
         self, device: torch.device | str | None = None, dtype: torch.dtype | None = None
@@ -158,7 +149,6 @@ class _NormalizationMixin:
         if dtype is not None:
             self.dtype = dtype
         self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)
-        self._reshape_visual_stats()
         return self
 
     def state_dict(self) -> dict[str, Tensor]:
@@ -208,7 +198,6 @@ class _NormalizationMixin:
             # Don't load from state_dict, keep the explicitly provided stats
             # But ensure _tensor_stats is properly initialized
             self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)  # type: ignore[assignment]
-            self._reshape_visual_stats()
             return
 
         # Normal behavior: load stats from state_dict
@@ -219,7 +208,6 @@ class _NormalizationMixin:
             self._tensor_stats.setdefault(key, {})[stat_name] = tensor.to(
                 dtype=torch.float32, device=self.device
             )
-        self._reshape_visual_stats()
 
         # Reconstruct the original stats dict from tensor stats for compatibility with to() method
         # and other functions that rely on self.stats

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/__init__.py

@@ -1,13 +0,0 @@
-# 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.

src/lerobot/rl/actor.py

@@ -61,7 +61,7 @@ from lerobot.cameras import opencv  # noqa: F401
 from lerobot.configs import parser
 from lerobot.configs.train import TrainRLServerPipelineConfig
 from lerobot.policies.factory import make_policy
-from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.sac.modeling_sac import SACPolicy
 from lerobot.processor import TransitionKey
 from lerobot.rl.process import ProcessSignalHandler
 from lerobot.rl.queue import get_last_item_from_queue
@@ -248,16 +248,16 @@ def act_with_policy(
 
     logging.info("make_policy")
 
-    policy = make_policy(
+    ### Instantiate the policy in both the actor and learner processes
+    ### To avoid sending a SACPolicy object through the port, we create a policy instance
+    ### on both sides, the learner sends the updated parameters every n steps to update the actor's parameters
+    policy: SACPolicy = make_policy(
         cfg=cfg.policy,
         env_cfg=cfg.env,
     )
     policy = policy.eval()
     assert isinstance(policy, nn.Module)
 
-    # TODO: Re-enable processor pipeline once refactoring is validated against main
-    # preprocessor, postprocessor = None, None
-
     obs, info = online_env.reset()
     env_processor.reset()
     action_processor.reset()
@@ -288,6 +288,7 @@ def act_with_policy(
 
         # Time policy inference and check if it meets FPS requirement
         with policy_timer:
+            # Extract observation from transition for policy
             action = policy.select_action(batch=observation)
         policy_fps = policy_timer.fps_last
 
@@ -648,12 +649,12 @@ def interactions_stream(
 #  Policy functions
 
 
-def update_policy_parameters(policy: PreTrainedPolicy, parameters_queue: Queue, device):
-    """Load the latest policy weights from the learner."""
+def update_policy_parameters(policy: SACPolicy, parameters_queue: Queue, device):
     bytes_state_dict = get_last_item_from_queue(parameters_queue, block=False)
     if bytes_state_dict is not None:
         logging.info("[ACTOR] Load new parameters from Learner.")
         state_dicts = bytes_to_state_dict(bytes_state_dict)
+
         # TODO: check encoder parameter synchronization possible issues:
         # 1. When shared_encoder=True, we're loading stale encoder params from actor's state_dict
         #    instead of the updated encoder params from critic (which is optimized separately)
@@ -663,9 +664,18 @@ def update_policy_parameters(policy: PreTrainedPolicy, parameters_queue: Queue,
         # - Send critic's encoder state when shared_encoder=True
         # - Skip encoder params entirely when freeze_vision_encoder=True
         # - Ensure discrete_critic gets correct encoder state (currently uses encoder_critic)
+
         # Load actor state dict
-        state_dicts = move_state_dict_to_device(state_dicts, device=device)
-        policy.load_state_dict(state_dicts)
+        actor_state_dict = move_state_dict_to_device(state_dicts["policy"], device=device)
+        policy.actor.load_state_dict(actor_state_dict)
+
+        # Load discrete critic if present
+        if hasattr(policy, "discrete_critic") and "discrete_critic" in state_dicts:
+            discrete_critic_state_dict = move_state_dict_to_device(
+                state_dicts["discrete_critic"], device=device
+            )
+            policy.discrete_critic.load_state_dict(discrete_critic_state_dict)
+            logging.info("[ACTOR] Loaded discrete critic parameters from Learner.")
 
 
 #  Utilities functions

src/lerobot/rl/algorithms/__init__.py

@@ -1,70 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import annotations
-
-import torch
-
-from lerobot.rl.algorithms.base import (
-    RLAlgorithm,
-    RLAlgorithmConfig,
-    TrainingStats,
-)
-from lerobot.rl.algorithms.rlt import RLTAlgorithm, RLTAlgorithmConfig
-from lerobot.rl.algorithms.sac import SACAlgorithm, SACAlgorithmConfig
-
-
-def make_algorithm(
-    policy: torch.nn.Module,
-    policy_cfg,
-    *,
-    algorithm_name: str,
-) -> RLAlgorithm:
-    """Construct an :class:`RLAlgorithm` from a policy and its config.
-
-    Algorithm selection is explicit via ``algorithm_name`` (from
-    ``cfg.algorithm``).
-
-    This is fully registry-driven — adding a new algorithm only requires
-    registering an ``RLAlgorithmConfig`` subclass; no changes here.
-
-    The returned algorithm has **no optimizers** yet.  On the learner side,
-    call ``algorithm.make_optimizers()`` afterwards to create them.  On the
-    actor side (inference-only), leave them empty.
-
-    Args:
-        policy: Instantiated policy (e.g. ``SACPolicy``).
-        policy_cfg: The policy's ``PreTrainedConfig`` with the hyper-parameters
-            expected by the algorithm config's ``from_policy_config`` class-method.
-        algorithm_name: Algorithm registry key to instantiate.
-    """
-    known = RLAlgorithmConfig.get_known_choices()
-    if algorithm_name not in known:
-        raise ValueError(f"No RLAlgorithmConfig registered for '{algorithm_name}'. Known: {list(known)}")
-
-    config_cls = RLAlgorithmConfig.get_choice_class(algorithm_name)
-    algo_config = config_cls.from_policy_config(policy_cfg)
-    return algo_config.build_algorithm(policy)
-
-
-__all__ = [
-    "RLAlgorithm",
-    "RLAlgorithmConfig",
-    "TrainingStats",
-    "SACAlgorithm",
-    "SACAlgorithmConfig",
-    "RLTAlgorithm",
-    "RLTAlgorithmConfig",
-    "make_algorithm",
-]

src/lerobot/rl/algorithms/base.py

@@ -1,183 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Base classes for RL algorithms.
-
-Defines the abstract interface that every algorithm must implement, a registry
-for algorithm configs, and a dataclass for training statistics.
-"""
-
-from __future__ import annotations
-
-import abc
-from collections.abc import Iterator
-from dataclasses import dataclass, field
-from typing import TYPE_CHECKING, Any
-
-import draccus
-import torch
-from torch import Tensor
-from torch.optim import Optimizer
-
-if TYPE_CHECKING:
-    from lerobot.rl.data_sources.data_mixer import DataMixer
-
-BatchType = dict[str, Any]
-
-
-@dataclass
-class TrainingStats:
-    """Returned by ``algorithm.update()`` for logging and checkpointing."""
-
-    # Generic containers for all algorithms
-    losses: dict[str, float] = field(default_factory=dict)
-    grad_norms: dict[str, float] = field(default_factory=dict)
-    extra: dict[str, float] = field(default_factory=dict)
-
-    def to_log_dict(self) -> dict[str, float]:
-        """Flatten all stats into a single dict for logging."""
-
-        d: dict[str, float] = {}
-        for name, val in self.losses.items():
-            d[name] = val
-        for name, val in self.grad_norms.items():
-            d[f"{name}_grad_norm"] = val
-        for name, val in self.extra.items():
-            d[name] = val
-        return d
-
-
-@dataclass
-class RLAlgorithmConfig(draccus.ChoiceRegistry):
-    """Registry for algorithm configs."""
-
-    def build_algorithm(self, policy: torch.nn.Module) -> RLAlgorithm:
-        """Construct the :class:`RLAlgorithm` for this config.
-
-        Must be overridden by every registered config subclass.
-        """
-        raise NotImplementedError(f"{type(self).__name__} must implement build_algorithm()")
-
-    @classmethod
-    def from_policy_config(cls, policy_cfg: Any) -> RLAlgorithmConfig:
-        """Build an algorithm config from a policy config.
-
-        Must be overridden by every registered config subclass.
-        """
-        raise NotImplementedError(f"{cls.__name__} must implement from_policy_config()")
-
-
-class RLAlgorithm(abc.ABC):
-    """Base for all RL algorithms."""
-
-    @abc.abstractmethod
-    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
-        """One complete training step.
-
-        The algorithm calls ``next(batch_iterator)`` as many times as it
-        needs (e.g. ``utd_ratio`` times for SAC) to obtain fresh batches.
-        The iterator is owned by the trainer; the algorithm just consumes
-        from it.
-        """
-        ...
-
-    def supports_offline_phase(self) -> bool:
-        """Whether this algorithm has an offline pretraining phase.
-
-        Algorithms like RLT (RL-token training) or ConRFT (Cal-QL pretraining)
-        return ``True`` here. The learner checks this before the main online
-        loop and routes to :meth:`offline_update` accordingly.
-        """
-        return False
-
-    def offline_update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
-        """One offline training step (called before any online collection).
-
-        Only called when :meth:`supports_offline_phase` returns ``True``.
-        Uses the same iterator protocol as :meth:`update`.
-        """
-        raise NotImplementedError(
-            f"{type(self).__name__} does not implement offline_update(). "
-            "Either override this method or return False from supports_offline_phase()."
-        )
-
-    def transition_to_online(self) -> None:  # noqa: B027
-        """Called once when switching from offline to online phase.
-
-        Use this to freeze modules trained offline, rebuild optimizers for the
-        online phase, reset step counters, etc.
-
-        Default is a no-op; subclasses override when they have an offline phase.
-        """
-
-    def configure_data_iterator(
-        self,
-        data_mixer: DataMixer,
-        batch_size: int,
-        *,
-        async_prefetch: bool = True,
-        queue_size: int = 2,
-    ) -> Iterator[BatchType]:
-        """Create the data iterator this algorithm needs.
-
-        The default implementation uses the standard ``data_mixer.get_iterator()``.
-        Algorithms that need specialised sampling should override this method.
-        """
-        return data_mixer.get_iterator(
-            batch_size=batch_size,
-            async_prefetch=async_prefetch,
-            queue_size=queue_size,
-        )
-
-    def make_optimizers(self) -> dict[str, Optimizer]:
-        """Create, store, and return the optimizers needed for training.
-
-        Called on the **learner** side after construction.  Subclasses must
-        override this with algorithm-specific optimizer setup.
-        """
-        return {}
-
-    def get_optimizers(self) -> dict[str, Optimizer]:
-        """Return optimizers for checkpointing / external scheduling."""
-        return {}
-
-    @property
-    def optimization_step(self) -> int:
-        """Current learner optimization step.
-
-        Part of the stable contract for checkpoint/resume. Algorithms can
-        either use this default storage or override for custom behavior.
-        """
-        return getattr(self, "_optimization_step", 0)
-
-    @optimization_step.setter
-    def optimization_step(self, value: int) -> None:
-        self._optimization_step = int(value)
-
-    def get_weights(self) -> dict[str, Any]:
-        """Policy state-dict to push to actors."""
-        return {}
-
-    @abc.abstractmethod
-    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
-        """Load policy state-dict received from the learner (inverse of ``get_weights``)."""
-
-    @torch.no_grad()
-    def get_observation_features(
-        self, observations: Tensor, next_observations: Tensor
-    ) -> tuple[Tensor | None, Tensor | None]:
-        """Pre-compute observation features (e.g. frozen encoder cache).
-
-        Returns ``(None, None)`` when caching is not applicable.
-        """
-        return None, None

