docs: reframe HIL guide around general intervention workflow

Add a note in the installation guide explaining that users on WSL need to install evdev to avoid build issues.
See: https://github.com/huggingface/lerobot/issues/2528

Signed-off-by: Steven Palma <imstevenpmwork@ieee.org>
Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>

benchmarks/video/README.md

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

docs/source/_toctree.yml

@@ -17,6 +17,8 @@
     title: Train RL in Simulation
   - local: multi_gpu_training
     title: Multi GPU training
+  - local: hil_collection
+    title: Human In the Loop Data Collection
   - local: peft_training
     title: Training with PEFT (e.g., LoRA)
   title: "Tutorials"
@@ -29,6 +31,8 @@
     title: Using the Dataset Tools
   - local: dataset_subtask
     title: Using Subtasks in the Dataset
+  - local: streaming_video_encoding
+    title: Streaming Video Encoding
   title: "Datasets"
 - sections:
   - local: act

docs/source/act.mdx

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

docs/source/earthrover_mini_plus.mdx

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

docs/source/groot.mdx

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

docs/source/hil_collection.mdx

@@ -0,0 +1,237 @@
+# Human-In-the-Loop Data Collection
+
+Human-In-the-Loop (HIL) data collection lets you improve a trained policy by deploying it on a real robot while a human operator monitors and intervenes when needed. The intervention data — recovery movements and corrections — is recorded alongside autonomous segments, producing a richer training dataset that teaches the policy how to handle failures.
+
+---
+
+## Why Human-In-the-Loop?
+
+Standard behavioral cloning trains policies on successful demonstrations only. During deployment, small errors can compound and push the robot into states never seen during training (distribution shift). HIL data collection addresses this by:
+
+- Running the trained policy on the real robot
+- Having a human intervene when the robot is about to fail
+- Recording the human's recovery and correction as training data
+- Fine-tuning the policy on the combined dataset
+
+This produces a policy that not only knows how to perform the task, but also how to recover when things go wrong.
+
+---
+
+## How It Works
+
+During a HIL session, the human operator follows this loop within each episode:
+
+1. **Watch** the policy run autonomously
+2. **Pause** when failure is imminent — the robot holds its position
+3. **Take control** — teleoperate the robot back to a good state (recovery), then correct the behavior
+4. **Return control to the policy** — the policy resumes autonomous execution
+5. Repeat steps 2–4 as many times as needed during the episode
+6. **End the episode** when the task is complete, save and move on to the next rollout
+
+Both autonomous and human-controlled segments are recorded. The policy and human can alternate control multiple times within a single episode, and the episode continues from the current state after each handoff (no reset required just because intervention happened). This captures autonomous execution, recovery, and correction in one continuous trajectory. After collection, the combined dataset (original demonstrations + HIL data) is used to fine-tune the policy.
+
+This process can be repeated iteratively: deploy, collect, fine-tune, repeat — each round targeting the current policy's failure modes.
+
+```
+┌─────────────────────────────────────────────────────────────────────────┐
+│  Policy v0 (trained on demos)                                           │
+│       ↓                                                                 │
+│  HIL Collection (target current failure modes) → Fine-tune → Policy v1  │
+│       ↓                                                                 │
+│  HIL Collection (target new failure modes) → Fine-tune → Policy v2      │
+│       ↓                                                                 │
+│  ... (repeat until satisfactory performance)                            │
+└─────────────────────────────────────────────────────────────────────────┘
+```
+
+---
+
+## Hardware Requirements
+
+### Teleoperator Requirements
+
+The HIL data collection scripts require **teleoperators with active motors** that can:
+
+- Enable/disable torque programmatically
+- Move to target positions (to mirror the robot state when pausing)
+
+**Compatible teleoperators:**
+
+- `so101_leader` - SO-101 Leader Arm
+- `openarms_mini` - OpenArms Mini (via third-party plugin)
+
+---
+
+## Scripts
+
+Two scripts are provided depending on your policy's inference speed:
+
+| Script                       | Use Case                                   | Models                |
+| ---------------------------- | ------------------------------------------ | --------------------- |
+| `hil_data_collection.py`     | Standard synchronous inference             | ACT, Diffusion Policy |
+| `hil_data_collection_rtc.py` | Real-Time Chunking for high-latency models | Pi0, Pi0.5, SmolVLA   |
+
+---
+
+## Step-by-Step Guide
+
+### Step 1: Pre-train a Base Policy
+
+First, train a policy on your demonstration dataset:
+
+```bash
+python src/lerobot/scripts/lerobot_train.py \
+    --dataset.repo_id=your-username/demo-dataset \
+    --policy.type=pi0 \
+    --output_dir=outputs/pretrain \
+    --batch_size=32 \
+    --steps=50000
+```
+
+### Step 2: Collect HIL Data
+
+**Standard inference (ACT, Diffusion Policy):**
+
+```bash
+python examples/rac/hil_data_collection.py \
+    --robot.type=so100_follower \
+    --robot.port=/dev/tty.usbmodem58760431541 \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --teleop.type=so100_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
+    --dataset.repo_id=your-username/hil-dataset \
+    --dataset.single_task="Pick up the cube and place it in the bowl" \
+    --dataset.num_episodes=50
+```
+
+**With RTC for large models (Pi0, Pi0.5, SmolVLA):**
+
+For models with high inference latency, use the RTC script for smooth execution:
+
+```bash
+python examples/rac/hil_data_collection_rtc.py \
+    --robot.type=so100_follower \
+    --teleop.type=so100_leader \
+    --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
+    --dataset.repo_id=your-username/hil-rtc-dataset \
+    --dataset.single_task="Pick up the cube" \
+    --rtc.execution_horizon=20 \
+    --interpolation=true
+```
+
+**Controls (Conceptual):**
+
+The interaction model is:
+
+- **Pause input**: pause autonomous policy execution
+- **Takeover input**: transfer control to the human operator and record intervention data
+- **Return-to-policy input**: hand control back to the policy and continue the same episode
+- **Episode control inputs**: save/re-record/stop/reset as needed
+
+Exact key/pedal bindings can differ across scripts and hardware integrations. Use each script's printed controls as the source of truth for the concrete mapping on your setup.
+
+**The HIL Protocol:**
+
+1. Watch the policy run autonomously (teleop is idle/free)
+2. When you see imminent failure, trigger the **pause input**
+   - Policy stops
+   - Teleoperator moves to match robot position (torque enabled)
+   - No frames recorded during pause
+3. Trigger the **takeover input** to take control
+   - Teleoperator torque disabled, free to move
+   - **Recovery**: Teleoperate the robot back to a good state
+   - **Correction**: Correct the behavior
+   - All movements are recorded
+4. Trigger the **return-to-policy input**
+   - Policy resumes autonomous execution from the current state
+   - You can intervene again at any time (repeat steps 2–4)
+5. End and save the episode when the task is complete (or episode time limit is reached)
+6. **Reset**: Teleop moves to robot position, you can move the robot to the starting position
+7. Start the next episode
+
+**Foot Pedal Setup (Linux):**
+
+If using a USB foot pedal (PCsensor FootSwitch), ensure access:
+
+```bash
+sudo setfacl -m u:$USER:rw /dev/input/by-id/usb-PCsensor_FootSwitch-event-kbd
+```
+
+### Step 3: Fine-tune the Policy
+
+Fine-tune on the combined demonstration + HIL data:
+
+```bash
+python src/lerobot/scripts/lerobot_train.py \
+    --dataset.repo_id=your-username/hil-dataset \
+    --policy.type=pi0 \
+    --policy.pretrained_path=outputs/pretrain/checkpoints/last/pretrained_model \
+    --output_dir=outputs/hil_finetune \
+    --steps=20000
+```
+
+Then deploy the fine-tuned policy and repeat from Step 2 to target its remaining failure modes.
+
+---
+
+## Tips for Effective HIL Collection
+
+### When to Intervene
+
+Intervene when you see:
+
+- Robot about to make an irreversible mistake
+- Robot hesitating or showing uncertain behavior
+- Robot deviating from the expected trajectory
+
+### Recovery: Teleoperating Back to a Good State
+
+During recovery, teleoperate the robot back to a state where:
+
+- The robot is in a familiar, in-distribution configuration
+- The current subtask can still be completed
+- The recovery trajectory itself is informative training data
+
+### Quality of Corrections
+
+During correction:
+
+- Provide **confident, clean** trajectories
+- Complete the current subtask fully
+- Don't overcorrect or add unnecessary movements
+
+---
+
+## Related Work
+
+This HIL data collection approach builds on ideas from interactive imitation learning, including DAgger (Ross et al., 2011), HG-DAgger (Kelly et al., 2019), RaC (Hu et al., 2025), and RECAP (Physical Intelligence, 2025). See those works for a deeper treatment of the theory behind human-in-the-loop policy improvement.
+
+```bibtex
+@article{ross2011dagger,
+  title={A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning},
+  author={Ross, Stéphane and Gordon, Geoffrey and Bagnell, Drew},
+  journal={Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics},
+  year={2011}
+}
+
+@article{kelly2019hgdagger,
+  title={HG-DAgger: Interactive Imitation Learning with Human Experts},
+  author={Kelly, Michael and Sidrane, Chelsea and Driggs-Campbell, Katherine and Kochenderfer, Mykel J},
+  journal={arXiv preprint arXiv:1810.02890},
+  year={2019}
+}
+
+@article{hu2025rac,
+  title={RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction},
+  author={Hu, Zheyuan and Wu, Robyn and Enock, Naveen and Li, Jasmine and Kadakia, Riya and Erickson, Zackory and Kumar, Aviral},
+  journal={arXiv preprint arXiv:2509.07953},
+  year={2025}
+}
+
+@article{pi2025recap,
+  title={π0.6: a VLA That Learns From Experience},
+  author={Physical Intelligence},
+  year={2025}
+}
+```

docs/source/hope_jr.mdx

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

docs/source/il_robots.mdx

@@ -165,7 +165,7 @@ huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(hf auth whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 
@@ -185,7 +185,10 @@ lerobot-record \
     --display_data=true \
     --dataset.repo_id=${HF_USER}/record-test \
     --dataset.num_episodes=5 \
-    --dataset.single_task="Grab the black cube"
+    --dataset.single_task="Grab the black cube" \
+    --dataset.streaming_encoding=true \
+    # --dataset.vcodec=auto \
+    --dataset.encoder_threads=2
 ```
 </hfoption>
 <hfoption id="API example">
@@ -515,6 +518,9 @@ lerobot-record  \
   --display_data=false \
   --dataset.repo_id=${HF_USER}/eval_so100 \
   --dataset.single_task="Put lego brick into the transparent box" \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

docs/source/installation.mdx

@@ -40,6 +40,13 @@ conda install ffmpeg -c conda-forge
 >
 > - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
 
+> [!NOTE]
+> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to install `evdev` with the following command:
+>
+> ```bash
+> conda install evdev -c conda-forge
+> ```
+
 ## Step 3: Install LeRobot 🤗
 
 ### From Source

docs/source/lerobot-dataset-v3.mdx

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

docs/source/phone_teleop.mdx

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

docs/source/pi0.mdx

@@ -60,7 +60,7 @@ policy.type=pi0
 For training π₀, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
     --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
-python src/lerobot/scripts/lerobot_train.py\
+lerobot-train \
     --dataset.repo_id=your_dataset \
     --policy.type=pi05 \
     --output_dir=./outputs/pi05_training \

docs/source/reachy2.mdx

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

docs/source/sarm.mdx

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

docs/source/smolvla.mdx

@@ -106,6 +106,9 @@ lerobot-record \
   --dataset.repo_id=${HF_USER}/eval_DATASET_NAME_test \  # <- This will be the dataset name on HF Hub
   --dataset.episode_time_s=50 \
   --dataset.num_episodes=10 \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

docs/source/streaming_video_encoding.mdx

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

docs/source/unitree_g1.mdx

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

docs/source/using_dataset_tools.mdx

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

docs/source/walloss.mdx

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

docs/source/xvla.mdx

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

examples/backward_compatibility/replay.py

@@ -22,7 +22,7 @@ lerobot-replay \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=black \
-    --dataset.repo_id=aliberts/record-test \
+    --dataset.repo_id=<USER>/record-test \
     --dataset.episode=2
 ```
 """