src/lerobot/rl/algorithms/rlt/__init__.py

@@ -1,18 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from lerobot.rl.algorithms.rlt.configuration_rlt import RLTAlgorithmConfig
-from lerobot.rl.algorithms.rlt.rlt_algorithm import RLTAlgorithm
-
-__all__ = ["RLTAlgorithm", "RLTAlgorithmConfig"]

src/lerobot/rl/algorithms/rlt/configuration_rlt.py

@@ -1,83 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""RLT algorithm configuration."""
-
-from __future__ import annotations
-
-from dataclasses import dataclass
-from typing import TYPE_CHECKING
-
-import torch
-
-from lerobot.rl.algorithms.base import RLAlgorithmConfig
-
-if TYPE_CHECKING:
-    from lerobot.rl.algorithms.rlt.rlt_algorithm import RLTAlgorithm
-
-
-@RLAlgorithmConfig.register_subclass("rlt")
-@dataclass
-class RLTAlgorithmConfig(RLAlgorithmConfig):
-    """RLT-specific hyper-parameters that control the update loop."""
-
-    # ── Action chunks ──
-    chunk_size: int = 10
-    chunk_stride: int = 2
-
-    # ── Update cadence ──
-    utd_ratio: int = 5
-    policy_update_freq: int = 2
-    clip_grad_norm: float = 10.0
-
-    # ── Learning rates ──
-    actor_lr: float = 3e-4
-    critic_lr: float = 3e-4
-    rl_token_lr: float = 1e-4
-
-    # ── TD learning ──
-    discount: float = 0.99
-    tau: float = 0.005
-    num_critics: int = 2
-
-    # ── Policy constraint (paper Eq. 5) ──
-    bc_reg_coeff: float = 0.1
-    ref_dropout: float = 0.5
-
-    # ── Offline RL-token training ──
-    vla_finetune_weight: float = 0.0
-
-    @classmethod
-    def from_policy_config(cls, policy_cfg) -> RLTAlgorithmConfig:
-        """Build from an existing ``RLTConfig`` (cfg.policy)."""
-        return cls(
-            chunk_size=policy_cfg.chunk_size,
-            chunk_stride=policy_cfg.chunk_stride,
-            utd_ratio=policy_cfg.utd_ratio,
-            policy_update_freq=policy_cfg.policy_update_freq,
-            clip_grad_norm=policy_cfg.clip_grad_norm,
-            actor_lr=policy_cfg.actor_lr,
-            critic_lr=policy_cfg.critic_lr,
-            rl_token_lr=policy_cfg.rl_token_lr,
-            discount=policy_cfg.discount,
-            tau=policy_cfg.tau,
-            num_critics=policy_cfg.num_critics,
-            bc_reg_coeff=policy_cfg.bc_reg_coeff,
-            ref_dropout=policy_cfg.ref_dropout,
-            vla_finetune_weight=policy_cfg.vla_finetune_weight,
-        )
-
-    def build_algorithm(self, policy: torch.nn.Module) -> RLTAlgorithm:
-        from lerobot.rl.algorithms.rlt.rlt_algorithm import RLTAlgorithm
-
-        return RLTAlgorithm(policy=policy, config=self)