examples/rac/hil_data_collection.py

@@ -0,0 +1,351 @@
+#!/usr/bin/env python
+"""
+Human-in-the-Loop (HIL) Data Collection with Policy Rollout.
+
+Implements the RaC paradigm (Hu et al., 2025) for LeRobot with standard synchronous
+inference. For large models with high inference latency, use hil_data_collection_rtc.py.
+
+The workflow:
+1. Policy runs autonomously
+2. Press SPACE to pause - robot holds position
+3. Press 'c' to take control - human provides RECOVERY + CORRECTION
+4. Press → to end episode (save and continue to next)
+5. Reset, then do next rollout
+
+Keyboard Controls:
+    SPACE  - Pause policy (robot holds position, no recording)
+    c      - Take control (start correction, recording resumes)
+    →      - End episode (save and continue to next)
+    ←      - Re-record episode
+    ESC    - Stop recording and push dataset to hub
+
+Usage:
+    python examples/rac/hil_data_collection.py \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58760431541 \
+        --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+        --teleop.type=so100_leader \
+        --teleop.port=/dev/tty.usbmodem58760431551 \
+        --policy.path=outputs/train/my_policy/checkpoints/last/pretrained_model \
+        --dataset.repo_id=my_user/hil_dataset \
+        --dataset.single_task="Pick up the cube"
+"""
+
+import logging
+import time
+from dataclasses import dataclass
+from pprint import pformat
+from typing import Any
+
+import torch
+from hil_utils import (
+    HILDatasetConfig,
+    init_keyboard_listener,
+    make_identity_processors,
+    print_controls,
+    reset_loop,
+    teleop_disable_torque,
+    teleop_smooth_move_to,
+)
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.datasets.image_writer import safe_stop_image_writer
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import build_dataset_frame, combine_feature_dicts
+from lerobot.datasets.video_utils import VideoEncodingManager
+from lerobot.policies.factory import make_policy, make_pre_post_processors
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rtc import ActionInterpolator
+from lerobot.policies.utils import make_robot_action
+from lerobot.processor import PolicyProcessorPipeline
+from lerobot.processor.rename_processor import rename_stats
+from lerobot.robots import Robot, RobotConfig, make_robot_from_config
+from lerobot.teleoperators import Teleoperator, TeleoperatorConfig, make_teleoperator_from_config
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.control_utils import is_headless, predict_action
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import get_safe_torch_device, init_logging, log_say
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class HILConfig:
+    robot: RobotConfig
+    teleop: TeleoperatorConfig
+    dataset: HILDatasetConfig
+    policy: PreTrainedConfig | None = None
+    interpolation_multiplier: int = 1  # Control rate multiplier (1=off, 2=2x, 3=3x)
+    display_data: bool = True
+    play_sounds: bool = True
+    resume: bool = False
+    device: str = "cuda"
+
+    def __post_init__(self):
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        if self.policy is None:
+            raise ValueError("policy.path is required")
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        return ["policy"]
+
+
+@safe_stop_image_writer
+def rollout_loop(
+    robot: Robot,
+    teleop: Teleoperator,
+    policy: PreTrainedPolicy,
+    preprocessor: PolicyProcessorPipeline,
+    postprocessor: PolicyProcessorPipeline,
+    dataset: LeRobotDataset,
+    events: dict,
+    cfg: HILConfig,
+):
+    """Rollout loop with standard synchronous inference."""
+    fps = cfg.dataset.fps
+    device = get_safe_torch_device(cfg.device)
+
+    policy.reset()
+    preprocessor.reset()
+    postprocessor.reset()
+
+    frame_buffer = []
+    teleop_disable_torque(teleop)
+
+    was_paused = False
+    waiting_for_takeover = False
+    last_action: dict[str, Any] | None = None
+    robot_action: dict[str, Any] = {}
+    action_keys = sorted(robot.action_features.keys())
+
+    interpolator = ActionInterpolator(multiplier=cfg.interpolation_multiplier)
+    control_interval = interpolator.get_control_interval(fps)
+
+    timestamp = 0
+    start_t = time.perf_counter()
+
+    while timestamp < cfg.dataset.episode_time_s:
+        loop_start = time.perf_counter()
+
+        if events["exit_early"]:
+            events["exit_early"] = False
+            events["policy_paused"] = False
+            events["correction_active"] = False
+            break
+
+        # Transition to paused state
+        if events["policy_paused"] and not was_paused:
+            obs = robot.get_observation()
+            robot_pos = {
+                k: v for k, v in obs.items() if k.endswith(".pos") and k in robot.observation_features
+            }
+            teleop_smooth_move_to(teleop, robot_pos, duration_s=2.0, fps=50)
+            events["start_next_episode"] = False
+            waiting_for_takeover = True
+            was_paused = True
+            interpolator.reset()
+
+        # Takeover
+        if waiting_for_takeover and events["start_next_episode"]:
+            teleop_disable_torque(teleop)
+            events["start_next_episode"] = False
+            events["correction_active"] = True
+            waiting_for_takeover = False
+
+        obs = robot.get_observation()
+        obs_filtered = {k: v for k, v in obs.items() if k in robot.observation_features}
+        obs_frame = build_dataset_frame(dataset.features, obs_filtered, prefix=OBS_STR)
+
+        if events["correction_active"]:
+            robot_action = teleop.get_action()
+            robot.send_action(robot_action)
+            action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+            frame_buffer.append({**obs_frame, **action_frame, "task": cfg.dataset.single_task})
+
+        elif waiting_for_takeover or events["policy_paused"]:
+            if last_action:
+                robot.send_action(last_action)
+
+        else:
+            # Policy execution with optional interpolation
+            if interpolator.needs_new_action():
+                action_values = predict_action(
+                    observation=obs_frame,
+                    policy=policy,
+                    device=device,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    use_amp=policy.config.use_amp,
+                    task=cfg.dataset.single_task,
+                    robot_type=robot.robot_type,
+                )
+                robot_action = make_robot_action(action_values, dataset.features)
+                action_tensor = torch.tensor([robot_action[k] for k in action_keys])
+                interpolator.add(action_tensor)
+
+            interp_action = interpolator.get()
+            if interp_action is not None:
+                robot_action = {k: interp_action[i].item() for i, k in enumerate(action_keys)}
+                robot.send_action(robot_action)
+                last_action = robot_action
+                action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+                frame_buffer.append({**obs_frame, **action_frame, "task": cfg.dataset.single_task})
+
+        if cfg.display_data and robot_action:
+            log_rerun_data(observation=obs_filtered, action=robot_action)
+
+        dt = time.perf_counter() - loop_start
+        if (sleep_time := control_interval - dt) > 0:
+            precise_sleep(sleep_time)
+        timestamp = time.perf_counter() - start_t
+
+    teleop_disable_torque(teleop)
+
+    for frame in frame_buffer:
+        dataset.add_frame(frame)
+
+
+@parser.wrap()
+def hil_collect(cfg: HILConfig) -> LeRobotDataset:
+    """Main HIL data collection function."""
+    init_logging()
+    logger.info(pformat(cfg.__dict__))
+
+    if cfg.display_data:
+        init_rerun(session_name="hil_collection")
+
+    robot = make_robot_from_config(cfg.robot)
+    teleop = make_teleoperator_from_config(cfg.teleop)
+
+    teleop_proc, obs_proc = make_identity_processors()
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_proc,
+            initial_features=create_initial_features(action=robot.action_features),
+            use_videos=cfg.dataset.video,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=obs_proc,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=cfg.dataset.video,
+        ),
+    )
+
+    dataset = None
+    listener = None
+
+    try:
+        if cfg.resume:
+            dataset = LeRobotDataset(
+                cfg.dataset.repo_id,
+                root=cfg.dataset.root,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+            )
+            if hasattr(robot, "cameras") and robot.cameras:
+                dataset.start_image_writer(
+                    num_processes=cfg.dataset.num_image_writer_processes,
+                    num_threads=cfg.dataset.num_image_writer_threads_per_camera * len(robot.cameras),
+                )
+        else:
+            dataset = LeRobotDataset.create(
+                cfg.dataset.repo_id,
+                cfg.dataset.fps,
+                root=cfg.dataset.root,
+                robot_type=robot.name,
+                features=dataset_features,
+                use_videos=cfg.dataset.video,
+                image_writer_processes=cfg.dataset.num_image_writer_processes,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera
+                * len(robot.cameras if hasattr(robot, "cameras") else []),
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+            )
+
+        policy = make_policy(cfg.policy, ds_meta=dataset.meta)
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=rename_stats(dataset.meta.stats, cfg.dataset.rename_map),
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.device},
+                "rename_observations_processor": {"rename_map": cfg.dataset.rename_map},
+            },
+        )
+
+        robot.connect()
+        teleop.connect()
+        listener, events = init_keyboard_listener()
+
+        print_controls(rtc=False)
+        print(f"  Policy: {cfg.policy.pretrained_path}")
+        print(f"  Task: {cfg.dataset.single_task}")
+        print(f"  Interpolation: {cfg.interpolation_multiplier}x\n")
+
+        with VideoEncodingManager(dataset):
+            recorded = 0
+            while recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                log_say(f"Episode {dataset.num_episodes}", cfg.play_sounds)
+
+                rollout_loop(
+                    robot=robot,
+                    teleop=teleop,
+                    policy=policy,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    dataset=dataset,
+                    events=events,
+                    cfg=cfg,
+                )
+
+                if events["rerecord_episode"]:
+                    log_say("Re-recording", cfg.play_sounds)
+                    events["rerecord_episode"] = False
+                    events["exit_early"] = False
+                    dataset.clear_episode_buffer()
+                    continue
+
+                dataset.save_episode()
+                recorded += 1
+
+                if recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                    reset_loop(robot, teleop, events, cfg.dataset.fps)
+
+    finally:
+        log_say("Stop recording", cfg.play_sounds, blocking=True)
+
+        if dataset:
+            dataset.finalize()
+
+        if robot.is_connected:
+            robot.disconnect()
+        if teleop.is_connected:
+            teleop.disconnect()
+
+        if not is_headless() and listener:
+            listener.stop()
+
+        if cfg.dataset.push_to_hub:
+            dataset.push_to_hub(tags=cfg.dataset.tags, private=cfg.dataset.private)
+
+    return dataset
+
+
+def main():
+    from lerobot.utils.import_utils import register_third_party_plugins
+
+    register_third_party_plugins()
+    hil_collect()
+
+
+if __name__ == "__main__":
+    main()

examples/rac/hil_data_collection_rtc.py

@@ -0,0 +1,513 @@
+#!/usr/bin/env python
+"""
+Human-in-the-Loop (HIL) Data Collection with Real-Time Chunking (RTC).
+
+Implements the RaC paradigm (Hu et al., 2025) with RTC for large flow-matching models
+(Pi0, Pi0.5, SmolVLA) that have high inference latency. RTC generates action chunks
+asynchronously in a background thread for smooth robot control.
+
+For fast models (ACT, Diffusion), use hil_data_collection.py instead.
+
+The workflow:
+1. Policy runs autonomously with RTC
+2. Press SPACE to pause - robot holds position
+3. Press 'c' to take control - human provides RECOVERY + CORRECTION
+4. Press → to end episode (save and continue to next)
+5. Reset, then do next rollout
+
+Keyboard Controls:
+    SPACE  - Pause policy (robot holds position, no recording)
+    c      - Take control (start correction, recording resumes)
+    →      - End episode (save and continue to next)
+    ←      - Re-record episode
+    ESC    - Stop recording and push dataset to hub
+
+Usage:
+    python examples/rac/hil_data_collection_rtc.py \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58760431541 \
+        --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+        --teleop.type=so100_leader \
+        --teleop.port=/dev/tty.usbmodem58760431551 \
+        --policy.path=outputs/train/pi0_policy/checkpoints/last/pretrained_model \
+        --dataset.repo_id=my_user/hil_rtc_dataset \
+        --dataset.single_task="Pick up the cube" \
+        --rtc.execution_horizon=20
+"""
+
+import logging
+import math
+import time
+from dataclasses import dataclass, field
+from pprint import pformat
+from threading import Event, Lock, Thread
+from typing import Any
+
+import torch
+from hil_utils import (
+    HILDatasetConfig,
+    init_keyboard_listener,
+    make_identity_processors,
+    print_controls,
+    reset_loop,
+    teleop_disable_torque,
+    teleop_smooth_move_to,
+)
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.datasets.image_writer import safe_stop_image_writer
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import build_dataset_frame, combine_feature_dicts, hw_to_dataset_features
+from lerobot.datasets.video_utils import VideoEncodingManager
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rtc import ActionInterpolator, ActionQueue, LatencyTracker, RTCConfig
+from lerobot.processor import PolicyProcessorPipeline
+from lerobot.processor.rename_processor import rename_stats
+from lerobot.robots import Robot, RobotConfig, make_robot_from_config
+from lerobot.teleoperators import Teleoperator, TeleoperatorConfig, make_teleoperator_from_config
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.control_utils import is_headless
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import init_logging, log_say
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class HILRTCConfig:
+    robot: RobotConfig
+    teleop: TeleoperatorConfig
+    dataset: HILDatasetConfig
+    policy: PreTrainedConfig | None = None
+    rtc: RTCConfig = field(default_factory=lambda: RTCConfig(enabled=True, execution_horizon=20))
+    interpolation_multiplier: int = 2  # Control rate multiplier (1=off, 2=2x, 3=3x)
+    display_data: bool = True
+    play_sounds: bool = True
+    resume: bool = False
+    device: str = "cuda"
+    use_torch_compile: bool = False  # First compile takes minutes, disable for real-time
+
+    def __post_init__(self):
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        if self.policy is None:
+            raise ValueError("policy.path is required")
+        self.rtc.enabled = True
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        return ["policy"]
+
+
+class ThreadSafeRobot:
+    """Thread-safe wrapper for robot operations."""
+
+    def __init__(self, robot: Robot):
+        self._robot = robot
+        self._lock = Lock()
+
+    def get_observation(self) -> dict[str, Any]:
+        with self._lock:
+            return self._robot.get_observation()
+
+    def send_action(self, action: dict) -> None:
+        with self._lock:
+            self._robot.send_action(action)
+
+    @property
+    def observation_features(self) -> dict:
+        return self._robot.observation_features
+
+    @property
+    def action_features(self) -> dict:
+        return self._robot.action_features
+
+    @property
+    def name(self) -> str:
+        return self._robot.name
+
+    @property
+    def robot_type(self) -> str:
+        return self._robot.robot_type
+
+    @property
+    def cameras(self):
+        return getattr(self._robot, "cameras", {})
+
+
+def rtc_inference_thread(
+    policy: PreTrainedPolicy,
+    obs_holder: dict,
+    hw_features: dict,
+    preprocessor: PolicyProcessorPipeline,
+    postprocessor: PolicyProcessorPipeline,
+    queue_holder: dict,
+    shutdown_event: Event,
+    policy_active: Event,
+    cfg: HILRTCConfig,
+):
+    """Background thread for RTC action chunk generation."""
+    latency_tracker = LatencyTracker()
+    time_per_chunk = 1.0 / cfg.dataset.fps
+    threshold = 30
+
+    while not shutdown_event.is_set():
+        if not policy_active.is_set():
+            time.sleep(0.01)
+            continue
+
+        queue = queue_holder.get("queue")
+        obs = obs_holder.get("obs")
+        if queue is None or obs is None:
+            time.sleep(0.01)
+            continue
+
+        if queue.qsize() <= threshold:
+            try:
+                current_time = time.perf_counter()
+                idx_before = queue.get_action_index()
+                prev_actions = queue.get_left_over()
+
+                latency = latency_tracker.max()
+                delay = math.ceil(latency / time_per_chunk) if latency else 0
+
+                obs_batch = build_dataset_frame(hw_features, obs, prefix="observation")
+                for name in obs_batch:
+                    obs_batch[name] = torch.from_numpy(obs_batch[name])
+                    if "image" in name:
+                        obs_batch[name] = obs_batch[name].float() / 255
+                        obs_batch[name] = obs_batch[name].permute(2, 0, 1).contiguous()
+                    obs_batch[name] = obs_batch[name].unsqueeze(0).to(cfg.device)
+
+                obs_batch["task"] = [cfg.dataset.single_task]
+                obs_batch["robot_type"] = obs_holder.get("robot_type", "unknown")
+
+                preprocessed = preprocessor(obs_batch)
+                actions = policy.predict_action_chunk(
+                    preprocessed, inference_delay=delay, prev_chunk_left_over=prev_actions
+                )
+
+                original = actions.squeeze(0).clone()
+                processed = postprocessor(actions).squeeze(0)
+                new_latency = time.perf_counter() - current_time
+                new_delay = math.ceil(new_latency / time_per_chunk)
+                latency_tracker.add(new_latency)
+                queue.merge(original, processed, new_delay, idx_before)
+                logger.debug(f"[RTC] Inference latency={new_latency:.2f}s, queue={queue.qsize()}")
+            except Exception as e:
+                logger.error(f"[RTC] Error: {e}")
+                time.sleep(0.5)
+        else:
+            time.sleep(0.01)
+
+
+@safe_stop_image_writer
+def rollout_loop(
+    robot: ThreadSafeRobot,
+    teleop: Teleoperator,
+    policy: PreTrainedPolicy,
+    preprocessor: PolicyProcessorPipeline,
+    postprocessor: PolicyProcessorPipeline,
+    dataset: LeRobotDataset,
+    events: dict,
+    cfg: HILRTCConfig,
+    queue_holder: dict,
+    obs_holder: dict,
+    policy_active: Event,
+    hw_features: dict,
+):
+    """Rollout loop with RTC for asynchronous inference."""
+    fps = cfg.dataset.fps
+
+    policy.reset()
+    preprocessor.reset()
+    postprocessor.reset()
+
+    frame_buffer = []
+    teleop_disable_torque(teleop)
+
+    was_paused = False
+    waiting_for_takeover = False
+    last_action: dict[str, Any] | None = None
+    action_keys = [k for k in robot.action_features if k.endswith(".pos")]
+
+    interpolator = ActionInterpolator(multiplier=cfg.interpolation_multiplier)
+    control_interval = interpolator.get_control_interval(fps)
+
+    robot_action: dict[str, Any] = {}
+    timestamp = 0
+    start_t = time.perf_counter()
+
+    while timestamp < cfg.dataset.episode_time_s:
+        loop_start = time.perf_counter()
+
+        if events["exit_early"]:
+            events["exit_early"] = False
+            events["policy_paused"] = False
+            events["correction_active"] = False
+            break
+
+        # Transition to paused state
+        if events["policy_paused"] and not was_paused:
+            policy_active.clear()
+            obs = robot.get_observation()
+            robot_pos = {
+                k: v for k, v in obs.items() if k.endswith(".pos") and k in robot.observation_features
+            }
+            teleop_smooth_move_to(teleop, robot_pos, duration_s=2.0, fps=50)
+            events["start_next_episode"] = False
+            waiting_for_takeover = True
+            was_paused = True
+            interpolator.reset()
+
+        # Takeover
+        if waiting_for_takeover and events["start_next_episode"]:
+            teleop_disable_torque(teleop)
+            events["start_next_episode"] = False
+            events["correction_active"] = True
+            waiting_for_takeover = False
+
+        obs = robot.get_observation()
+        obs_filtered = {k: v for k, v in obs.items() if k in robot.observation_features}
+        obs_frame = build_dataset_frame(dataset.features, obs_filtered, prefix=OBS_STR)
+
+        obs_holder["obs"] = obs_filtered
+
+        if events["correction_active"]:
+            robot_action = teleop.get_action()
+            robot.send_action(robot_action)
+            action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+            frame_buffer.append({**obs_frame, **action_frame, "task": cfg.dataset.single_task})
+
+        elif waiting_for_takeover or events["policy_paused"]:
+            if last_action:
+                robot.send_action(last_action)
+
+        else:
+            # Policy execution with RTC
+            if not policy_active.is_set():
+                policy_active.set()
+
+            queue = queue_holder["queue"]
+
+            if interpolator.needs_new_action():
+                new_action = queue.get() if queue else None
+                if new_action is not None:
+                    interpolator.add(new_action.cpu())
+
+            action_tensor = interpolator.get()
+            if action_tensor is not None:
+                robot_action = {
+                    k: action_tensor[i].item() for i, k in enumerate(action_keys) if i < len(action_tensor)
+                }
+                robot.send_action(robot_action)
+                last_action = robot_action
+                action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+                frame_buffer.append({**obs_frame, **action_frame, "task": cfg.dataset.single_task})
+
+        if cfg.display_data and robot_action:
+            log_rerun_data(observation=obs_filtered, action=robot_action)
+
+        dt = time.perf_counter() - loop_start
+        if (sleep_time := control_interval - dt) > 0:
+            precise_sleep(sleep_time)
+        timestamp = time.perf_counter() - start_t
+
+    policy_active.clear()
+    teleop_disable_torque(teleop)
+
+    for frame in frame_buffer:
+        dataset.add_frame(frame)
+
+
+@parser.wrap()
+def hil_rtc_collect(cfg: HILRTCConfig) -> LeRobotDataset:
+    """Main HIL data collection function with RTC."""
+    init_logging()
+    logger.info(pformat(cfg.__dict__))
+
+    if cfg.display_data:
+        init_rerun(session_name="hil_rtc_collection")
+
+    robot_raw = make_robot_from_config(cfg.robot)
+    teleop = make_teleoperator_from_config(cfg.teleop)
+
+    teleop_proc, obs_proc = make_identity_processors()
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_proc,
+            initial_features=create_initial_features(action=robot_raw.action_features),
+            use_videos=cfg.dataset.video,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=obs_proc,
+            initial_features=create_initial_features(observation=robot_raw.observation_features),
+            use_videos=cfg.dataset.video,
+        ),
+    )
+
+    dataset = None
+    listener = None
+    shutdown_event = Event()
+    policy_active = Event()
+    rtc_thread = None
+
+    try:
+        if cfg.resume:
+            dataset = LeRobotDataset(
+                cfg.dataset.repo_id,
+                root=cfg.dataset.root,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+            )
+            if hasattr(robot_raw, "cameras") and robot_raw.cameras:
+                dataset.start_image_writer(
+                    num_processes=cfg.dataset.num_image_writer_processes,
+                    num_threads=cfg.dataset.num_image_writer_threads_per_camera * len(robot_raw.cameras),
+                )
+        else:
+            dataset = LeRobotDataset.create(
+                cfg.dataset.repo_id,
+                cfg.dataset.fps,
+                root=cfg.dataset.root,
+                robot_type=robot_raw.name,
+                features=dataset_features,
+                use_videos=cfg.dataset.video,
+                image_writer_processes=cfg.dataset.num_image_writer_processes,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera
+                * len(robot_raw.cameras if hasattr(robot_raw, "cameras") else []),
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+            )
+
+        # Load policy with RTC
+        policy_class = get_policy_class(cfg.policy.type)
+        policy_config = PreTrainedConfig.from_pretrained(cfg.policy.pretrained_path)
+        if hasattr(policy_config, "compile_model"):
+            policy_config.compile_model = cfg.use_torch_compile
+        policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=policy_config)
+        policy.config.rtc_config = cfg.rtc
+        if hasattr(policy, "init_rtc_processor"):
+            policy.init_rtc_processor()
+        policy = policy.to(cfg.device)
+        policy.eval()
+
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=rename_stats(dataset.meta.stats, cfg.dataset.rename_map),
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.device},
+                "rename_observations_processor": {"rename_map": cfg.dataset.rename_map},
+            },
+        )
+
+        robot_raw.connect()
+        robot = ThreadSafeRobot(robot_raw)
+        teleop.connect()
+        listener, events = init_keyboard_listener()
+
+        queue_holder = {"queue": ActionQueue(cfg.rtc)}
+        obs_holder = {"obs": None, "robot_type": robot.robot_type}
+        hw_features = hw_to_dataset_features(robot_raw.observation_features, "observation")
+
+        rtc_thread = Thread(
+            target=rtc_inference_thread,
+            args=(
+                policy,
+                obs_holder,
+                hw_features,
+                preprocessor,
+                postprocessor,
+                queue_holder,
+                shutdown_event,
+                policy_active,
+                cfg,
+            ),
+            daemon=True,
+        )
+        rtc_thread.start()
+
+        print_controls(rtc=True)
+        print(f"  Policy: {cfg.policy.pretrained_path}")
+        print(f"  Task: {cfg.dataset.single_task}")
+        print(f"  Interpolation: {cfg.interpolation_multiplier}x\n")
+
+        with VideoEncodingManager(dataset):
+            recorded = 0
+            while recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                log_say(f"Episode {dataset.num_episodes}", cfg.play_sounds)
+
+                queue_holder["queue"] = ActionQueue(cfg.rtc)
+
+                rollout_loop(
+                    robot=robot,
+                    teleop=teleop,
+                    policy=policy,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    dataset=dataset,
+                    events=events,
+                    cfg=cfg,
+                    queue_holder=queue_holder,
+                    obs_holder=obs_holder,
+                    policy_active=policy_active,
+                    hw_features=hw_features,
+                )
+
+                if events["rerecord_episode"]:
+                    log_say("Re-recording", cfg.play_sounds)
+                    events["rerecord_episode"] = False
+                    events["exit_early"] = False
+                    dataset.clear_episode_buffer()
+                    continue
+
+                dataset.save_episode()
+                recorded += 1
+
+                if recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                    reset_loop(robot, teleop, events, cfg.dataset.fps)
+
+    finally:
+        log_say("Stop recording", cfg.play_sounds, blocking=True)
+
+        shutdown_event.set()
+        policy_active.clear()
+
+        if rtc_thread and rtc_thread.is_alive():
+            rtc_thread.join(timeout=2.0)
+
+        if dataset:
+            dataset.finalize()
+
+        if robot_raw.is_connected:
+            robot_raw.disconnect()
+        if teleop.is_connected:
+            teleop.disconnect()
+
+        if not is_headless() and listener:
+            listener.stop()
+
+        if cfg.dataset.push_to_hub:
+            dataset.push_to_hub(tags=cfg.dataset.tags, private=cfg.dataset.private)
+
+    return dataset
+
+
+def main():
+    from lerobot.utils.import_utils import register_third_party_plugins
+
+    register_third_party_plugins()
+    hil_rtc_collect()
+
+
+if __name__ == "__main__":
+    main()