src/lerobot/rl/algorithms/rlt/rlt_algorithm.py

@@ -1,319 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""RLT (RL Token) algorithm.
-
-Implements the two-stage training from "RL Token: Bootstrapping Online RL
-with Vision-Language-Action Models" (Xu et al., Physical Intelligence, 2026).
-
-Stage 1 (offline): Train RL-token encoder/decoder via reconstruction loss.
-Stage 2 (online):  Train actor-critic with chunked TD, BC regularization,
-                   reference-action pass-through, and reference-action dropout.
-"""
-
-from __future__ import annotations
-
-import copy
-from collections.abc import Iterator
-from typing import Any
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F  # noqa: N812
-from torch import Tensor
-from torch.optim import Optimizer
-
-from lerobot.policies.rlt.modeling_rlt import MLP, RLTPolicy
-from lerobot.policies.utils import get_device_from_parameters
-from lerobot.rl.algorithms.base import (
-    BatchType,
-    RLAlgorithm,
-    TrainingStats,
-)
-from lerobot.rl.algorithms.rlt.configuration_rlt import RLTAlgorithmConfig
-from lerobot.utils.constants import ACTION
-
-
-class RLTCritic(nn.Module):
-    """Q-function over (state, action_chunk) pairs.
-
-    Paper Eq. 3: Q_psi(x, a_{1:C})
-
-    Training-only component — lives on the algorithm side, not in the policy.
-    """
-
-    def __init__(self, state_dim: int, action_chunk_dim: int, hidden_dims: list[int]):
-        super().__init__()
-        self.net = MLP(state_dim + action_chunk_dim, hidden_dims, output_dim=1)
-
-    def forward(self, state: Tensor, action_chunk: Tensor) -> Tensor:
-        x = torch.cat([state, action_chunk], dim=-1)
-        return self.net(x)
-
-
-class RLTAlgorithm(RLAlgorithm):
-    """RL Token: lightweight actor-critic on frozen VLA features.
-
-    Owns the ``RLTPolicy`` (RL-token encoder/decoder + actor), a critic
-    ensemble, and target networks.  All VLA-specific logic (embedding
-    extraction, reference actions) lives in ``_prepare_forward_batch``.
-    """
-
-    def __init__(self, policy: RLTPolicy, config: RLTAlgorithmConfig):
-        self.policy = policy
-        self.config = config
-        self.optimizers: dict[str, Optimizer] = {}
-        self._optimization_step: int = 0
-        self._device = get_device_from_parameters(self.policy)
-        self._is_online = False
-
-        self._init_critics()
-        self._move_to_device()
-
-    # ── Initialization ───────────────────────────────────────────────
-
-    def _init_critics(self) -> None:
-        state_dim = self.policy._state_dim
-        action_chunk_dim = self.policy._action_chunk_dim
-        hidden_dims = self.policy.config.critic.hidden_dims
-
-        self.critics = torch.nn.ModuleList(
-            [RLTCritic(state_dim, action_chunk_dim, hidden_dims) for _ in range(self.config.num_critics)]
-        )
-        self.critic_targets = torch.nn.ModuleList([copy.deepcopy(c) for c in self.critics])
-        for ct in self.critic_targets:
-            ct.requires_grad_(False)
-
-    def _move_to_device(self) -> None:
-        self.critics.to(self._device)
-        self.critic_targets.to(self._device)
-
-    # ── Offline phase (Stage 1): RL-token training ───────────────────
-
-    def supports_offline_phase(self) -> bool:
-        return True
-
-    def offline_update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
-        """Train RL-token encoder/decoder on demonstration data.
-
-        Paper Eq. 2: L_ro = E[ sum_i || h(d([z_rl, z_bar_{1:i-1}]))_i - z_bar_i ||^2 ]
-        """
-        batch = next(batch_iterator)
-
-        vla_embeddings = batch["state"]["observation.vla_embeddings"].to(self._device)
-        z_vla = vla_embeddings.detach()  # stop-gradient on VLA embeddings
-
-        z_rl = self.policy.rl_token_encoder(z_vla)
-        z_reconstructed = self.policy.rl_token_decoder(z_rl, z_vla)
-
-        loss_ro = F.mse_loss(z_reconstructed, z_vla)
-
-        self.optimizers["rl_token"].zero_grad()
-        loss_ro.backward()
-        torch.nn.utils.clip_grad_norm_(
-            list(self.policy.rl_token_encoder.parameters()) + list(self.policy.rl_token_decoder.parameters()),
-            max_norm=self.config.clip_grad_norm,
-        )
-        self.optimizers["rl_token"].step()
-
-        self._optimization_step += 1
-        return TrainingStats(losses={"loss_rl_token": loss_ro.item()})
-
-    def transition_to_online(self) -> None:
-        """Freeze RL-token modules; rebuild optimizers for actor-critic only."""
-        self.policy.rl_token_encoder.requires_grad_(False)
-        self.policy.rl_token_decoder.requires_grad_(False)
-        self._is_online = True
-
-        self.optimizers = {
-            "actor": torch.optim.Adam(self.policy.actor.parameters(), lr=self.config.actor_lr),
-            "critic": torch.optim.Adam(self.critics.parameters(), lr=self.config.critic_lr),
-        }
-        self._optimization_step = 0
-
-    # ── Online phase (Stage 2): Actor-Critic ─────────────────────────
-
-    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
-        """One full RLT update step with UTD critic warm-up.
-
-        Pulls ``utd_ratio`` batches. First ``utd_ratio - 1`` are critic-only;
-        the last batch also updates the actor (every ``policy_update_freq`` steps).
-        """
-        for _ in range(self.config.utd_ratio - 1):
-            batch = next(batch_iterator)
-            fb = self._prepare_forward_batch(batch)
-            self._critic_step(fb)
-            self._update_target_networks()
-
-        batch = next(batch_iterator)
-        fb = self._prepare_forward_batch(batch)
-        critic_loss = self._critic_step(fb)
-
-        stats = TrainingStats(losses={"loss_critic": critic_loss})
-
-        if self._optimization_step % self.config.policy_update_freq == 0:
-            actor_loss, bc_loss, q_val = self._actor_step(fb)
-            stats.losses["loss_actor"] = actor_loss
-            stats.extra["bc_loss"] = bc_loss
-            stats.extra["q_value_mean"] = q_val
-
-        self._update_target_networks()
-        self._optimization_step += 1
-        return stats
-
-    def _prepare_forward_batch(self, batch: BatchType) -> dict[str, Any]:
-        """Convert a replay batch into algorithm-ready tensors.
-
-        Extracts RL-token from VLA embeddings, builds RL state, reads
-        reference action from complementary_info.
-        """
-        obs = batch["state"]
-        next_obs = batch["next_state"]
-        device = self._device
-
-        vla_emb = obs["observation.vla_embeddings"].to(device)
-        next_vla_emb = next_obs["observation.vla_embeddings"].to(device)
-
-        with torch.no_grad():
-            z_rl = self.policy.rl_token_encoder(vla_emb)
-            z_rl_next = self.policy.rl_token_encoder(next_vla_emb)
-
-        parts = [z_rl]
-        next_parts = [z_rl_next]
-        if "observation.state" in obs and self.policy._proprioception_dim > 0:
-            prop = obs["observation.state"].to(device)
-            next_prop = next_obs["observation.state"].to(device)
-            parts.append(prop)
-            next_parts.append(next_prop)
-
-        state = torch.cat(parts, dim=-1)
-        next_state = torch.cat(next_parts, dim=-1)
-
-        action = batch[ACTION].to(device)
-        reward = batch["reward"].to(device)
-        done = batch["done"].to(device)
-
-        ref_action = None
-        comp_info = batch.get("complementary_info")
-        if comp_info is not None and "reference_action" in comp_info:
-            ref_action = comp_info["reference_action"].to(device)
-
-        return {
-            "state": state,
-            "next_state": next_state,
-            "action": action,
-            "reward": reward,
-            "done": done,
-            "reference_action": ref_action,
-        }
-
-    def _critic_step(self, fb: dict[str, Any]) -> float:
-        """Paper Eq. 3: chunked TD with clipped double-Q target."""
-        state = fb["state"]
-        next_state = fb["next_state"]
-        action = fb["action"]
-        reward = fb["reward"]
-        done = fb["done"]
-
-        with torch.no_grad():
-            ref = fb.get("reference_action")
-            if ref is None:
-                ref = torch.zeros_like(action)
-            next_action = self.policy.actor(next_state, ref)
-
-            target_qs = [ct(next_state, next_action) for ct in self.critic_targets]
-            min_target_q = torch.min(torch.cat(target_qs, dim=-1), dim=-1, keepdim=True).values
-
-            discount_chunk = self.config.discount**self.config.chunk_size
-            td_target = reward.unsqueeze(-1) + (1 - done.unsqueeze(-1)) * discount_chunk * min_target_q
-
-        q_preds = [c(state, action) for c in self.critics]
-        loss = sum(F.mse_loss(q, td_target) for q in q_preds)
-
-        self.optimizers["critic"].zero_grad()
-        loss.backward()
-        torch.nn.utils.clip_grad_norm_(self.critics.parameters(), max_norm=self.config.clip_grad_norm)
-        self.optimizers["critic"].step()
-        return loss.item()
-
-    def _actor_step(self, fb: dict[str, Any]) -> tuple[float, float, float]:
-        """Paper Eq. 5: maximize Q while staying near VLA reference.
-
-        L_pi(theta) = E[ -Q(x, a) + beta * ||a - a_tilde||^2 ]
-        With reference-action dropout applied to the actor's ref input.
-        """
-        state = fb["state"]
-        ref = fb.get("reference_action")
-        if ref is None:
-            ref = torch.zeros(state.shape[0], self.policy._action_chunk_dim, device=self._device)
-
-        # Reference-action dropout (paper Section IV-B)
-        mask = (torch.rand(ref.shape[0], 1, device=self._device) > self.config.ref_dropout).float()
-        ref_input = ref * mask
-
-        action = self.policy.actor(state, ref_input)
-
-        q_value = self.critics[0](state, action)
-
-        bc_loss = F.mse_loss(action, ref)
-
-        loss = -q_value.mean() + self.config.bc_reg_coeff * bc_loss
-
-        self.optimizers["actor"].zero_grad()
-        loss.backward()
-        torch.nn.utils.clip_grad_norm_(self.policy.actor.parameters(), max_norm=self.config.clip_grad_norm)
-        self.optimizers["actor"].step()
-
-        return loss.item(), bc_loss.item(), q_value.mean().item()
-
-    def _update_target_networks(self) -> None:
-        tau = self.config.tau
-        for critic, target in zip(self.critics, self.critic_targets, strict=True):
-            for p, tp in zip(critic.parameters(), target.parameters(), strict=True):
-                tp.data.copy_(tau * p.data + (1 - tau) * tp.data)
-
-    # ── Optimizer management ─────────────────────────────────────────
-
-    def make_optimizers(self) -> dict[str, Optimizer]:
-        """Create optimizers. Initially for RL-token (Stage 1)."""
-        self.optimizers = {
-            "rl_token": torch.optim.Adam(
-                list(self.policy.rl_token_encoder.parameters())
-                + list(self.policy.rl_token_decoder.parameters()),
-                lr=self.config.rl_token_lr,
-            ),
-            "actor": torch.optim.Adam(self.policy.actor.parameters(), lr=self.config.actor_lr),
-            "critic": torch.optim.Adam(self.critics.parameters(), lr=self.config.critic_lr),
-        }
-        return self.optimizers
-
-    def get_optimizers(self) -> dict[str, Optimizer]:
-        return self.optimizers
-
-    # ── Weight sync ──────────────────────────────────────────────────
-
-    def get_weights(self) -> dict[str, Any]:
-        """Push actor + RL-token encoder to actors (small footprint)."""
-        weights = {
-            "actor": self.policy.actor.state_dict(),
-            "rl_token_encoder": self.policy.rl_token_encoder.state_dict(),
-        }
-        return {k: {kk: vv.cpu() for kk, vv in v.items()} for k, v in weights.items()}
-
-    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
-        if "actor" in weights:
-            self.policy.actor.load_state_dict({k: v.to(device) for k, v in weights["actor"].items()})
-        if "rl_token_encoder" in weights:
-            self.policy.rl_token_encoder.load_state_dict(
-                {k: v.to(device) for k, v in weights["rl_token_encoder"].items()}
-            )

src/lerobot/rl/algorithms/sac/__init__.py

@@ -1,18 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from lerobot.rl.algorithms.sac.configuration_sac import SACAlgorithmConfig
-from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
-
-__all__ = ["SACAlgorithm", "SACAlgorithmConfig"]