examples/rac/hil_utils.py

@@ -0,0 +1,206 @@
+"""Shared utilities for Human-in-the-Loop data collection scripts."""
+
+import logging
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+
+from lerobot.processor import (
+    IdentityProcessorStep,
+    RobotAction,
+    RobotObservation,
+    RobotProcessorPipeline,
+)
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots import Robot
+from lerobot.teleoperators import Teleoperator
+from lerobot.utils.control_utils import is_headless
+from lerobot.utils.robot_utils import precise_sleep
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class HILDatasetConfig:
+    repo_id: str
+    single_task: str
+    root: str | Path | None = None
+    fps: int = 30
+    episode_time_s: float = 120
+    num_episodes: int = 50
+    video: bool = True
+    push_to_hub: bool = True
+    private: bool = False
+    tags: list[str] | None = None
+    num_image_writer_processes: int = 0
+    num_image_writer_threads_per_camera: int = 4
+    video_encoding_batch_size: int = 1
+    rename_map: dict[str, str] = field(default_factory=dict)
+
+
+def teleop_has_motor_control(teleop: Teleoperator) -> bool:
+    """Check if teleoperator has motor control capabilities."""
+    return hasattr(teleop, "bus") and hasattr(teleop.bus, "disable_torque")
+
+
+def teleop_disable_torque(teleop: Teleoperator) -> None:
+    """Disable teleop torque if supported."""
+    if teleop_has_motor_control(teleop):
+        teleop.bus.disable_torque()
+
+
+def teleop_enable_torque(teleop: Teleoperator) -> None:
+    """Enable teleop torque if supported."""
+    if teleop_has_motor_control(teleop):
+        teleop.bus.enable_torque()
+
+
+def teleop_smooth_move_to(teleop: Teleoperator, target_pos: dict, duration_s: float = 2.0, fps: int = 50):
+    """Smoothly move teleop to target position if motor control is available."""
+    if not teleop_has_motor_control(teleop):
+        logger.warning("Teleop does not support motor control - cannot mirror robot position")
+        return
+
+    teleop_enable_torque(teleop)
+    current = teleop.get_action()
+    steps = int(duration_s * fps)
+
+    for step in range(steps + 1):
+        t = step / steps
+        interp = {}
+        for k in current:
+            if k in target_pos:
+                interp[k] = current[k] * (1 - t) + target_pos[k] * t
+            else:
+                interp[k] = current[k]
+        teleop.bus.sync_write("Goal_Position", {k.replace(".pos", ""): v for k, v in interp.items()})
+        time.sleep(1 / fps)
+
+
+def init_keyboard_listener():
+    """Initialize keyboard listener with HIL controls."""
+    events = {
+        "exit_early": False,
+        "rerecord_episode": False,
+        "stop_recording": False,
+        "policy_paused": False,
+        "correction_active": False,
+        "in_reset": False,
+        "start_next_episode": False,
+    }
+
+    if is_headless():
+        logger.warning("Headless environment - keyboard controls unavailable")
+        return None, events
+
+    from pynput import keyboard
+
+    def on_press(key):
+        try:
+            if events["in_reset"]:
+                if key in [keyboard.Key.space, keyboard.Key.right]:
+                    print("\n[HIL] Starting next episode...")
+                    events["start_next_episode"] = True
+                elif hasattr(key, "char") and key.char == "c":
+                    events["start_next_episode"] = True
+                elif key == keyboard.Key.esc:
+                    print("[HIL] ESC - Stop recording, pushing to hub...")
+                    events["stop_recording"] = True
+                    events["start_next_episode"] = True
+            else:
+                if key == keyboard.Key.space:
+                    if not events["policy_paused"] and not events["correction_active"]:
+                        print("\n[HIL] ⏸ PAUSED - Press 'c' to take control")
+                        events["policy_paused"] = True
+                elif hasattr(key, "char") and key.char == "c":
+                    if events["policy_paused"] and not events["correction_active"]:
+                        print("\n[HIL] ▶ Taking control...")
+                        events["start_next_episode"] = True
+                elif key == keyboard.Key.right:
+                    print("[HIL] → End episode")
+                    events["exit_early"] = True
+                elif key == keyboard.Key.left:
+                    print("[HIL] ← Re-record episode")
+                    events["rerecord_episode"] = True
+                    events["exit_early"] = True
+                elif key == keyboard.Key.esc:
+                    print("[HIL] ESC - Stop recording...")
+                    events["stop_recording"] = True
+                    events["exit_early"] = True
+        except Exception as e:
+            print(f"Key error: {e}")
+
+    listener = keyboard.Listener(on_press=on_press)
+    listener.start()
+    return listener, events
+
+
+def make_identity_processors():
+    """Create identity processors for recording."""
+    teleop_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[IdentityProcessorStep()],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+    obs_proc = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[IdentityProcessorStep()],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+    return teleop_proc, obs_proc
+
+
+def reset_loop(robot: Robot, teleop: Teleoperator, events: dict, fps: int):
+    """Reset period where human repositions environment."""
+    print("\n" + "=" * 60)
+    print("  [HIL] RESET")
+    print("=" * 60)
+
+    events["in_reset"] = True
+    events["start_next_episode"] = False
+
+    obs = robot.get_observation()
+    robot_pos = {k: v for k, v in obs.items() if k.endswith(".pos") and k in robot.observation_features}
+    teleop_smooth_move_to(teleop, robot_pos, duration_s=2.0, fps=50)
+
+    print("  Press any key to enable teleoperation")
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        precise_sleep(0.05)
+
+    if events["stop_recording"]:
+        return
+
+    events["start_next_episode"] = False
+    teleop_disable_torque(teleop)
+    print("  Teleop enabled - press any key to start episode")
+
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        loop_start = time.perf_counter()
+        action = teleop.get_action()
+        robot.send_action(action)
+        precise_sleep(1 / fps - (time.perf_counter() - loop_start))
+
+    events["in_reset"] = False
+    events["start_next_episode"] = False
+    events["exit_early"] = False
+    events["policy_paused"] = False
+    events["correction_active"] = False
+
+
+def print_controls(rtc: bool = False):
+    """Print control instructions."""
+    print("\n" + "=" * 60)
+    print("  Human-in-the-Loop Data Collection" + (" (RTC)" if rtc else ""))
+    print("=" * 60)
+    print()
+    print("  Controls:")
+    print("    SPACE  - Pause policy")
+    print("    c      - Take control")
+    print("    →      - End episode")
+    print("    ESC    - Stop and push to hub")
+    print("=" * 60 + "\n")

examples/rtc/eval_dataset.py

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

examples/rtc/eval_with_real_robot.py

@@ -28,7 +28,7 @@ For simulation environments, see eval_with_simulation.py
 Usage:
     # Run RTC with Real robot with RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
@@ -41,7 +41,7 @@ Usage:
 
     # Run RTC with Real robot without RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=false \
         --robot.type=so100_follower \
@@ -53,7 +53,7 @@ Usage:
 
     # Run RTC with Real robot with pi0.5 policy
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/pi05_check_rtc \
+        --policy.path=<USER>/pi05_check_rtc \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
@@ -83,9 +83,7 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import RTCAttentionSchedule
 from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
 from lerobot.policies.factory import get_policy_class, make_pre_post_processors
-from lerobot.policies.rtc.action_queue import ActionQueue
-from lerobot.policies.rtc.configuration_rtc import RTCConfig
-from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.policies.rtc import ActionInterpolator, ActionQueue, LatencyTracker, RTCConfig
 from lerobot.processor.factory import (
     make_default_robot_action_processor,
     make_default_robot_observation_processor,
@@ -151,6 +149,7 @@ class RTCDemoConfig(HubMixin):
     # Demo parameters
     duration: float = 30.0  # Duration to run the demo (seconds)
     fps: float = 10.0  # Action execution frequency (Hz)
+    interpolation_multiplier: int = 1  # Control rate multiplier (1=off, 2=2x, 3=3x)
 
     # Compute device
     device: str | None = None  # Device to run on (cuda, cpu, auto)
@@ -351,20 +350,22 @@ def actor_control(
         logger.info("[ACTOR] Starting actor thread")
 
         action_count = 0
-        action_interval = 1.0 / cfg.fps
+        interpolator = ActionInterpolator(multiplier=cfg.interpolation_multiplier)
+        action_interval = interpolator.get_control_interval(cfg.fps)
 
         while not shutdown_event.is_set():
             start_time = time.perf_counter()
 
-            # Try to get an action from the queue with timeout
-            action = action_queue.get()
+            if interpolator.needs_new_action():
+                new_action = action_queue.get()
+                if new_action is not None:
+                    interpolator.add(new_action.cpu())
 
+            action = interpolator.get()
             if action is not None:
-                action = action.cpu()
                 action_dict = {key: action[i].item() for i, key in enumerate(robot.action_features())}
                 action_processed = robot_action_processor((action_dict, None))
                 robot.send_action(action_processed)
-
                 action_count += 1
 
             dt_s = time.perf_counter() - start_time

pyproject.toml

@@ -59,7 +59,7 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=4.0.0,<4.2.0",
+    "datasets>=4.0.0,<5.0.0",
     "diffusers>=0.27.2,<0.36.0",
     "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
     "accelerate>=1.10.0,<2.0.0",
@@ -76,9 +76,9 @@ dependencies = [
     "pyserial>=3.5,<4.0",
     "wandb>=0.24.0,<0.25.0",
 
-    "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
-    "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
-    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
+    "torch>=2.2.1,<2.11.0", # TODO: Bump dependency
+    "torchcodec>=0.2.1,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bump dependency
+    "torchvision>=0.21.0,<0.26.0", # TODO: Bump dependency
 
     "draccus==0.10.0", # TODO: Remove ==
     "gymnasium>=1.1.1,<2.0.0",
@@ -98,11 +98,13 @@ pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.10.0"]
 transformers-dep = ["transformers>=4.57.1,<5.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
+can-dep = ["python-can>=4.2.0,<5.0.0"]
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
-damiao = ["python-can>=4.2.0,<5.0.0"]
+damiao = ["lerobot[can-dep]"]
+robstride = ["lerobot[can-dep]"]
 
 # Robots
 openarms = ["lerobot[damiao]"]

src/lerobot/cameras/camera.py

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

src/lerobot/cameras/opencv/camera_opencv.py

@@ -530,7 +530,7 @@ class OpenCVCamera(Camera):
         return frame
 
     @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -201,7 +201,7 @@ class Reachy2Camera(Camera):
         return self.read()
 
     @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/cameras/realsense/camera_realsense.py

@@ -573,7 +573,7 @@ class RealSenseCamera(Camera):
 
     # NOTE(Steven): Missing implementation for depth for now
     @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
         """Return the most recent (color) frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/datasets/lerobot_dataset.py

@@ -68,6 +68,7 @@ from lerobot.datasets.utils import (
     write_tasks,
 )
 from lerobot.datasets.video_utils import (
+    StreamingVideoEncoder,
     VideoFrame,
     concatenate_video_files,
     decode_video_frames,
@@ -75,11 +76,11 @@ from lerobot.datasets.video_utils import (
     get_safe_default_codec,
     get_video_duration_in_s,
     get_video_info,
+    resolve_vcodec,
 )
 from lerobot.utils.constants import HF_LEROBOT_HOME
 
 CODEBASE_VERSION = "v3.0"
-VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1"}
 
 
 class LeRobotDatasetMetadata:
@@ -545,12 +546,19 @@ class LeRobotDatasetMetadata:
 
 
 def _encode_video_worker(
-    video_key: str, episode_index: int, root: Path, fps: int, vcodec: str = "libsvtav1"
+    video_key: str,
+    episode_index: int,
+    root: Path,
+    fps: int,
+    vcodec: str = "libsvtav1",
+    encoder_threads: int | None = None,
 ) -> Path:
     temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
     fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
     img_dir = (root / fpath).parent
-    encode_video_frames(img_dir, temp_path, fps, vcodec=vcodec, overwrite=True)
+    encode_video_frames(
+        img_dir, temp_path, fps, vcodec=vcodec, overwrite=True, encoder_threads=encoder_threads
+    )
     shutil.rmtree(img_dir)
     return temp_path
 
@@ -570,6 +578,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
+        streaming_encoding: bool = False,
+        encoder_queue_maxsize: int = 30,
+        encoder_threads: int | None = None,
     ):
         """
         2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -656,7 +667,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             repo_id (str): This is the repo id that will be used to fetch the dataset. Locally, the dataset
                 will be stored under root/repo_id.
             root (Path | None, optional): Local directory to use for downloading/writing files. You can also
-                set the LEROBOT_HOME environment variable to point to a different location. Defaults to
+                set the HF_LEROBOT_HOME environment variable to point to a different location. Defaults to
                 '~/.cache/huggingface/lerobot'.
             episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                 their episode_index in this list. Defaults to None.
@@ -683,12 +694,17 @@ class LeRobotDataset(torch.utils.data.Dataset):
             batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                 Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
             vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
-                'libsvtav1'. Defaults to 'libsvtav1'. Use 'h264' for faster encoding on systems where AV1
-                encoding is CPU-heavy.
+                'libsvtav1', 'auto', or hardware-specific codecs like 'h264_videotoolbox', 'h264_nvenc'.
+                Defaults to 'libsvtav1'. Use 'auto' to auto-detect the best available hardware encoder.
+            streaming_encoding (bool, optional): If True, encode video frames in real-time during capture
+                instead of writing PNG images first. This makes save_episode() near-instant. Defaults to False.
+            encoder_queue_maxsize (int, optional): Maximum number of frames to buffer per camera when using
+                streaming encoding. Defaults to 30 (~1s at 30fps).
+            encoder_threads (int | None, optional): Number of threads per encoder instance. None lets the
+                codec auto-detect (default). Lower values reduce CPU usage per encoder. Maps to 'lp' (via svtav1-params) for
+                libsvtav1 and 'threads' for h264/hevc.
         """
         super().__init__()
-        if vcodec not in VALID_VIDEO_CODECS:
-            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         self.repo_id = repo_id
         self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
         self.image_transforms = image_transforms
@@ -700,7 +716,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.delta_indices = None
         self.batch_encoding_size = batch_encoding_size
         self.episodes_since_last_encoding = 0
-        self.vcodec = vcodec
+        self.vcodec = resolve_vcodec(vcodec)
+        self._encoder_threads = encoder_threads
 
         # Unused attributes
         self.image_writer = None
@@ -708,6 +725,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.writer = None
         self.latest_episode = None
         self._current_file_start_frame = None  # Track the starting frame index of the current parquet file
+        self._streaming_encoder = None
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -749,6 +767,19 @@ class LeRobotDataset(torch.utils.data.Dataset):
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
             self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
 
+        # Initialize streaming encoder for resumed recording
+        if streaming_encoding and len(self.meta.video_keys) > 0:
+            self._streaming_encoder = StreamingVideoEncoder(
+                fps=self.meta.fps,
+                vcodec=self.vcodec,
+                pix_fmt="yuv420p",
+                g=2,
+                crf=30,
+                preset=None,
+                queue_maxsize=encoder_queue_maxsize,
+                encoder_threads=encoder_threads,
+            )
+
     def _close_writer(self) -> None:
         """Close and cleanup the parquet writer if it exists."""
         writer = getattr(self, "writer", None)
@@ -1104,6 +1135,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         self._close_writer()
         self.meta._close_writer()
+        if self._streaming_encoder is not None:
+            self._streaming_encoder.close()
 
     def create_episode_buffer(self, episode_index: int | None = None) -> dict:
         current_ep_idx = self.meta.total_episodes if episode_index is None else episode_index
@@ -1158,6 +1191,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.episode_buffer["timestamp"].append(timestamp)
         self.episode_buffer["task"].append(frame.pop("task"))  # Remove task from frame after processing
 
+        # Start streaming encoder on first frame of episode (once, before iterating keys)
+        if frame_index == 0 and self._streaming_encoder is not None:
+            self._streaming_encoder.start_episode(
+                video_keys=list(self.meta.video_keys),
+                temp_dir=self.root,
+            )
+
         # Add frame features to episode_buffer
         for key in frame:
             if key not in self.features:
@@ -1165,7 +1205,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
                     f"An element of the frame is not in the features. '{key}' not in '{self.features.keys()}'."
                 )
 
-            if self.features[key]["dtype"] in ["image", "video"]:
+            if self.features[key]["dtype"] == "video" and self._streaming_encoder is not None:
+                self._streaming_encoder.feed_frame(key, frame[key])
+                self.episode_buffer[key].append(None)  # Placeholder (video keys are skipped in parquet)
+            elif self.features[key]["dtype"] in ["image", "video"]:
                 img_path = self._get_image_file_path(
                     episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
                 )
@@ -1226,13 +1269,38 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Wait for image writer to end, so that episode stats over images can be computed
         self._wait_image_writer()
-        ep_stats = compute_episode_stats(episode_buffer, self.features)
 
-        ep_metadata = self._save_episode_data(episode_buffer)
         has_video_keys = len(self.meta.video_keys) > 0
+        use_streaming = self._streaming_encoder is not None and has_video_keys
         use_batched_encoding = self.batch_encoding_size > 1
 
-        if has_video_keys and not use_batched_encoding:
+        if use_streaming:
+            # Compute stats for non-video features only (video stats come from encoder)
+            non_video_buffer = {
+                k: v
+                for k, v in episode_buffer.items()
+                if self.features.get(k, {}).get("dtype") not in ("video",)
+            }
+            non_video_features = {k: v for k, v in self.features.items() if v["dtype"] != "video"}
+            ep_stats = compute_episode_stats(non_video_buffer, non_video_features)
+        else:
+            ep_stats = compute_episode_stats(episode_buffer, self.features)
+
+        ep_metadata = self._save_episode_data(episode_buffer)
+
+        if use_streaming:
+            # Finish streaming encoding and collect results
+            streaming_results = self._streaming_encoder.finish_episode()
+            for video_key in self.meta.video_keys:
+                temp_path, video_stats = streaming_results[video_key]
+                if video_stats is not None:
+                    # Format stats same as compute_episode_stats: normalize to [0,1], reshape to (C,1,1)
+                    ep_stats[video_key] = {
+                        k: v if k == "count" else np.squeeze(v.reshape(1, -1, 1, 1) / 255.0, axis=0)
+                        for k, v in video_stats.items()
+                    }
+                ep_metadata.update(self._save_episode_video(video_key, episode_index, temp_path=temp_path))
+        elif has_video_keys and not use_batched_encoding:
             num_cameras = len(self.meta.video_keys)
             if parallel_encoding and num_cameras > 1:
                 # TODO(Steven): Ideally we would like to control the number of threads per encoding such that:
@@ -1246,6 +1314,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                             self.root,
                             self.fps,
                             self.vcodec,
+                            self._encoder_threads,
                         ): video_key
                         for video_key in self.meta.video_keys
                     }
@@ -1514,6 +1583,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return metadata
 
     def clear_episode_buffer(self, delete_images: bool = True) -> None:
+        # Cancel streaming encoder if active
+        if self._streaming_encoder is not None:
+            self._streaming_encoder.cancel_episode()
+
         # Clean up image files for the current episode buffer
         if delete_images:
             # Wait for the async image writer to finish
@@ -1561,7 +1634,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        return _encode_video_worker(video_key, episode_index, self.root, self.fps, self.vcodec)
+        return _encode_video_worker(
+            video_key, episode_index, self.root, self.fps, self.vcodec, self._encoder_threads
+        )
 
     @classmethod
     def create(
@@ -1578,10 +1653,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
+        metadata_buffer_size: int = 10,
+        streaming_encoding: bool = False,
+        encoder_queue_maxsize: int = 30,
+        encoder_threads: int | None = None,
     ) -> "LeRobotDataset":
         """Create a LeRobot Dataset from scratch in order to record data."""
-        if vcodec not in VALID_VIDEO_CODECS:
-            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
+        vcodec = resolve_vcodec(vcodec)
         obj = cls.__new__(cls)
         obj.meta = LeRobotDatasetMetadata.create(
             repo_id=repo_id,
@@ -1590,6 +1668,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             features=features,
             root=root,
             use_videos=use_videos,
+            metadata_buffer_size=metadata_buffer_size,
         )
         obj.repo_id = obj.meta.repo_id
         obj.root = obj.meta.root
@@ -1599,6 +1678,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.batch_encoding_size = batch_encoding_size
         obj.episodes_since_last_encoding = 0
         obj.vcodec = vcodec
+        obj._encoder_threads = encoder_threads
 
         if image_writer_processes or image_writer_threads:
             obj.start_image_writer(image_writer_processes, image_writer_threads)
@@ -1620,6 +1700,22 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj._lazy_loading = False
         obj._recorded_frames = 0
         obj._writer_closed_for_reading = False
+
+        # Initialize streaming encoder
+        if streaming_encoding and len(obj.meta.video_keys) > 0:
+            obj._streaming_encoder = StreamingVideoEncoder(
+                fps=fps,
+                vcodec=vcodec,
+                pix_fmt="yuv420p",
+                g=2,
+                crf=30,
+                preset=None,
+                queue_maxsize=encoder_queue_maxsize,
+                encoder_threads=encoder_threads,
+            )
+        else:
+            obj._streaming_encoder = None
+
         return obj
 
 

src/lerobot/datasets/utils.py

@@ -122,19 +122,9 @@ def load_nested_dataset(
         raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
 
     with SuppressProgressBars():
-        # When no filtering needed, Dataset uses memory-mapped loading for efficiency
-        # PyArrow loads the entire dataset into memory
-        if episodes is None:
-            return Dataset.from_parquet([str(path) for path in paths], features=features)
-
-        arrow_dataset = pa_ds.dataset(paths, format="parquet")
-        filter_expr = pa_ds.field("episode_index").isin(episodes)
-        table = arrow_dataset.to_table(filter=filter_expr)
-
-        if features is not None:
-            table = table.cast(features.arrow_schema)
-
-        return Dataset(table)
+        # We use .from_parquet() memory-mapped loading for efficiency
+        filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
+        return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

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

src/lerobot/datasets/video_utils.py

@@ -13,25 +13,106 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import contextlib
 import glob
 import importlib
 import logging
+import queue
 import shutil
 import tempfile
+import threading
 import warnings
 from dataclasses import dataclass, field
+from fractions import Fraction
 from pathlib import Path
 from threading import Lock
 from typing import Any, ClassVar
 
 import av
 import fsspec
+import numpy as np
 import pyarrow as pa
 import torch
 import torchvision
 from datasets.features.features import register_feature
 from PIL import Image
 
+# List of hardware encoders to probe for auto-selection. Availability depends on the platform and FFmpeg build.
+# Determines the order of preference for auto-selection when vcodec="auto" is used.
+HW_ENCODERS = [
+    "h264_videotoolbox",  # macOS
+    "hevc_videotoolbox",  # macOS
+    "h264_nvenc",  # NVIDIA GPU
+    "hevc_nvenc",  # NVIDIA GPU
+    "h264_vaapi",  # Linux Intel/AMD
+    "h264_qsv",  # Intel Quick Sync
+]
+
+VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1", "auto"} | set(HW_ENCODERS)
+
+
+def _get_codec_options(
+    vcodec: str,
+    g: int | None = 2,
+    crf: int | None = 30,
+    preset: int | None = None,
+) -> dict:
+    """Build codec-specific options dict for video encoding."""
+    options = {}
+
+    # GOP size (keyframe interval) - supported by VideoToolbox and software encoders
+    if g is not None and (vcodec in ("h264_videotoolbox", "hevc_videotoolbox") or vcodec not in HW_ENCODERS):
+        options["g"] = str(g)
+
+    # Quality control (codec-specific parameter names)
+    if crf is not None:
+        if vcodec in ("h264", "hevc", "libsvtav1"):
+            options["crf"] = str(crf)
+        elif vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
+            quality = max(1, min(100, int(100 - crf * 2)))
+            options["q:v"] = str(quality)
+        elif vcodec in ("h264_nvenc", "hevc_nvenc"):
+            options["rc"] = "constqp"
+            options["qp"] = str(crf)
+        elif vcodec in ("h264_vaapi",):
+            options["qp"] = str(crf)
+        elif vcodec in ("h264_qsv",):
+            options["global_quality"] = str(crf)
+
+    # Preset (only for libsvtav1)
+    if vcodec == "libsvtav1":
+        options["preset"] = str(preset) if preset is not None else "12"
+
+    return options
+
+
+def detect_available_hw_encoders() -> list[str]:
+    """Probe PyAV/FFmpeg for available hardware video encoders."""
+    available = []
+    for codec_name in HW_ENCODERS:
+        try:
+            av.codec.Codec(codec_name, "w")
+            available.append(codec_name)
+        except Exception:  # nosec B110
+            pass  # nosec B110
+    return available
+
+
+def resolve_vcodec(vcodec: str) -> str:
+    """Validate vcodec and resolve 'auto' to best available HW encoder, fallback to libsvtav1."""
+    if vcodec not in VALID_VIDEO_CODECS:
+        raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
+    if vcodec != "auto":
+        logging.info(f"Using video codec: {vcodec}")
+        return vcodec
+    available = detect_available_hw_encoders()
+    for encoder in HW_ENCODERS:
+        if encoder in available:
+            logging.info(f"Auto-selected video codec: {encoder}")
+            return encoder
+    logging.info("No hardware encoder available, falling back to software encoder 'libsvtav1'")
+    return "libsvtav1"
+
 
 def get_safe_default_codec():
     if importlib.util.find_spec("torchcodec"):
@@ -309,14 +390,13 @@ def encode_video_frames(
     g: int | None = 2,
     crf: int | None = 30,
     fast_decode: int = 0,
-    log_level: int | None = av.logging.ERROR,
+    log_level: int | None = av.logging.WARNING,
     overwrite: bool = False,
     preset: int | None = None,
+    encoder_threads: int | None = None,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
-    # Check encoder availability
-    if vcodec not in ["h264", "hevc", "libsvtav1"]:
-        raise ValueError(f"Unsupported video codec: {vcodec}. Supported codecs are: h264, hevc, libsvtav1.")
+    vcodec = resolve_vcodec(vcodec)
 
     video_path = Path(video_path)
     imgs_dir = Path(imgs_dir)
@@ -347,21 +427,22 @@ def encode_video_frames(
         width, height = dummy_image.size
 
     # Define video codec options
-    video_options = {}
-
-    if g is not None:
-        video_options["g"] = str(g)
-
-    if crf is not None:
-        video_options["crf"] = str(crf)
+    video_options = _get_codec_options(vcodec, g, crf, preset)
 
     if fast_decode:
         key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
         value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
         video_options[key] = value
 
-    if vcodec == "libsvtav1":
-        video_options["preset"] = str(preset) if preset is not None else "12"
+    if encoder_threads is not None:
+        if vcodec == "libsvtav1":
+            lp_param = f"lp={encoder_threads}"
+            if "svtav1-params" in video_options:
+                video_options["svtav1-params"] += f":{lp_param}"
+            else:
+                video_options["svtav1-params"] = lp_param
+        else:
+            video_options["threads"] = str(encoder_threads)
 
     # Set logging level
     if log_level is not None:
@@ -480,6 +561,348 @@ def concatenate_video_files(
     Path(tmp_concatenate_path).unlink()
 