src/lerobot/rl/algorithms/sac/configuration_sac.py

@@ -1,81 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""SAC algorithm configuration."""
-
-from __future__ import annotations
-
-from dataclasses import dataclass, field
-from typing import TYPE_CHECKING
-
-import torch
-
-from lerobot.policies.sac.configuration_sac import CriticNetworkConfig
-from lerobot.rl.algorithms.base import RLAlgorithmConfig
-
-if TYPE_CHECKING:
-    from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
-
-
-@RLAlgorithmConfig.register_subclass("sac")
-@dataclass
-class SACAlgorithmConfig(RLAlgorithmConfig):
-    """SAC-specific hyper-parameters that control the update loop."""
-
-    utd_ratio: int = 1
-    policy_update_freq: int = 1
-    clip_grad_norm: float = 40.0
-    actor_lr: float = 3e-4
-    critic_lr: float = 3e-4
-    temperature_lr: float = 3e-4
-    discount: float = 0.99
-    temperature_init: float = 1.0
-    target_entropy: float | None = None
-    use_backup_entropy: bool = True
-    critic_target_update_weight: float = 0.005
-    num_critics: int = 2
-    num_subsample_critics: int | None = None
-    num_discrete_actions: int | None = None
-    shared_encoder: bool = True
-    critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
-    discrete_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
-    use_torch_compile: bool = True
-
-    @classmethod
-    def from_policy_config(cls, policy_cfg) -> SACAlgorithmConfig:
-        """Build from an existing ``SACConfig`` (cfg.policy) for backwards compat."""
-        return cls(
-            utd_ratio=policy_cfg.utd_ratio,
-            policy_update_freq=policy_cfg.policy_update_freq,
-            clip_grad_norm=policy_cfg.grad_clip_norm,
-            actor_lr=policy_cfg.actor_lr,
-            critic_lr=policy_cfg.critic_lr,
-            temperature_lr=policy_cfg.temperature_lr,
-            discount=policy_cfg.discount,
-            temperature_init=policy_cfg.temperature_init,
-            target_entropy=policy_cfg.target_entropy,
-            use_backup_entropy=policy_cfg.use_backup_entropy,
-            critic_target_update_weight=policy_cfg.critic_target_update_weight,
-            num_critics=policy_cfg.num_critics,
-            num_subsample_critics=policy_cfg.num_subsample_critics,
-            num_discrete_actions=policy_cfg.num_discrete_actions,
-            shared_encoder=policy_cfg.shared_encoder,
-            critic_network_kwargs=policy_cfg.critic_network_kwargs,
-            discrete_critic_network_kwargs=policy_cfg.discrete_critic_network_kwargs,
-            use_torch_compile=policy_cfg.use_torch_compile,
-        )
-
-    def build_algorithm(self, policy: torch.nn.Module) -> SACAlgorithm:
-        from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
-
-        return SACAlgorithm(policy=policy, config=self)

src/lerobot/rl/algorithms/sac/sac_algorithm.py

@@ -1,409 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""SAC (Soft Actor-Critic) algorithm.
-
-This module encapsulates all SAC-specific training logic (critic, actor,
-temperature, and discrete-critic updates) behind the ``RLAlgorithm`` interface.
-"""
-
-from __future__ import annotations
-
-import math
-from collections.abc import Iterator
-from dataclasses import asdict
-from typing import Any
-
-import einops
-import numpy as np
-import torch
-import torch.nn as nn
-import torch.nn.functional as F  # noqa: N812
-from torch import Tensor
-from torch.optim import Optimizer
-
-from lerobot.policies.sac.modeling_sac import (
-    DISCRETE_DIMENSION_INDEX,
-    CriticEnsemble,
-    CriticHead,
-    DiscreteCritic,
-    SACObservationEncoder,
-    SACPolicy,
-)
-from lerobot.policies.utils import get_device_from_parameters
-from lerobot.rl.algorithms.base import (
-    BatchType,
-    RLAlgorithm,
-    TrainingStats,
-)
-from lerobot.rl.algorithms.sac.configuration_sac import SACAlgorithmConfig
-from lerobot.utils.constants import ACTION
-from lerobot.utils.transition import move_state_dict_to_device
-
-
-class SACAlgorithm(RLAlgorithm):
-    """Soft Actor-Critic with optional discrete-critic head.
-
-    Owns the ``SACPolicy`` and its optimizers.  All loss methods call
-    ``self.policy(batch_dict)`` rather than reaching into ``self.policy.actor``
-    directly, so any policy that returns ``{"action", "log_prob"}`` from its
-    ``forward()`` is compatible.
-    """
-
-    def __init__(
-        self,
-        policy: SACPolicy,
-        config: SACAlgorithmConfig,
-    ):
-        self.policy = policy
-        self.config = config
-        self.optimizers: dict[str, Optimizer] = {}
-        self._optimization_step: int = 0
-
-        self._device = get_device_from_parameters(self.policy)
-        self._init_critic_encoder()
-        self._init_critics()
-        self._init_temperature()
-        self._move_to_device()
-
-    def _init_critic_encoder(self) -> None:
-        """Build or share the encoder used by critics."""
-        if self.config.shared_encoder:
-            self.critic_encoder = self.policy.encoder
-            self.policy.actor.encoder_is_shared = True
-        else:
-            self.critic_encoder = SACObservationEncoder(self.policy.config)
-
-    def _init_critics(self) -> None:
-        """Build critic ensemble, targets, and optional discrete critic."""
-        action_dim = self.policy.config.output_features[ACTION].shape[0]
-        input_dim = self.critic_encoder.output_dim + action_dim
-
-        heads = [
-            CriticHead(input_dim=input_dim, **asdict(self.config.critic_network_kwargs))
-            for _ in range(self.config.num_critics)
-        ]
-        self.critic_ensemble = CriticEnsemble(encoder=self.critic_encoder, ensemble=heads)
-
-        target_heads = [
-            CriticHead(input_dim=input_dim, **asdict(self.config.critic_network_kwargs))
-            for _ in range(self.config.num_critics)
-        ]
-        self.critic_target = CriticEnsemble(encoder=self.critic_encoder, ensemble=target_heads)
-        self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
-
-        if self.config.use_torch_compile:
-            self.critic_ensemble = torch.compile(self.critic_ensemble)
-            self.critic_target = torch.compile(self.critic_target)
-
-        if self.config.num_discrete_actions is not None:
-            self._init_discrete_critic_target()
-
-    def _init_discrete_critic_target(self) -> None:
-        """Build only the target discrete critic."""
-        input_dim = self.critic_encoder.output_dim
-        self.discrete_critic_target = DiscreteCritic(
-            encoder=self.critic_encoder,
-            input_dim=input_dim,
-            output_dim=self.config.num_discrete_actions,
-            **asdict(self.config.discrete_critic_network_kwargs),
-        )
-        # TODO: (kmeftah) Compile the discrete critic
-        self.discrete_critic_target.load_state_dict(self.policy.discrete_critic.state_dict())
-
-    def _init_temperature(self) -> None:
-        """Set up temperature parameter (log_alpha) and default target entropy."""
-        temp_init = self.config.temperature_init
-        self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
-
-        action_dim = self.policy.config.output_features[ACTION].shape[0]
-        self.target_entropy = self.config.target_entropy
-        if self.target_entropy is None:
-            dim = action_dim + (1 if self.config.num_discrete_actions is not None else 0)
-            self.target_entropy = -np.prod(dim) / 2
-
-    def _move_to_device(self) -> None:
-        """Move algorithm-owned modules to the policy device."""
-        self.critic_ensemble.to(self._device)
-        self.critic_target.to(self._device)
-        self.log_alpha = nn.Parameter(self.log_alpha.data.to(self._device))
-        if hasattr(self, "discrete_critic_target"):
-            self.discrete_critic_target.to(self._device)
-
-    @property
-    def temperature(self) -> float:
-        return self.log_alpha.exp().item()
-
-    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
-        """Run one full SAC update with UTD critic warm-up.
-
-        Pulls ``utd_ratio`` batches from ``batch_iterator``.  The first
-        ``utd_ratio - 1`` batches are used for critic-only warm-up steps;
-        the last batch drives the full update (critic + actor + temperature).
-        """
-        for _ in range(self.config.utd_ratio - 1):
-            batch = next(batch_iterator)
-            forward_batch = self._prepare_forward_batch(batch)
-
-            loss_critic = self._compute_loss_critic(forward_batch)
-            self.optimizers["critic"].zero_grad()
-            loss_critic.backward()
-            torch.nn.utils.clip_grad_norm_(
-                self.critic_ensemble.parameters(),
-                max_norm=self.config.clip_grad_norm,
-            ).item()
-            self.optimizers["critic"].step()
-
-            if self.config.num_discrete_actions is not None:
-                loss_discrete = self._compute_loss_discrete_critic(forward_batch)
-                self.optimizers["discrete_critic"].zero_grad()
-                loss_discrete.backward()
-                torch.nn.utils.clip_grad_norm_(
-                    self.policy.discrete_critic.parameters(),
-                    max_norm=self.config.clip_grad_norm,
-                ).item()
-                self.optimizers["discrete_critic"].step()
-            self._update_target_networks()
-
-        batch = next(batch_iterator)
-        forward_batch = self._prepare_forward_batch(batch)
-
-        loss_critic = self._compute_loss_critic(forward_batch)
-        self.optimizers["critic"].zero_grad()
-        loss_critic.backward()
-        critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-            self.critic_ensemble.parameters(),
-            max_norm=self.config.clip_grad_norm,
-        ).item()
-        self.optimizers["critic"].step()
-
-        critic_loss_val = loss_critic.item()
-        stats = TrainingStats(
-            losses={"loss_critic": critic_loss_val},
-            grad_norms={"critic": critic_grad_norm},
-        )
-
-        if self.config.num_discrete_actions is not None:
-            loss_discrete = self._compute_loss_discrete_critic(forward_batch)
-            self.optimizers["discrete_critic"].zero_grad()
-            loss_discrete.backward()
-            dc_grad = torch.nn.utils.clip_grad_norm_(
-                self.policy.discrete_critic.parameters(),
-                max_norm=self.config.clip_grad_norm,
-            ).item()
-            self.optimizers["discrete_critic"].step()
-            stats.losses["loss_discrete_critic"] = loss_discrete.item()
-            stats.grad_norms["discrete_critic"] = dc_grad
-
-        if self._optimization_step % self.config.policy_update_freq == 0:
-            for _ in range(self.config.policy_update_freq):
-                actor_loss = self._compute_loss_actor(forward_batch)
-                self.optimizers["actor"].zero_grad()
-                actor_loss.backward()
-                actor_grad = torch.nn.utils.clip_grad_norm_(
-                    self.policy.actor.parameters(),
-                    max_norm=self.config.clip_grad_norm,
-                ).item()
-                self.optimizers["actor"].step()
-
-                temp_loss = self._compute_loss_temperature(forward_batch)
-                self.optimizers["temperature"].zero_grad()
-                temp_loss.backward()
-                temp_grad = torch.nn.utils.clip_grad_norm_(
-                    [self.log_alpha],
-                    max_norm=self.config.clip_grad_norm,
-                ).item()
-                self.optimizers["temperature"].step()
-
-            stats.losses["loss_actor"] = actor_loss.item()
-            stats.losses["loss_temperature"] = temp_loss.item()
-            stats.grad_norms["actor"] = actor_grad
-            stats.grad_norms["temperature"] = temp_grad
-            stats.extra["temperature"] = self.temperature
-
-        self._update_target_networks()
-
-        self._optimization_step += 1
-        return stats
-
-    def _compute_loss_critic(self, batch: dict[str, Any]) -> Tensor:
-        observations = batch["state"]
-        actions = batch[ACTION]
-        rewards = batch["reward"]
-        next_observations = batch["next_state"]
-        done = batch["done"]
-        obs_features = batch.get("observation_feature")
-        next_obs_features = batch.get("next_observation_feature")
-
-        with torch.no_grad():
-            next_output = self.policy({"state": next_observations, "observation_feature": next_obs_features})
-            next_actions = next_output["action"]
-            next_log_probs = next_output["log_prob"]
-
-            q_targets = self.critic_target(next_observations, next_actions, next_obs_features)
-
-            if self.config.num_subsample_critics is not None:
-                indices = torch.randperm(self.config.num_critics)
-                indices = indices[: self.config.num_subsample_critics]
-                q_targets = q_targets[indices]
-
-            min_q, _ = q_targets.min(dim=0)
-            if self.config.use_backup_entropy:
-                min_q = min_q - (self.temperature * next_log_probs)
-
-            td_target = rewards + (1 - done) * self.config.discount * min_q
-
-        if self.config.num_discrete_actions is not None:
-            actions = actions[:, :DISCRETE_DIMENSION_INDEX]
-
-        q_preds = self.critic_ensemble(observations, actions, obs_features)
-
-        td_target_dup = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
-        critics_loss = (F.mse_loss(input=q_preds, target=td_target_dup, reduction="none").mean(dim=1)).sum()
-        return critics_loss
-
-    def _compute_loss_discrete_critic(self, batch: dict[str, Any]) -> Tensor:
-        observations = batch["state"]
-        actions = batch[ACTION]
-        rewards = batch["reward"]
-        next_observations = batch["next_state"]
-        done = batch["done"]
-        obs_features = batch.get("observation_feature")
-        next_obs_features = batch.get("next_observation_feature")
-        complementary_info = batch.get("complementary_info")
-
-        actions_discrete: Tensor = actions[:, DISCRETE_DIMENSION_INDEX:].clone()
-        actions_discrete = torch.round(actions_discrete).long()
-
-        discrete_penalties: Tensor | None = None
-        if complementary_info is not None:
-            discrete_penalties = complementary_info.get("discrete_penalty")
-
-        with torch.no_grad():
-            next_discrete_qs = self.policy.discrete_critic(next_observations, next_obs_features)
-            best_next_action = torch.argmax(next_discrete_qs, dim=-1, keepdim=True)
-
-            target_next_qs = self.discrete_critic_target(next_observations, next_obs_features)
-            target_next_q = torch.gather(target_next_qs, dim=1, index=best_next_action).squeeze(-1)
-
-            rewards_disc = rewards
-            if discrete_penalties is not None:
-                rewards_disc = rewards + discrete_penalties
-            target_q = rewards_disc + (1 - done) * self.config.discount * target_next_q
-
-        predicted_qs = self.policy.discrete_critic(observations, obs_features)
-        predicted_q = torch.gather(predicted_qs, dim=1, index=actions_discrete).squeeze(-1)
-
-        return F.mse_loss(input=predicted_q, target=target_q)
-
-    def _compute_loss_actor(self, batch: dict[str, Any]) -> Tensor:
-        observations = batch["state"]
-        obs_features = batch.get("observation_feature")
-
-        output = self.policy({"state": observations, "observation_feature": obs_features})
-        actions_pi = output["action"]
-        log_probs = output["log_prob"]
-
-        q_preds = self.critic_ensemble(observations, actions_pi, obs_features)
-        min_q = q_preds.min(dim=0)[0]
-
-        return ((self.temperature * log_probs) - min_q).mean()
-
-    def _compute_loss_temperature(self, batch: dict[str, Any]) -> Tensor:
-        observations = batch["state"]
-        obs_features = batch.get("observation_feature")
-
-        with torch.no_grad():
-            output = self.policy({"state": observations, "observation_feature": obs_features})
-            log_probs = output["log_prob"]
-
-        return (-self.log_alpha.exp() * (log_probs + self.target_entropy)).mean()
-
-    def _update_target_networks(self) -> None:
-        tau = self.config.critic_target_update_weight
-        for target_p, p in zip(
-            self.critic_target.parameters(), self.critic_ensemble.parameters(), strict=True
-        ):
-            target_p.data.copy_(p.data * tau + target_p.data * (1.0 - tau))
-        if self.config.num_discrete_actions is not None:
-            for target_p, p in zip(
-                self.discrete_critic_target.parameters(),
-                self.policy.discrete_critic.parameters(),
-                strict=True,
-            ):
-                target_p.data.copy_(p.data * tau + target_p.data * (1.0 - tau))
-
-    def _prepare_forward_batch(self, batch: BatchType) -> dict[str, Any]:
-        """Build the dict expected by loss computation from a sampled batch."""
-        observations = batch["state"]
-        next_observations = batch["next_state"]
-
-        observation_features, next_observation_features = self.get_observation_features(
-            observations, next_observations
-        )
-        forward_batch: dict[str, Any] = {
-            ACTION: batch[ACTION],
-            "reward": batch["reward"],
-            "state": observations,
-            "next_state": next_observations,
-            "done": batch["done"],
-            "observation_feature": observation_features,
-            "next_observation_feature": next_observation_features,
-        }
-        if "complementary_info" in batch:
-            forward_batch["complementary_info"] = batch["complementary_info"]
-        return forward_batch
-
-    def make_optimizers(self) -> dict[str, Optimizer]:
-        """Create Adam optimizers for the SAC components and store them."""
-        actor_params = [
-            p
-            for n, p in self.policy.actor.named_parameters()
-            if not self.config.shared_encoder or not n.startswith("encoder")
-        ]
-        self.optimizers = {
-            "actor": torch.optim.Adam(actor_params, lr=self.config.actor_lr),
-            "critic": torch.optim.Adam(self.critic_ensemble.parameters(), lr=self.config.critic_lr),
-            "temperature": torch.optim.Adam([self.log_alpha], lr=self.config.temperature_lr),
-        }
-        if self.config.num_discrete_actions is not None:
-            self.optimizers["discrete_critic"] = torch.optim.Adam(
-                self.policy.discrete_critic.parameters(), lr=self.config.critic_lr
-            )
-        return self.optimizers
-
-    def get_optimizers(self) -> dict[str, Optimizer]:
-        return self.optimizers
-
-    def get_weights(self) -> dict[str, Any]:
-        """Policy state-dict to push to actors (includes actor + discrete critic)."""
-        return move_state_dict_to_device(self.policy.state_dict(), device="cpu")
-
-    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
-        """Load policy state-dict received from the learner."""
-        state = move_state_dict_to_device(weights, device=device)
-        self.policy.load_state_dict(state)
-
-    @torch.no_grad()
-    def get_observation_features(
-        self, observations: Tensor, next_observations: Tensor
-    ) -> tuple[Tensor | None, Tensor | None]:
-        if not self.config.shared_encoder:
-            return None, None
-        if self.policy.config.vision_encoder_name is None or not self.policy.config.freeze_vision_encoder:
-            return None, None
-        if not self.policy.encoder.has_images:
-            return None, None
-        observation_features = self.policy.encoder.get_cached_image_features(observations)
-        next_observation_features = self.policy.encoder.get_cached_image_features(next_observations)
-        return observation_features, next_observation_features

src/lerobot/rl/data_sources/__init__.py

@@ -1,17 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from lerobot.rl.data_sources.data_mixer import BatchType, DataMixer, OnlineOfflineMixer
-
-__all__ = ["BatchType", "DataMixer", "OnlineOfflineMixer"]