 
+class _CameraEncoderThread(threading.Thread):
+    """A thread that encodes video frames streamed via a queue into an MP4 file.
+
+    One instance is created per camera per episode. Frames are received as numpy arrays
+    from the main thread, encoded in real-time using PyAV (which releases the GIL during
+    encoding), and written to disk. Stats are computed incrementally using
+    RunningQuantileStats and returned via result_queue.
+    """
+
+    def __init__(
+        self,
+        video_path: Path,
+        fps: int,
+        vcodec: str,
+        pix_fmt: str,
+        g: int | None,
+        crf: int | None,
+        preset: int | None,
+        frame_queue: queue.Queue,
+        result_queue: queue.Queue,
+        stop_event: threading.Event,
+        encoder_threads: int | None = None,
+    ):
+        super().__init__(daemon=True)
+        self.video_path = video_path
+        self.fps = fps
+        self.vcodec = vcodec
+        self.pix_fmt = pix_fmt
+        self.g = g
+        self.crf = crf
+        self.preset = preset
+        self.frame_queue = frame_queue
+        self.result_queue = result_queue
+        self.stop_event = stop_event
+        self.encoder_threads = encoder_threads
+
+    def run(self) -> None:
+        from lerobot.datasets.compute_stats import RunningQuantileStats, auto_downsample_height_width
+
+        container = None
+        output_stream = None
+        stats_tracker = RunningQuantileStats()
+        frame_count = 0
+
+        try:
+            logging.getLogger("libav").setLevel(av.logging.WARNING)
+
+            while True:
+                try:
+                    frame_data = self.frame_queue.get(timeout=1)
+                except queue.Empty:
+                    if self.stop_event.is_set():
+                        break
+                    continue
+
+                if frame_data is None:
+                    # Sentinel: flush and close
+                    break
+
+                # Ensure HWC uint8 numpy array
+                if isinstance(frame_data, np.ndarray):
+                    if frame_data.ndim == 3 and frame_data.shape[0] == 3:
+                        # CHW -> HWC
+                        frame_data = frame_data.transpose(1, 2, 0)
+                    if frame_data.dtype != np.uint8:
+                        frame_data = (frame_data * 255).astype(np.uint8)
+
+                # Open container on first frame (to get width/height)
+                if container is None:
+                    height, width = frame_data.shape[:2]
+                    video_options = _get_codec_options(self.vcodec, self.g, self.crf, self.preset)
+                    if self.encoder_threads is not None:
+                        if self.vcodec == "libsvtav1":
+                            lp_param = f"lp={self.encoder_threads}"
+                            if "svtav1-params" in video_options:
+                                video_options["svtav1-params"] += f":{lp_param}"
+                            else:
+                                video_options["svtav1-params"] = lp_param
+                        else:
+                            video_options["threads"] = str(self.encoder_threads)
+                    Path(self.video_path).parent.mkdir(parents=True, exist_ok=True)
+                    container = av.open(str(self.video_path), "w")
+                    output_stream = container.add_stream(self.vcodec, self.fps, options=video_options)
+                    output_stream.pix_fmt = self.pix_fmt
+                    output_stream.width = width
+                    output_stream.height = height
+                    output_stream.time_base = Fraction(1, self.fps)
+
+                # Encode frame with explicit timestamps
+                pil_img = Image.fromarray(frame_data)
+                video_frame = av.VideoFrame.from_image(pil_img)
+                video_frame.pts = frame_count
+                video_frame.time_base = Fraction(1, self.fps)
+                packet = output_stream.encode(video_frame)
+                if packet:
+                    container.mux(packet)
+
+                # Update stats with downsampled frame (per-channel stats like compute_episode_stats)
+                img_chw = frame_data.transpose(2, 0, 1)  # HWC -> CHW
+                img_downsampled = auto_downsample_height_width(img_chw)
+                # Reshape CHW to (H*W, C) for per-channel stats
+                channels = img_downsampled.shape[0]
+                img_for_stats = img_downsampled.transpose(1, 2, 0).reshape(-1, channels)
+                stats_tracker.update(img_for_stats)
+
+                frame_count += 1
+
+            # Flush encoder
+            if output_stream is not None:
+                packet = output_stream.encode()
+                if packet:
+                    container.mux(packet)
+
+            if container is not None:
+                container.close()
+
+            av.logging.restore_default_callback()
+
+            # Get stats and put on result queue
+            if frame_count >= 2:
+                stats = stats_tracker.get_statistics()
+                self.result_queue.put(("ok", stats))
+            else:
+                self.result_queue.put(("ok", None))
+
+        except Exception as e:
+            logging.error(f"Encoder thread error: {e}")
+            if container is not None:
+                with contextlib.suppress(Exception):
+                    container.close()
+            self.result_queue.put(("error", str(e)))
+
+
+class StreamingVideoEncoder:
+    """Manages per-camera encoder threads for real-time video encoding during recording.
+
+    Instead of writing frames as PNG images and then encoding to MP4 at episode end,
+    this class streams frames directly to encoder threads, eliminating the
+    PNG round-trip and making save_episode() near-instant.
+
+    Uses threading instead of multiprocessing to avoid the overhead of pickling large
+    numpy arrays through multiprocessing.Queue. PyAV's encode() releases the GIL,
+    so encoding runs in parallel with the main recording loop.
+    """
+
+    def __init__(
+        self,
+        fps: int,
+        vcodec: str = "libsvtav1",
+        pix_fmt: str = "yuv420p",
+        g: int | None = 2,
+        crf: int | None = 30,
+        preset: int | None = None,
+        queue_maxsize: int = 30,
+        encoder_threads: int | None = None,
+    ):
+        self.fps = fps
+        self.vcodec = resolve_vcodec(vcodec)
+        self.pix_fmt = pix_fmt
+        self.g = g
+        self.crf = crf
+        self.preset = preset
+        self.queue_maxsize = queue_maxsize
+        self.encoder_threads = encoder_threads
+
+        self._frame_queues: dict[str, queue.Queue] = {}
+        self._result_queues: dict[str, queue.Queue] = {}
+        self._threads: dict[str, _CameraEncoderThread] = {}
+        self._stop_events: dict[str, threading.Event] = {}
+        self._video_paths: dict[str, Path] = {}
+        self._dropped_frames: dict[str, int] = {}
+        self._episode_active = False
+
+    def start_episode(self, video_keys: list[str], temp_dir: Path) -> None:
+        """Start encoder threads for a new episode.
+
+        Args:
+            video_keys: List of video feature keys (e.g. ["observation.images.laptop"])
+            temp_dir: Base directory for temporary MP4 files
+        """
+        if self._episode_active:
+            self.cancel_episode()
+
+        self._dropped_frames.clear()
+
+        for video_key in video_keys:
+            frame_queue: queue.Queue = queue.Queue(maxsize=self.queue_maxsize)
+            result_queue: queue.Queue = queue.Queue(maxsize=1)
+            stop_event = threading.Event()
+
+            temp_video_dir = Path(tempfile.mkdtemp(dir=temp_dir))
+            video_path = temp_video_dir / f"{video_key.replace('/', '_')}_streaming.mp4"
+
+            encoder_thread = _CameraEncoderThread(
+                video_path=video_path,
+                fps=self.fps,
+                vcodec=self.vcodec,
+                pix_fmt=self.pix_fmt,
+                g=self.g,
+                crf=self.crf,
+                preset=self.preset,
+                frame_queue=frame_queue,
+                result_queue=result_queue,
+                stop_event=stop_event,
+                encoder_threads=self.encoder_threads,
+            )
+            encoder_thread.start()
+
+            self._frame_queues[video_key] = frame_queue
+            self._result_queues[video_key] = result_queue
+            self._threads[video_key] = encoder_thread
+            self._stop_events[video_key] = stop_event
+            self._video_paths[video_key] = video_path
+
+        self._episode_active = True
+
+    def feed_frame(self, video_key: str, image: np.ndarray) -> None:
+        """Feed a frame to the encoder for a specific camera.
+
+        A copy of the image is made before enqueueing to prevent race conditions
+        with camera drivers that may reuse buffers. If the encoder queue is full
+        (encoder can't keep up), the frame is dropped with a warning instead of
+        crashing the recording session.
+
+        Args:
+            video_key: The video feature key
+            image: numpy array in (H,W,C) or (C,H,W) format, uint8 or float
+
+        Raises:
+            RuntimeError: If the encoder thread has crashed
+        """
+        if not self._episode_active:
+            raise RuntimeError("No active episode. Call start_episode() first.")
+
+        thread = self._threads[video_key]
+        if not thread.is_alive():
+            # Check for error
+            try:
+                status, msg = self._result_queues[video_key].get_nowait()
+                if status == "error":
+                    raise RuntimeError(f"Encoder thread for {video_key} crashed: {msg}")
+            except queue.Empty:
+                pass
+            raise RuntimeError(f"Encoder thread for {video_key} is not alive")
+
+        try:
+            self._frame_queues[video_key].put(image.copy(), timeout=0.1)
+        except queue.Full:
+            self._dropped_frames[video_key] = self._dropped_frames.get(video_key, 0) + 1
+            count = self._dropped_frames[video_key]
+            # Log periodically to avoid spam (1st, then every 10th)
+            if count == 1 or count % 10 == 0:
+                logging.warning(
+                    f"Encoder queue full for {video_key}, dropped {count} frame(s). "
+                    f"Consider using vcodec='auto' for hardware encoding or increasing encoder_queue_maxsize."
+                )
+
+    def finish_episode(self) -> dict[str, tuple[Path, dict | None]]:
+        """Finish encoding the current episode.
+
+        Sends sentinel values, waits for encoder threads to complete,
+        and collects results.
+
+        Returns:
+            Dict mapping video_key to (mp4_path, stats_dict_or_None)
+        """
+        if not self._episode_active:
+            raise RuntimeError("No active episode to finish.")
+
+        results = {}
+
+        # Report dropped frames
+        for video_key, count in self._dropped_frames.items():
+            if count > 0:
+                logging.warning(f"Episode finished with {count} dropped frame(s) for {video_key}.")
+
+        # Send sentinel to all queues
+        for video_key in self._frame_queues:
+            self._frame_queues[video_key].put(None)
+
+        # Wait for all threads and collect results
+        for video_key in self._threads:
+            self._threads[video_key].join(timeout=120)
+            if self._threads[video_key].is_alive():
+                logging.error(f"Encoder thread for {video_key} did not finish in time")
+                self._stop_events[video_key].set()
+                self._threads[video_key].join(timeout=5)
+                results[video_key] = (self._video_paths[video_key], None)
+                continue
+
+            try:
+                status, data = self._result_queues[video_key].get(timeout=5)
+                if status == "error":
+                    raise RuntimeError(f"Encoder thread for {video_key} failed: {data}")
+                results[video_key] = (self._video_paths[video_key], data)
+            except queue.Empty:
+                logging.error(f"No result from encoder thread for {video_key}")
+                results[video_key] = (self._video_paths[video_key], None)
+
+        self._cleanup()
+        self._episode_active = False
+        return results
+
+    def cancel_episode(self) -> None:
+        """Cancel the current episode, stopping encoder threads and cleaning up."""
+        if not self._episode_active:
+            return
+
+        # Signal all threads to stop
+        for video_key in self._stop_events:
+            self._stop_events[video_key].set()
+
+        # Wait for threads to finish
+        for video_key in self._threads:
+            self._threads[video_key].join(timeout=5)
+
+            # Clean up temp MP4 files
+            video_path = self._video_paths.get(video_key)
+            if video_path is not None and video_path.exists():
+                shutil.rmtree(str(video_path.parent), ignore_errors=True)
+
+        self._cleanup()
+        self._episode_active = False
+
+    def close(self) -> None:
+        """Close the encoder, canceling any in-progress episode."""
+        if self._episode_active:
+            self.cancel_episode()
+
+    def _cleanup(self) -> None:
+        """Clean up queues and thread tracking dicts."""
+        for q in self._frame_queues.values():
+            with contextlib.suppress(Exception):
+                while not q.empty():
+                    q.get_nowait()
+        self._frame_queues.clear()
+        self._result_queues.clear()
+        self._threads.clear()
+        self._stop_events.clear()
+        self._video_paths.clear()
+
+
 @dataclass
 class VideoFrame:
     # TODO(rcadene, lhoestq): move to Hugging Face `datasets` repo
@@ -514,7 +937,7 @@ with warnings.catch_warnings():
 
 def get_audio_info(video_path: Path | str) -> dict:
     # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.ERROR)
+    logging.getLogger("libav").setLevel(av.logging.WARNING)
 
     # Getting audio stream information
     audio_info = {}
@@ -546,7 +969,7 @@ def get_audio_info(video_path: Path | str) -> dict:
 
 def get_video_info(video_path: Path | str) -> dict:
     # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.ERROR)
+    logging.getLogger("libav").setLevel(av.logging.WARNING)
 
     # Getting video stream information
     video_info = {}
@@ -632,8 +1055,15 @@ class VideoEncodingManager:
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
-        # Handle any remaining episodes that haven't been batch encoded
-        if self.dataset.episodes_since_last_encoding > 0:
+        streaming_encoder = getattr(self.dataset, "_streaming_encoder", None)
+
+        if streaming_encoder is not None:
+            # Handle streaming encoder cleanup
+            if exc_type is not None:
+                streaming_encoder.cancel_episode()
+            streaming_encoder.close()
+        elif self.dataset.episodes_since_last_encoding > 0:
+            # Handle any remaining episodes that haven't been batch encoded
             if exc_type is not None:
                 logging.info("Exception occurred. Encoding remaining episodes before exit...")
             else:
@@ -650,8 +1080,8 @@ class VideoEncodingManager:
         # Finalize the dataset to properly close all writers
         self.dataset.finalize()
 
-        # Clean up episode images if recording was interrupted
-        if exc_type is not None:
+        # Clean up episode images if recording was interrupted (only for non-streaming mode)
+        if exc_type is not None and streaming_encoder is None:
             interrupted_episode_index = self.dataset.num_episodes
             for key in self.dataset.meta.video_keys:
                 img_dir = self.dataset._get_image_file_path(
@@ -665,14 +1095,12 @@ class VideoEncodingManager:
 