src/lerobot/rl/data_sources/data_mixer.py

@@ -1,94 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import annotations
-
-import abc
-from typing import Any
-
-from lerobot.rl.buffer import ReplayBuffer, concatenate_batch_transitions
-
-BatchType = dict[str, Any]
-
-
-class DataMixer(abc.ABC):
-    """Abstract interface for all data mixing strategies.
-
-    Subclasses must implement ``sample(batch_size)`` and may override
-    ``get_iterator`` for specialised iteration.
-    """
-
-    @abc.abstractmethod
-    def sample(self, batch_size: int) -> BatchType:
-        """Draw one batch of ``batch_size`` transitions."""
-        ...
-
-    def get_iterator(
-        self,
-        batch_size: int,
-        async_prefetch: bool = True,
-        queue_size: int = 2,
-    ):
-        """Infinite iterator that yields batches.
-
-        The default implementation repeatedly calls ``self.sample()``.
-        Subclasses with underlying buffer iterators (async prefetch)
-        should override this for better throughput.
-        """
-        while True:
-            yield self.sample(batch_size)
-
-
-class OnlineOfflineMixer(DataMixer):
-    """Mixes transitions from an online and an optional offline replay buffer.
-
-    When both buffers are present, each batch is constructed by sampling
-    ``ceil(batch_size * online_ratio)`` from the online buffer and the
-    remainder from the offline buffer, then concatenating.
-
-    This mixer assumes both online and offline buffers are present.
-    """
-
-    def __init__(
-        self,
-        online_buffer: ReplayBuffer,
-        offline_buffer: ReplayBuffer | None = None,
-        online_ratio: float = 1.0,
-    ):
-        if not 0.0 <= online_ratio <= 1.0:
-            raise ValueError(f"online_ratio must be in [0, 1], got {online_ratio}")
-        self.online_buffer = online_buffer
-        self.offline_buffer = offline_buffer
-        self.online_ratio = online_ratio
-
-    def sample(self, batch_size: int) -> BatchType:
-        if self.offline_buffer is None:
-            return self.online_buffer.sample(batch_size)
-
-        n_online = max(1, int(batch_size * self.online_ratio))
-        n_offline = batch_size - n_online
-
-        online_batch = self.online_buffer.sample(n_online)
-        offline_batch = self.offline_buffer.sample(n_offline)
-        return concatenate_batch_transitions(online_batch, offline_batch)
-
-    def get_iterator(
-        self,
-        batch_size: int,
-        async_prefetch: bool = True,
-        queue_size: int = 2,
-    ):
-        """Yield batches from online/offline mixed sampling."""
-        while True:
-            yield self.sample(batch_size)

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
-        else:
-            action_features = {
-                "dtype": "float32",
-                "shape": (4,),
-                "names": ["delta_x", "delta_y", "delta_z", "gripper"],
-            }
+        action_features = teleop_device.action_features
         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/rl/learner.py

@@ -65,11 +65,9 @@ from lerobot.configs.train import TrainRLServerPipelineConfig
 from lerobot.datasets.factory import make_dataset
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.policies.factory import make_policy
-from lerobot.rl.algorithms import make_algorithm
-from lerobot.rl.buffer import ReplayBuffer
-from lerobot.rl.data_sources import OnlineOfflineMixer
+from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.rl.buffer import ReplayBuffer, concatenate_batch_transitions
 from lerobot.rl.process import ProcessSignalHandler
-from lerobot.rl.trainer import RLTrainer
 from lerobot.rl.wandb_utils import WandBLogger
 from lerobot.robots import so_follower  # noqa: F401
 from lerobot.teleoperators import gamepad, so_leader  # noqa: F401
@@ -95,7 +93,7 @@ from lerobot.utils.train_utils import (
     save_checkpoint,
     update_last_checkpoint,
 )
-from lerobot.utils.transition import move_transition_to_device
+from lerobot.utils.transition import move_state_dict_to_device, move_transition_to_device
 from lerobot.utils.utils import (
     format_big_number,
     get_safe_torch_device,
@@ -266,8 +264,8 @@ def add_actor_information_and_train(
     - Transfers transitions from the actor to the replay buffer.
     - Logs received interaction messages.
     - Ensures training begins only when the replay buffer has a sufficient number of transitions.
-    - Delegates training updates to an ``RLAlgorithm`` (currently ``SACAlgorithm``).
-    - Periodically pushes updated weights to actors.
+    - Samples batches from the replay buffer and performs multiple critic updates.
+    - Periodically updates the actor, critic, and temperature optimizers.
     - Logs training statistics, including loss values and optimization frequency.
 
     NOTE: This function doesn't have a single responsibility, it should be split into multiple functions
@@ -286,15 +284,17 @@ def add_actor_information_and_train(
     # of 7%
     device = get_safe_torch_device(try_device=cfg.policy.device, log=True)
     storage_device = get_safe_torch_device(try_device=cfg.policy.storage_device)
+    clip_grad_norm_value = cfg.policy.grad_clip_norm
     online_step_before_learning = cfg.policy.online_step_before_learning
+    utd_ratio = cfg.policy.utd_ratio
     fps = cfg.env.fps
     log_freq = cfg.log_freq
     save_freq = cfg.save_freq
+    policy_update_freq = cfg.policy.policy_update_freq
     policy_parameters_push_frequency = cfg.policy.actor_learner_config.policy_parameters_push_frequency
     saving_checkpoint = cfg.save_checkpoint
     online_steps = cfg.policy.online_steps
-    async_prefetch = cfg.async_prefetch
-    queue_size = cfg.queue_size
+    async_prefetch = cfg.policy.async_prefetch
 
     # Initialize logging for multiprocessing
     if not use_threads(cfg):
@@ -306,7 +306,7 @@ def add_actor_information_and_train(
 
     logging.info("Initializing policy")
 
-    policy = make_policy(
+    policy: SACPolicy = make_policy(
         cfg=cfg.policy,
         env_cfg=cfg.env,
     )
@@ -315,24 +315,19 @@ def add_actor_information_and_train(
 
     policy.train()
 
-    algorithm = make_algorithm(
-        policy=policy,
-        policy_cfg=cfg.policy,
-        algorithm_name=cfg.algorithm,
-    )
+    push_actor_policy_to_queue(parameters_queue=parameters_queue, policy=policy)
 
-    # TODO: Re-enable processor pipeline once refactoring is validated against main
-    preprocessor, postprocessor = None, None
-
-    # Push initial policy weights to actors (same path as periodic push)
-    state_bytes = state_to_bytes(algorithm.get_weights())
-    parameters_queue.put(state_bytes)
     last_time_policy_pushed = time.time()
 
+    optimizers, lr_scheduler = make_optimizers_and_scheduler(cfg=cfg, policy=policy)
+
+    # If we are resuming, we need to load the training state
+    resume_optimization_step, resume_interaction_step = load_training_state(cfg=cfg, optimizers=optimizers)
+
     log_training_info(cfg=cfg, policy=policy)
 
     replay_buffer = initialize_replay_buffer(cfg, device, storage_device)
-    total_batch_size = cfg.batch_size
+    batch_size = cfg.batch_size
     offline_replay_buffer = None
 
     if cfg.dataset is not None:
@@ -341,70 +336,20 @@ def add_actor_information_and_train(
             device=device,
             storage_device=storage_device,
         )
-
-    # DataMixer: online-only or online/offline 50-50 mix
-    data_mixer = OnlineOfflineMixer(
-        online_buffer=replay_buffer,
-        offline_buffer=offline_replay_buffer,
-        online_ratio=cfg.online_ratio,
-    )
-    # RLTrainer owns the iterator, preprocessor, and creates optimizers.
-    trainer = RLTrainer(
-        algorithm=algorithm,
-        data_mixer=data_mixer,
-        batch_size=total_batch_size,
-        preprocessor=preprocessor,
-        action_dim=cfg.policy.output_features["action"].shape[0],
-        async_prefetch=async_prefetch,
-        queue_size=queue_size,
-    )
-
-    # If we are resuming, we need to load the training state
-    optimizers = algorithm.get_optimizers()
-    resume_optimization_step, resume_interaction_step = load_training_state(cfg=cfg, optimizers=optimizers)
+        batch_size: int = batch_size // 2  # We will sample from both replay buffer
 
     logging.info("Starting learner thread")
     interaction_message = None
     optimization_step = resume_optimization_step if resume_optimization_step is not None else 0
-    algorithm.optimization_step = optimization_step
     interaction_step_shift = resume_interaction_step if resume_interaction_step is not None else 0
 
     dataset_repo_id = None
     if cfg.dataset is not None:
         dataset_repo_id = cfg.dataset.repo_id
 
-    # ── Offline phase (e.g. RLT RL-token training, ConRFT Cal-QL pretraining) ──
-    offline_steps = getattr(cfg.policy, "offline_steps", 0)
-    if algorithm.supports_offline_phase() and offline_steps > 0 and offline_replay_buffer is not None:
-        logging.info(f"[LEARNER] Starting offline phase ({offline_steps} steps)")
-        offline_mixer = OnlineOfflineMixer(
-            online_buffer=offline_replay_buffer,
-            offline_buffer=None,
-            online_ratio=1.0,
-        )
-        offline_iterator = algorithm.configure_data_iterator(
-            data_mixer=offline_mixer,
-            batch_size=total_batch_size,
-            async_prefetch=async_prefetch,
-            queue_size=queue_size,
-        )
-        for step in range(offline_steps):
-            if shutdown_event is not None and shutdown_event.is_set():
-                logging.info("[LEARNER] Shutdown during offline phase. Exiting...")
-                return
-
-            stats = algorithm.offline_update(offline_iterator)
-
-            if step % log_freq == 0:
-                logging.info(f"[LEARNER] Offline step {step}/{offline_steps}: {stats.to_log_dict()}")
-                if wandb_logger:
-                    log_dict = stats.to_log_dict()
-                    log_dict["offline_step"] = step
-                    wandb_logger.log_dict(d=log_dict, mode="train", custom_step_key="offline_step")
-
-        algorithm.transition_to_online()
-        optimizers = algorithm.get_optimizers()
-        logging.info("[LEARNER] Offline phase complete, transitioned to online")
+    # Initialize iterators
+    online_iterator = None
+    offline_iterator = None
 
     # NOTE: THIS IS THE MAIN LOOP OF THE LEARNER
     while True:
@@ -435,22 +380,180 @@ def add_actor_information_and_train(
         if len(replay_buffer) < online_step_before_learning:
             continue
 
-        time_for_one_optimization_step = time.time()
+        if online_iterator is None:
+            online_iterator = replay_buffer.get_iterator(
+                batch_size=batch_size, async_prefetch=async_prefetch, queue_size=2
+            )
 
-        # One training step (trainer owns data_mixer iterator; algorithm owns UTD loop)
-        stats = trainer.training_step()
+        if offline_replay_buffer is not None and offline_iterator is None:
+            offline_iterator = offline_replay_buffer.get_iterator(
+                batch_size=batch_size, async_prefetch=async_prefetch, queue_size=2
+            )
+
+        time_for_one_optimization_step = time.time()
+        for _ in range(utd_ratio - 1):
+            # Sample from the iterators
+            batch = next(online_iterator)
+
+            if dataset_repo_id is not None:
+                batch_offline = next(offline_iterator)
+                batch = concatenate_batch_transitions(
+                    left_batch_transitions=batch, right_batch_transition=batch_offline
+                )
+
+            actions = batch[ACTION]
+            rewards = batch["reward"]
+            observations = batch["state"]
+            next_observations = batch["next_state"]
+            done = batch["done"]
+            check_nan_in_transition(observations=observations, actions=actions, next_state=next_observations)
+
+            observation_features, next_observation_features = get_observation_features(
+                policy=policy, observations=observations, next_observations=next_observations
+            )
+
+            # Create a batch dictionary with all required elements for the forward method
+            forward_batch = {
+                ACTION: actions,
+                "reward": rewards,
+                "state": observations,
+                "next_state": next_observations,
+                "done": done,
+                "observation_feature": observation_features,
+                "next_observation_feature": next_observation_features,
+                "complementary_info": batch["complementary_info"],
+            }
+
+            # Use the forward method for critic loss
+            critic_output = policy.forward(forward_batch, model="critic")
+
+            # Main critic optimization
+            loss_critic = critic_output["loss_critic"]
+            optimizers["critic"].zero_grad()
+            loss_critic.backward()
+            critic_grad_norm = torch.nn.utils.clip_grad_norm_(
+                parameters=policy.critic_ensemble.parameters(), max_norm=clip_grad_norm_value
+            )
+            optimizers["critic"].step()
+
+            # Discrete critic optimization (if available)
+            if policy.config.num_discrete_actions is not None:
+                discrete_critic_output = policy.forward(forward_batch, model="discrete_critic")
+                loss_discrete_critic = discrete_critic_output["loss_discrete_critic"]
+                optimizers["discrete_critic"].zero_grad()
+                loss_discrete_critic.backward()
+                discrete_critic_grad_norm = torch.nn.utils.clip_grad_norm_(
+                    parameters=policy.discrete_critic.parameters(), max_norm=clip_grad_norm_value
+                )
+                optimizers["discrete_critic"].step()
+
+            # Update target networks (main and discrete)
+            policy.update_target_networks()
+
+        # Sample for the last update in the UTD ratio
+        batch = next(online_iterator)
+
+        if dataset_repo_id is not None:
+            batch_offline = next(offline_iterator)
+            batch = concatenate_batch_transitions(
+                left_batch_transitions=batch, right_batch_transition=batch_offline
+            )
+
+        actions = batch[ACTION]
+        rewards = batch["reward"]
+        observations = batch["state"]
+        next_observations = batch["next_state"]
+        done = batch["done"]
+
+        check_nan_in_transition(observations=observations, actions=actions, next_state=next_observations)
+
+        observation_features, next_observation_features = get_observation_features(
+            policy=policy, observations=observations, next_observations=next_observations
+        )
+
+        # Create a batch dictionary with all required elements for the forward method
+        forward_batch = {
+            ACTION: actions,
+            "reward": rewards,
+            "state": observations,
+            "next_state": next_observations,
+            "done": done,
+            "observation_feature": observation_features,
+            "next_observation_feature": next_observation_features,
+        }
+
+        critic_output = policy.forward(forward_batch, model="critic")
+
+        loss_critic = critic_output["loss_critic"]
+        optimizers["critic"].zero_grad()
+        loss_critic.backward()
+        critic_grad_norm = torch.nn.utils.clip_grad_norm_(
+            parameters=policy.critic_ensemble.parameters(), max_norm=clip_grad_norm_value
+        ).item()
+        optimizers["critic"].step()
+
+        # Initialize training info dictionary
+        training_infos = {
+            "loss_critic": loss_critic.item(),
+            "critic_grad_norm": critic_grad_norm,
+        }
+
+        # Discrete critic optimization (if available)
+        if policy.config.num_discrete_actions is not None:
+            discrete_critic_output = policy.forward(forward_batch, model="discrete_critic")
+            loss_discrete_critic = discrete_critic_output["loss_discrete_critic"]
+            optimizers["discrete_critic"].zero_grad()
+            loss_discrete_critic.backward()
+            discrete_critic_grad_norm = torch.nn.utils.clip_grad_norm_(
+                parameters=policy.discrete_critic.parameters(), max_norm=clip_grad_norm_value
+            ).item()
+            optimizers["discrete_critic"].step()
+
+            # Add discrete critic info to training info
+            training_infos["loss_discrete_critic"] = loss_discrete_critic.item()
+            training_infos["discrete_critic_grad_norm"] = discrete_critic_grad_norm
+
+        # Actor and temperature optimization (at specified frequency)
+        if optimization_step % policy_update_freq == 0:
+            for _ in range(policy_update_freq):
+                # Actor optimization
+                actor_output = policy.forward(forward_batch, model="actor")
+                loss_actor = actor_output["loss_actor"]
+                optimizers["actor"].zero_grad()
+                loss_actor.backward()
+                actor_grad_norm = torch.nn.utils.clip_grad_norm_(
+                    parameters=policy.actor.parameters(), max_norm=clip_grad_norm_value
+                ).item()
+                optimizers["actor"].step()
+
+                # Add actor info to training info
+                training_infos["loss_actor"] = loss_actor.item()
+                training_infos["actor_grad_norm"] = actor_grad_norm
+
+                # Temperature optimization
+                temperature_output = policy.forward(forward_batch, model="temperature")
+                loss_temperature = temperature_output["loss_temperature"]
+                optimizers["temperature"].zero_grad()
+                loss_temperature.backward()
+                temp_grad_norm = torch.nn.utils.clip_grad_norm_(
+                    parameters=[policy.log_alpha], max_norm=clip_grad_norm_value
+                ).item()
+                optimizers["temperature"].step()
+
+                # Add temperature info to training info
+                training_infos["loss_temperature"] = loss_temperature.item()
+                training_infos["temperature_grad_norm"] = temp_grad_norm
+                training_infos["temperature"] = policy.temperature
 
         # Push policy to actors if needed
         if time.time() - last_time_policy_pushed > policy_parameters_push_frequency:
-            state_dicts = algorithm.get_weights()
-            state_bytes = state_to_bytes(state_dicts)
-            parameters_queue.put(state_bytes)
+            push_actor_policy_to_queue(parameters_queue=parameters_queue, policy=policy)
             last_time_policy_pushed = time.time()
 
-        training_infos = stats.to_log_dict()
+        # Update target networks (main and discrete)
+        policy.update_target_networks()
 
         # Log training metrics at specified intervals
-        optimization_step = algorithm.optimization_step
         if optimization_step % log_freq == 0:
             training_infos["replay_buffer_size"] = len(replay_buffer)
             if offline_replay_buffer is not None:
@@ -478,6 +581,7 @@ def add_actor_information_and_train(
                 custom_step_key="Optimization step",
             )
 
+        optimization_step += 1
         if optimization_step % log_freq == 0:
             logging.info(f"[LEARNER] Number of optimization step: {optimization_step}")
 
@@ -494,8 +598,6 @@ def add_actor_information_and_train(
                 offline_replay_buffer=offline_replay_buffer,
                 dataset_repo_id=dataset_repo_id,
                 fps=fps,
-                preprocessor=preprocessor,
-                postprocessor=postprocessor,
             )
 
 
@@ -580,8 +682,6 @@ def save_training_checkpoint(
     offline_replay_buffer: ReplayBuffer | None = None,
     dataset_repo_id: str | None = None,
     fps: int = 30,
-    preprocessor=None,
-    postprocessor=None,
 ) -> None:
     """
     Save training checkpoint and associated data.
@@ -605,8 +705,6 @@ def save_training_checkpoint(
         offline_replay_buffer: Optional offline replay buffer to save
         dataset_repo_id: Repository ID for dataset
         fps: Frames per second for dataset
-        preprocessor: Optional preprocessor pipeline to save
-        postprocessor: Optional postprocessor pipeline to save
     """
     logging.info(f"Checkpoint policy after step {optimization_step}")
     _num_digits = max(6, len(str(online_steps)))
@@ -623,8 +721,6 @@ def save_training_checkpoint(
         policy=policy,
         optimizer=optimizers,
         scheduler=None,
-        preprocessor=preprocessor,
-        postprocessor=postprocessor,
     )
 