         # Clean up any remaining images directory if it's empty
         img_dir = self.dataset.root / "images"
-        # Check for any remaining PNG files
-        png_files = list(img_dir.rglob("*.png"))
-        if len(png_files) == 0:
-            # Only remove the images directory if no PNG files remain
-            if img_dir.exists():
+        if img_dir.exists():
+            png_files = list(img_dir.rglob("*.png"))
+            if len(png_files) == 0:
                 shutil.rmtree(img_dir)
                 logging.debug("Cleaned up empty images directory")
-        else:
-            logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
+            else:
+                logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
 
         return False  # Don't suppress the original exception

src/lerobot/motors/robstride/__init__.py

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

src/lerobot/motors/robstride/tables.py

@@ -0,0 +1,120 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Configuration tables for Damiao motors."""
+
+from enum import IntEnum
+
+
+# Motor type definitions
+class MotorType(IntEnum):
+    O0 = 0
+    O1 = 1
+    O2 = 2
+    O3 = 3
+    O4 = 4
+    O5 = 5
+    ELO5 = 6
+    O6 = 7
+
+
+class CommMode(IntEnum):
+    PrivateProtocole = 0
+    CANopen = 1
+    MIT = 2
+
+
+# Control modes
+class ControlMode(IntEnum):
+    MIT = 0
+    POS_VEL = 1
+    VEL = 2
+
+
+# Motor limit parameters [PMAX, VMAX, TMAX]
+# PMAX: Maximum position (rad)
+# VMAX: Maximum velocity (rad/s)
+# TMAX: Maximum torque (N·m)
+MOTOR_LIMIT_PARAMS: dict[MotorType, tuple[float, float, float]] = {
+    MotorType.O0: (12.57, 33, 14),
+    MotorType.O1: (12.57, 44, 17),
+    MotorType.O2: (12.57, 33, 20),
+    MotorType.O3: (12.57, 33, 60),
+    MotorType.O4: (12.57, 33, 120),
+    MotorType.O5: (12.57, 50, 5.5),
+    MotorType.ELO5: (12.57, 50, 6),
+    MotorType.O6: (112.5, 50, 36),
+}
+
+# Motor model names
+MODEL_NAMES = {
+    MotorType.O0: "O0",
+    MotorType.O1: "O1",
+    MotorType.O2: "O2",
+    MotorType.O3: "O3",
+    MotorType.O4: "O4",
+    MotorType.O5: "O5",
+    MotorType.ELO5: "ELO5",
+    MotorType.O6: "O6",
+}
+
+# Motor resolution table (encoder counts per revolution)
+MODEL_RESOLUTION = {
+    "O0": 65536,
+    "O1": 65536,
+    "O2": 65536,
+    "O3": 65536,
+    "O4": 65536,
+    "O5": 65536,
+    "ELO5": 65536,
+    "O6": 65536,
+}
+
+# CAN baudrates supported by Robstride motors
+AVAILABLE_BAUDRATES = [
+    1000000,  # 4: 1 mbps (default)
+]
+DEFAULT_BAUDRATE = 1000000
+
+# Default timeout in milliseconds
+DEFAULT_TIMEOUT_MS = 0  # disabled by default, otherwise 20000 is 1s
+
+
+# Data that should be normalized
+NORMALIZED_DATA = ["Present_Position", "Goal_Position"]
+
+
+# MIT control parameter ranges
+MIT_KP_RANGE = (0.0, 500.0)
+MIT_KD_RANGE = (0.0, 5.0)
+
+# CAN frame command IDs
+CAN_CMD_ENABLE = 0xFC
+CAN_CMD_DISABLE = 0xFD
+CAN_CMD_SET_ZERO = 0xFE
+CAN_CMD_CLEAR_FAULT = 0xFB
+
+
+CAN_CMD_QUERY_PARAM = 0x33
+CAN_CMD_WRITE_PARAM = 0x55
+CAN_CMD_SAVE_PARAM = 0xAA
+
+# CAN ID for parameter operations
+CAN_PARAM_ID = 0x7FF
+
+
+RUNNING_TIMEOUT = 0.001
+PARAM_TIMEOUT = 0.01
+
+STATE_CACHE_TTL_S = 0.02

src/lerobot/policies/rtc/__init__.py

@@ -0,0 +1,29 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Real-Time Chunking (RTC) utilities for action-chunking policies."""
+
+from lerobot.policies.rtc.action_interpolator import ActionInterpolator
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.policies.rtc.modeling_rtc import RTCProcessor
+
+__all__ = [
+    "ActionInterpolator",
+    "ActionQueue",
+    "LatencyTracker",
+    "RTCConfig",
+    "RTCProcessor",
+]

src/lerobot/policies/rtc/action_interpolator.py

@@ -0,0 +1,117 @@
+#!/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.
+
+"""Action interpolation for smoother robot control.
+
+Provides configurable Nx control rate by interpolating between consecutive actions.
+Useful with RTC and action-chunking policies to reduce jerkiness.
+"""
+
+from torch import Tensor
+
+
+class ActionInterpolator:
+    """Interpolates between consecutive actions for smoother control.
+
+    When enabled with multiplier N, produces N actions per policy action
+    by linearly interpolating between the previous and current action.
+
+    Example with multiplier=3:
+        prev_action -> [1/3 interpolated, 2/3 interpolated, current_action]
+
+    This effectively multiplies the control rate for smoother motion.
+
+    Usage:
+        interpolator = ActionInterpolator(multiplier=2)  # 2x control rate
+
+        # In control loop:
+        if interpolator.needs_new_action():
+            new_action = queue.get()
+            if new_action:
+                interpolator.add(new_action.cpu())
+
+        action = interpolator.get()
+        if action:
+            robot.send_action(action)
+    """
+
+    def __init__(self, multiplier: int = 1):
+        """Initialize the interpolator.
+
+        Args:
+            multiplier: Control rate multiplier (1 = no interpolation, 2 = 2x, 3 = 3x, etc.)
+        """
+        if multiplier < 1:
+            raise ValueError(f"multiplier must be >= 1, got {multiplier}")
+        self.multiplier = multiplier
+        self._prev: Tensor | None = None
+        self._buffer: list[Tensor] = []
+        self._idx = 0
+
+    @property
+    def enabled(self) -> bool:
+        """Whether interpolation is active (multiplier > 1)."""
+        return self.multiplier > 1
+
+    def reset(self):
+        """Reset interpolation state (call between episodes)."""
+        self._prev = None
+        self._buffer = []
+        self._idx = 0
+
+    def needs_new_action(self) -> bool:
+        """Check if a new action is needed from the queue."""
+        return self._idx >= len(self._buffer)
+
+    def add(self, action: Tensor) -> None:
+        """Add a new action and compute interpolated sequence.
+
+        Args:
+            action: New action tensor from policy/queue (already on CPU).
+        """
+        if self.multiplier > 1 and self._prev is not None:
+            self._buffer = []
+            for i in range(1, self.multiplier + 1):
+                t = i / self.multiplier
+                interp = self._prev + t * (action - self._prev)
+                self._buffer.append(interp)
+        else:
+            self._buffer = [action]
+        self._prev = action
+        self._idx = 0
+
+    def get(self) -> Tensor | None:
+        """Get the next interpolated action.
+
+        Returns:
+            Next action tensor, or None if buffer is exhausted.
+        """
+        if self._idx >= len(self._buffer):
+            return None
+        action = self._buffer[self._idx]
+        self._idx += 1
+        return action
+
+    def get_control_interval(self, fps: float) -> float:
+        """Get the control interval based on interpolation multiplier.
+
+        Args:
+            fps: Base frames per second.
+
+        Returns:
+            Control interval in seconds (divided by multiplier).
+        """
+        return 1.0 / (fps * self.multiplier)

src/lerobot/policies/sarm/compute_rabc_weights.py

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

src/lerobot/policies/smolvla/configuration_smolvla.py

@@ -85,7 +85,7 @@ class SmolVLAConfig(PreTrainedConfig):
     scheduler_decay_lr: float = 2.5e-6
 
     vlm_model_name: str = "HuggingFaceTB/SmolVLM2-500M-Video-Instruct"  # Select the VLM backbone.
-    load_vlm_weights: bool = False  # Set to True in case of training the expert from scratch. True when init from pretrained SmolVLA weights
+    load_vlm_weights: bool = False  # Set to False in case of training the expert from scratch. True when init from pretrained SmolVLA weights
 
     add_image_special_tokens: bool = False  # Whether to use special image tokens around image features.
 

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -30,7 +30,7 @@ Example of finetuning the smolvla pretrained model (`smolvla_base`):
 ```bash
 lerobot-train \
 --policy.path=lerobot/smolvla_base \
---dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
+--dataset.repo_id=<USER>/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```
@@ -40,7 +40,7 @@ and an action expert.
 ```bash
 lerobot-train \
 --policy.type=smolvla \
---dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
+--dataset.repo_id=<USER>/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```

src/lerobot/processor/__init__.py

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

src/lerobot/processor/gym_action_processor.py

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

src/lerobot/processor/hil_processor.py

@@ -312,6 +312,37 @@ 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/rl/gym_manipulator.py

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

src/lerobot/robots/hope_jr/hope_jr_arm.py

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

src/lerobot/robots/hope_jr/hope_jr_hand.py

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

src/lerobot/robots/koch_follower/koch_follower.py

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

src/lerobot/robots/lekiwi/lekiwi.py

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

src/lerobot/robots/omx_follower/omx_follower.py

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

src/lerobot/robots/openarm_follower/openarm_follower.py

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

src/lerobot/robots/reachy2/robot_reachy2.py

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

src/lerobot/robots/so_follower/so_follower.py

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

src/lerobot/robots/unitree_g1/unitree_g1.py

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

src/lerobot/scripts/lerobot_dataset_viz.py

@@ -47,16 +47,14 @@ local$ rerun lerobot_pusht_episode_0.rrd
 ```
 
 - Visualize data stored on a distant machine through streaming:
-(You need to forward the websocket port to the distant machine, with
-`ssh -L 9087:localhost:9087 username@remote-host`)
 ```
 distant$ lerobot-dataset-viz \
     --repo-id lerobot/pusht \
     --episode-index 0 \
     --mode distant \
-    --ws-port 9087
+    --grpc-port 9876
 