     # Save interaction step manually
@@ -662,6 +758,58 @@ def save_training_checkpoint(
     logging.info("Resume training")
 
 
+def make_optimizers_and_scheduler(cfg: TrainRLServerPipelineConfig, policy: nn.Module):
+    """
+    Creates and returns optimizers for the actor, critic, and temperature components of a reinforcement learning policy.
+
+    This function sets up Adam optimizers for:
+    - The **actor network**, ensuring that only relevant parameters are optimized.
+    - The **critic ensemble**, which evaluates the value function.
+    - The **temperature parameter**, which controls the entropy in soft actor-critic (SAC)-like methods.
+
+    It also initializes a learning rate scheduler, though currently, it is set to `None`.
+
+    NOTE:
+    - If the encoder is shared, its parameters are excluded from the actor's optimization process.
+    - The policy's log temperature (`log_alpha`) is wrapped in a list to ensure proper optimization as a standalone tensor.
+
+    Args:
+        cfg: Configuration object containing hyperparameters.
+        policy (nn.Module): The policy model containing the actor, critic, and temperature components.
+
+    Returns:
+        Tuple[Dict[str, torch.optim.Optimizer], Optional[torch.optim.lr_scheduler._LRScheduler]]:
+        A tuple containing:
+        - `optimizers`: A dictionary mapping component names ("actor", "critic", "temperature") to their respective Adam optimizers.
+        - `lr_scheduler`: Currently set to `None` but can be extended to support learning rate scheduling.
+
+    """
+    optimizer_actor = torch.optim.Adam(
+        params=[
+            p
+            for n, p in policy.actor.named_parameters()
+            if not policy.config.shared_encoder or not n.startswith("encoder")
+        ],
+        lr=cfg.policy.actor_lr,
+    )
+    optimizer_critic = torch.optim.Adam(params=policy.critic_ensemble.parameters(), lr=cfg.policy.critic_lr)
+
+    if cfg.policy.num_discrete_actions is not None:
+        optimizer_discrete_critic = torch.optim.Adam(
+            params=policy.discrete_critic.parameters(), lr=cfg.policy.critic_lr
+        )
+    optimizer_temperature = torch.optim.Adam(params=[policy.log_alpha], lr=cfg.policy.critic_lr)
+    lr_scheduler = None
+    optimizers = {
+        "actor": optimizer_actor,
+        "critic": optimizer_critic,
+        "temperature": optimizer_temperature,
+    }
+    if cfg.policy.num_discrete_actions is not None:
+        optimizers["discrete_critic"] = optimizer_discrete_critic
+    return optimizers, lr_scheduler
+
+
 # Training setup functions
 
 
@@ -866,6 +1014,33 @@ def initialize_offline_replay_buffer(
 # Utilities/Helpers functions
 
 
+def get_observation_features(
+    policy: SACPolicy, observations: torch.Tensor, next_observations: torch.Tensor
+) -> tuple[torch.Tensor | None, torch.Tensor | None]:
+    """
+    Get observation features from the policy encoder. It act as cache for the observation features.
+    when the encoder is frozen, the observation features are not updated.
+    We can save compute by caching the observation features.
+
+    Args:
+        policy: The policy model
+        observations: The current observations
+        next_observations: The next observations
+
+    Returns:
+        tuple: observation_features, next_observation_features
+    """
+
+    if policy.config.vision_encoder_name is None or not policy.config.freeze_vision_encoder:
+        return None, None
+
+    with torch.no_grad():
+        observation_features = policy.actor.encoder.get_cached_image_features(observations)
+        next_observation_features = policy.actor.encoder.get_cached_image_features(next_observations)
+
+    return observation_features, next_observation_features
+
+
 def use_threads(cfg: TrainRLServerPipelineConfig) -> bool:
     return cfg.policy.concurrency.learner == "threads"
 
@@ -916,6 +1091,23 @@ def check_nan_in_transition(
     return nan_detected
 
 
+def push_actor_policy_to_queue(parameters_queue: Queue, policy: nn.Module):
+    logging.debug("[LEARNER] Pushing actor policy to the queue")
+
+    # Create a dictionary to hold all the state dicts
+    state_dicts = {"policy": move_state_dict_to_device(policy.actor.state_dict(), device="cpu")}
+
+    # Add discrete critic if it exists
+    if hasattr(policy, "discrete_critic") and policy.discrete_critic is not None:
+        state_dicts["discrete_critic"] = move_state_dict_to_device(
+            policy.discrete_critic.state_dict(), device="cpu"
+        )
+        logging.debug("[LEARNER] Including discrete critic in state dict push")
+
+    state_bytes = state_to_bytes(state_dicts)
+    parameters_queue.put(state_bytes)
+
+
 def process_interaction_message(
     message, interaction_step_shift: int, wandb_logger: WandBLogger | None = None
 ):

src/lerobot/rl/trainer.py

@@ -1,132 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import annotations
-
-from collections.abc import Iterator
-from typing import Any
-
-import torch
-
-from lerobot.rl.algorithms.base import (
-    BatchType,
-    RLAlgorithm,
-    TrainingStats,
-)
-from lerobot.rl.data_sources.data_mixer import DataMixer
-from lerobot.utils.constants import ACTION
-
-
-def preprocess_rl_batch(preprocessor: Any, batch: BatchType, *, action_dim: int | None = None) -> BatchType:
-    """Apply a policy preprocessor to an RL batch."""
-    observations = batch["state"]
-    next_observations = batch["next_state"]
-    actions = batch[ACTION]
-
-    extra_action = None
-    if action_dim is not None and actions.shape[-1] > action_dim:
-        extra_action = actions[..., action_dim:]
-        actions = actions[..., :action_dim]
-
-    obs_action = {**observations, ACTION: actions}
-    obs_action = preprocessor(obs_action)
-    batch["state"] = {k: v for k, v in obs_action.items() if k.startswith("observation.")}
-    batch[ACTION] = obs_action[ACTION]
-
-    if extra_action is not None:
-        batch[ACTION] = torch.cat([batch[ACTION], extra_action], dim=-1)
-
-    next_obs = {**next_observations}
-    next_obs = preprocessor(next_obs)
-    batch["next_state"] = {k: v for k, v in next_obs.items() if k.startswith("observation.")}
-
-    return batch
-
-
-class _PreprocessedIterator:
-    """Iterator wrapper that preprocesses each sampled RL batch."""
-
-    __slots__ = ("_raw", "_preprocessor", "_action_dim")
-
-    def __init__(
-        self, raw_iterator: Iterator[BatchType], preprocessor: Any, action_dim: int | None = None
-    ) -> None:
-        self._raw = raw_iterator
-        self._preprocessor = preprocessor
-        self._action_dim = action_dim
-
-    def __iter__(self) -> _PreprocessedIterator:
-        return self
-
-    def __next__(self) -> BatchType:
-        batch = next(self._raw)
-        return preprocess_rl_batch(self._preprocessor, batch, action_dim=self._action_dim)
-
-
-class RLTrainer:
-    """Unified training step orchestrator.
-
-    Holds the algorithm, a DataMixer, and an optional preprocessor.
-    """
-
-    def __init__(
-        self,
-        algorithm: RLAlgorithm,
-        data_mixer: DataMixer,
-        batch_size: int,
-        *,
-        preprocessor: Any | None = None,
-        action_dim: int | None = None,
-        async_prefetch: bool = True,
-        queue_size: int = 2,
-    ):
-        self.algorithm = algorithm
-        self.data_mixer = data_mixer
-        self.batch_size = batch_size
-        self._preprocessor = preprocessor
-        self._action_dim = action_dim
-        self.async_prefetch = async_prefetch
-        self.queue_size = queue_size
-
-        self._iterator: Iterator[BatchType] | None = None
-
-        self.algorithm.make_optimizers()
-
-    def _build_data_iterator(self) -> Iterator[BatchType]:
-        """Create a fresh algorithm-configured iterator (optionally preprocessed)."""
-        raw = self.algorithm.configure_data_iterator(
-            data_mixer=self.data_mixer,
-            batch_size=self.batch_size,
-            async_prefetch=self.async_prefetch,
-            queue_size=self.queue_size,
-        )
-        if self._preprocessor is not None:
-            return _PreprocessedIterator(raw, self._preprocessor, self._action_dim)
-        return raw
-
-    def reset_data_iterator(self) -> None:
-        """Discard the current iterator so it will be rebuilt lazily next step."""
-        self._iterator = None
-
-    def set_data_mixer(self, data_mixer: DataMixer, *, reset: bool = True) -> None:
-        """Swap the active data mixer, optionally resetting the iterator."""
-        self.data_mixer = data_mixer
-        if reset:
-            self.reset_data_iterator()
-
-    def training_step(self) -> TrainingStats:
-        """Run one training step (algorithm-agnostic)."""
-        if self._iterator is None:
-            self._iterator = self._build_data_iterator()
-        return self.algorithm.update(self._iterator)

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
 

src/lerobot/robots/so_follower/config_so_follower.py

@@ -40,7 +40,7 @@ class SOFollowerConfig:
     cameras: dict[str, CameraConfig] = field(default_factory=dict)
 
     # Set to `True` for backward compatibility with previous policies/dataset
-    use_degrees: bool = True
+    use_degrees: bool = False
 
 
 @RobotConfig.register_subclass("so101_follower")

src/lerobot/robots/so_follower/so_follower.py

@@ -187,7 +187,7 @@ class SOFollower(Robot):
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
             start = time.perf_counter()
-            obs_dict[cam_key] = cam.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/unitree_g1/unitree_g1.py

@@ -324,7 +324,7 @@ class UnitreeG1(Robot):
 
         # Cameras - read images from ZMQ cameras
         for cam_name, cam in self._cameras.items():
-            obs[cam_name] = cam.read_latest()
+            obs[cam_name] = cam.async_read()
 
         return obs