-local$ rerun ws://localhost:9087
+local$ rerun rerun+http://IP:GRPC_PORT/proxy
 ```
 
 """
@@ -75,6 +73,7 @@ import tqdm
 
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.utils.constants import ACTION, DONE, OBS_STATE, REWARD
+from lerobot.utils.utils import init_logging
 
 
 def to_hwc_uint8_numpy(chw_float32_torch: torch.Tensor) -> np.ndarray:
@@ -93,10 +92,11 @@ def visualize_dataset(
     num_workers: int = 0,
     mode: str = "local",
     web_port: int = 9090,
-    ws_port: int = 9087,
+    grpc_port: int = 9876,
     save: bool = False,
     output_dir: Path | None = None,
     display_compressed_images: bool = False,
+    **kwargs,
 ) -> Path | None:
     if save:
         assert output_dir is not None, (
@@ -126,7 +126,9 @@ def visualize_dataset(
     gc.collect()
 
     if mode == "distant":
-        rr.serve_web_viewer(open_browser=False, web_port=web_port)
+        server_uri = rr.serve_grpc(grpc_port=grpc_port)
+        logging.info(f"Connect to a Rerun Server: rerun rerun+http://IP:{grpc_port}/proxy")
+        rr.serve_web_viewer(open_browser=False, web_port=web_port, connect_to=server_uri)
 
     logging.info("Logging to Rerun")
 
@@ -226,7 +228,7 @@ def main():
             "Mode of viewing between 'local' or 'distant'. "
             "'local' requires data to be on a local machine. It spawns a viewer to visualize the data locally. "
             "'distant' creates a server on the distant machine where the data is stored. "
-            "Visualize the data by connecting to the server with `rerun ws://localhost:PORT` on the local machine."
+            "Visualize the data by connecting to the server with `rerun rerun+http://IP:GRPC_PORT/proxy` on the local machine."
         ),
     )
     parser.add_argument(
@@ -238,8 +240,13 @@ def main():
     parser.add_argument(
         "--ws-port",
         type=int,
-        default=9087,
-        help="Web socket port for rerun.io when `--mode distant` is set.",
+        help="deprecated, please use --grpc-port instead.",
+    )
+    parser.add_argument(
+        "--grpc-port",
+        type=int,
+        default=9876,
+        help="gRPC port for rerun.io when `--mode distant` is set.",
     )
     parser.add_argument(
         "--save",
@@ -265,9 +272,7 @@ def main():
 
     parser.add_argument(
         "--display-compressed-images",
-        type=bool,
-        required=True,
-        default=False,
+        action="store_true",
         help="If set, display compressed images in Rerun instead of uncompressed ones.",
     )
 
@@ -277,6 +282,14 @@ def main():
     root = kwargs.pop("root")
     tolerance_s = kwargs.pop("tolerance_s")
 
+    if kwargs["ws_port"] is not None:
+        logging.warning(
+            "--ws-port is deprecated and will be removed in future versions. Please use --grpc-port instead."
+        )
+        logging.warning("Setting grpc_port to ws_port value.")
+        kwargs["grpc_port"] = kwargs.pop("ws_port")
+
+    init_logging()
     logging.info("Loading dataset")
     dataset = LeRobotDataset(repo_id, episodes=[args.episode_index], root=root, tolerance_s=tolerance_s)
 

src/lerobot/scripts/lerobot_edit_dataset.py

@@ -24,94 +24,107 @@ When new_repo_id is specified, creates a new dataset.
 Usage Examples:
 
 Delete episodes 0, 2, and 5 from a dataset:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --operation.type delete_episodes \
         --operation.episode_indices "[0, 2, 5]"
 
 Delete episodes and save to a new dataset:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --new_repo_id lerobot/pusht_filtered \
         --operation.type delete_episodes \
         --operation.episode_indices "[0, 2, 5]"
 
 Split dataset by fractions:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --operation.type split \
         --operation.splits '{"train": 0.8, "val": 0.2}'
 
 Split dataset by episode indices:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --operation.type split \
         --operation.splits '{"train": [0, 1, 2, 3], "val": [4, 5]}'
 
 Split into more than two splits:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --operation.type split \
         --operation.splits '{"train": 0.6, "val": 0.2, "test": 0.2}'
 
 Merge multiple datasets:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht_merged \
         --operation.type merge \
         --operation.repo_ids "['lerobot/pusht_train', 'lerobot/pusht_val']"
 
 Remove camera feature:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --operation.type remove_feature \
         --operation.feature_names "['observation.images.top']"
 
 Modify tasks - set a single task for all episodes (WARNING: modifies in-place):
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --operation.type modify_tasks \
         --operation.new_task "Pick up the cube and place it"
 
 Modify tasks - set different tasks for specific episodes (WARNING: modifies in-place):
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --operation.type modify_tasks \
         --operation.episode_tasks '{"0": "Task A", "1": "Task B", "2": "Task A"}'
 
 Modify tasks - set default task with overrides for specific episodes (WARNING: modifies in-place):
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --operation.type modify_tasks \
         --operation.new_task "Default task" \
         --operation.episode_tasks '{"5": "Special task for episode 5"}'
 
 Convert image dataset to video format and save locally:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht_image \
         --operation.type convert_image_to_video \
         --operation.output_dir /path/to/output/pusht_video
 
 Convert image dataset to video format and save with new repo_id:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht_image \
         --new_repo_id lerobot/pusht_video \
         --operation.type convert_image_to_video
 
 Convert image dataset to video format and push to hub:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --repo_id lerobot/pusht_image \
         --new_repo_id lerobot/pusht_video \
         --operation.type convert_image_to_video \
         --push_to_hub true
 
+Show dataset information:
+    lerobot-edit-dataset \
+        --repo_id lerobot/pusht_image \
+        --operation.type info \
+        --operation.show_features true
+
+Show dataset information without feature details:
+    lerobot-edit-dataset \
+        --repo_id lerobot/pusht_image \
+        --operation.type info \
+        --operation.show_features false
+
 Using JSON config file:
-    python -m lerobot.scripts.lerobot_edit_dataset \
+    lerobot-edit-dataset \
         --config_path path/to/edit_config.json
 """
 
 import abc
 import logging
 import shutil
+import sys
 from dataclasses import dataclass
 from pathlib import Path
 
@@ -184,6 +197,13 @@ class ConvertImageToVideoConfig(OperationConfig):
     max_frames_per_batch: int | None = None
 
 
+@OperationConfig.register_subclass("info")
+@dataclass
+class InfoConfig(OperationConfig):
+    type: str = "info"
+    show_features: bool = False
+
+
 @dataclass
 class EditDatasetConfig:
     repo_id: str
@@ -436,6 +456,49 @@ def handle_convert_image_to_video(cfg: EditDatasetConfig) -> None:
         logging.info("Dataset saved locally (not pushed to hub)")
 
 
+def _get_dataset_size(repo_path):
+    import os
+
+    total = 0
+    with os.scandir(repo_path) as it:
+        for entry in it:
+            if entry.is_file():
+                total += entry.stat().st_size
+            elif entry.is_dir():
+                total += _get_dataset_size(entry.path)
+    return total
+
+
+def handle_info(cfg: EditDatasetConfig):
+    if not isinstance(cfg.operation, InfoConfig):
+        raise ValueError("Operation config must be InfoConfig")
+
+    dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
+    sys.stdout.write(f"======Info {dataset.meta.repo_id}\n")
+    sys.stdout.write(f"Repository ID: {dataset.meta.repo_id} \n")
+    sys.stdout.write(f"Total episode: {dataset.meta.total_episodes} \n")
+    sys.stdout.write(f"Total task: {dataset.meta.total_tasks} \n")
+    sys.stdout.write(f"Total frame(Actual Count): {dataset.meta.total_frames}({len(dataset)}) \n")
+    sys.stdout.write(
+        f"Average frame per episode: {dataset.meta.total_frames / dataset.meta.total_episodes:.1f}\n"
+    )
+    sys.stdout.write(
+        f"Average episode time(sec): {(dataset.meta.total_frames / dataset.meta.total_episodes) / dataset.meta.fps:.1f}\n"
+    )
+    sys.stdout.write(f"FPS: {dataset.meta.fps}\n")
+
+    total_file_size = _get_dataset_size(dataset.root)
+    sys.stdout.write(f"Size: {total_file_size / (1024 * 1024):.1f} MB\n")
+    if cfg.operation.show_features:
+        import json
+
+        feature_dump_str = json.dumps(
+            dataset.meta.features, ensure_ascii=False, indent=4, sort_keys=True, separators=(",", ": ")
+        )
+        sys.stdout.write("Features:\n")
+        sys.stdout.write(f"{feature_dump_str}\n")
+
+
 @parser.wrap()
 def edit_dataset(cfg: EditDatasetConfig) -> None:
     operation_type = cfg.operation.type
@@ -452,6 +515,8 @@ def edit_dataset(cfg: EditDatasetConfig) -> None:
         handle_modify_tasks(cfg)
     elif operation_type == "convert_image_to_video":
         handle_convert_image_to_video(cfg)
+    elif operation_type == "info":
+        handle_info(cfg)
     else:
         available = ", ".join(OperationConfig.get_known_choices())
         raise ValueError(f"Unknown operation: {operation_type}\nAvailable operations: {available}")

src/lerobot/scripts/lerobot_record.py

@@ -26,8 +26,10 @@ lerobot-record \
     --dataset.repo_id=<my_username>/<my_dataset_name> \
     --dataset.num_episodes=2 \
     --dataset.single_task="Grab the cube" \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
     --display_data=true
-    # <- Optional: specify video codec (h264, hevc, libsvtav1). Default is libsvtav1. \
+    # <- Optional: specify video codec (auto, h264, hevc, libsvtav1). Default is libsvtav1. \
     # --dataset.vcodec=h264 \
     # <- Teleop optional if you want to teleoperate to record or in between episodes with a policy \
     # --teleop.type=so100_leader \
@@ -58,7 +60,10 @@ lerobot-record \
   --display_data=true \
   --dataset.repo_id=${HF_USER}/bimanual-so-handover-cube \
   --dataset.num_episodes=25 \
-  --dataset.single_task="Grab and handover the red cube to the other arm"
+  --dataset.single_task="Grab and handover the red cube to the other arm" \
+  --dataset.streaming_encoding=true \
+  # --dataset.vcodec=auto \
+  --dataset.encoder_threads=2
 ```
 """
 
@@ -69,6 +74,8 @@ from pathlib import Path
 from pprint import pformat
 from typing import Any
 
+import torch
+
 from lerobot.cameras import (  # noqa: F401
     CameraConfig,  # noqa: F401
 )
@@ -85,6 +92,7 @@ from lerobot.datasets.utils import build_dataset_frame, combine_feature_dicts
 from lerobot.datasets.video_utils import VideoEncodingManager
 from lerobot.policies.factory import make_policy, make_pre_post_processors
 from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rtc import ActionInterpolator
 from lerobot.policies.utils import make_robot_action
 from lerobot.processor import (
     PolicyAction,
@@ -179,9 +187,19 @@ class DatasetRecordConfig:
     # Number of episodes to record before batch encoding videos
     # Set to 1 for immediate encoding (default behavior), or higher for batched encoding
     video_encoding_batch_size: int = 1
-    # Video codec for encoding videos. Options: 'h264', 'hevc', 'libsvtav1'.
-    # Use 'h264' for faster encoding on systems where AV1 encoding is CPU-heavy.
+    # Video codec for encoding videos. Options: 'h264', 'hevc', 'libsvtav1', 'auto',
+    # or hardware-specific: 'h264_videotoolbox', 'h264_nvenc', 'h264_vaapi', 'h264_qsv'.
+    # Use 'auto' to auto-detect the best available hardware encoder.
     vcodec: str = "libsvtav1"
+    # Enable streaming video encoding: encode frames in real-time during capture instead
+    # of writing PNG images first. Makes save_episode() near-instant. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding
+    streaming_encoding: bool = False
+    # Maximum number of frames to buffer per camera when using streaming encoding.
+    # ~1s buffer at 30fps. Provides backpressure if the encoder can't keep up.
+    encoder_queue_maxsize: int = 30
+    # Number of threads per encoder instance. None = auto (codec default).
+    # Lower values reduce CPU usage, maps to 'lp' (via svtav1-params) for libsvtav1 and 'threads' for h264/hevc..
+    encoder_threads: int | None = None
     # Rename map for the observation to override the image and state keys
     rename_map: dict[str, str] = field(default_factory=dict)
 
@@ -210,6 +228,9 @@ class RecordConfig:
     play_sounds: bool = True
     # Resume recording on an existing dataset.
     resume: bool = False
+    # Action interpolation multiplier for smoother policy control (1=off, 2=2x, 3=3x)
+    # Only applies when using a policy (not teleop)
+    interpolation_multiplier: int = 1
 
     def __post_init__(self):
         # HACK: We parse again the cli args here to get the pretrained path if there was one.
@@ -282,6 +303,7 @@ def record_loop(
     control_time_s: int | None = None,
     single_task: str | None = None,
     display_data: bool = False,
+    interpolator: ActionInterpolator | None = None,
     display_compressed_images: bool = False,
 ):
     if dataset is not None and dataset.fps != fps:
@@ -318,6 +340,14 @@ def record_loop(
         preprocessor.reset()
         postprocessor.reset()
 
+    # Reset interpolator if provided
+    if interpolator is not None:
+        interpolator.reset()
+
+    # Calculate control interval based on interpolation
+    use_interpolation = interpolator is not None and interpolator.enabled and policy is not None
+    control_interval = interpolator.get_control_interval(fps) if interpolator else 1 / fps
+
     timestamp = 0
     start_episode_t = time.perf_counter()
     while timestamp < control_time_s:
@@ -338,24 +368,58 @@ def record_loop(
 
         # Get action from either policy or teleop
         if policy is not None and preprocessor is not None and postprocessor is not None:
-            action_values = predict_action(
-                observation=observation_frame,
-                policy=policy,
-                device=get_safe_torch_device(policy.config.device),
-                preprocessor=preprocessor,
-                postprocessor=postprocessor,
-                use_amp=policy.config.use_amp,
-                task=single_task,
-                robot_type=robot.robot_type,
-            )
+            # With interpolation: only call policy when interpolator needs new action
+            if use_interpolation:
+                # Get action keys from robot
+                action_keys = sorted(robot.action_features)
 
-            act_processed_policy: RobotAction = make_robot_action(action_values, dataset.features)
+                if interpolator.needs_new_action():
+                    action_values = predict_action(
+                        observation=observation_frame,
+                        policy=policy,
+                        device=get_safe_torch_device(policy.config.device),
+                        preprocessor=preprocessor,
+                        postprocessor=postprocessor,
+                        use_amp=policy.config.use_amp,
+                        task=single_task,
+                        robot_type=robot.robot_type,
+                    )
+                    act_processed_policy = make_robot_action(action_values, dataset.features)
+                    robot_action_to_send = robot_action_processor((act_processed_policy, obs))
+
+                    # Convert to tensor for interpolator
+                    action_tensor = torch.tensor([robot_action_to_send[k] for k in action_keys])
+                    interpolator.add(action_tensor)
+
+                # Get interpolated action
+                interp_action = interpolator.get()
+                if interp_action is not None:
+                    robot_action_to_send = {k: interp_action[i].item() for i, k in enumerate(action_keys)}
+                    action_values = robot_action_to_send
+                else:
+                    # No action available yet, skip this iteration
+                    continue
+            else:
+                action_values = predict_action(
+                    observation=observation_frame,
+                    policy=policy,
+                    device=get_safe_torch_device(policy.config.device),
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    use_amp=policy.config.use_amp,
+                    task=single_task,
+                    robot_type=robot.robot_type,
+                )
+                act_processed_policy: RobotAction = make_robot_action(action_values, dataset.features)
+                robot_action_to_send = robot_action_processor((act_processed_policy, obs))
 
         elif policy is None and isinstance(teleop, Teleoperator):
             act = teleop.get_action()
 
             # Applies a pipeline to the raw teleop action, default is IdentityProcessor
             act_processed_teleop = teleop_action_processor((act, obs))
+            action_values = act_processed_teleop
+            robot_action_to_send = robot_action_processor((act_processed_teleop, obs))
 
         elif policy is None and isinstance(teleop, list):
             arm_action = teleop_arm.get_action()
@@ -364,6 +428,8 @@ def record_loop(
             base_action = robot._from_keyboard_to_base_action(keyboard_action)
             act = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
             act_processed_teleop = teleop_action_processor((act, obs))
+            action_values = act_processed_teleop
+            robot_action_to_send = robot_action_processor((act_processed_teleop, obs))
         else:
             logging.info(
                 "No policy or teleoperator provided, skipping action generation."
@@ -372,14 +438,6 @@ def record_loop(
             )
             continue
 
-        # Applies a pipeline to the action, default is IdentityProcessor
-        if policy is not None and act_processed_policy is not None:
-            action_values = act_processed_policy
-            robot_action_to_send = robot_action_processor((act_processed_policy, obs))
-        else:
-            action_values = act_processed_teleop
-            robot_action_to_send = robot_action_processor((act_processed_teleop, obs))
-
         # Send action to robot
         # Action can eventually be clipped using `max_relative_target`,
         # so action actually sent is saved in the dataset. action = postprocessor.process(action)
@@ -398,7 +456,14 @@ def record_loop(
             )
 
         dt_s = time.perf_counter() - start_loop_t
-        precise_sleep(max(1 / fps - dt_s, 0.0))
+
+        sleep_time_s: float = control_interval - dt_s
+        if sleep_time_s < 0:
+            logging.warning(
+                f"Record loop is running slower ({1 / dt_s:.1f} Hz) than the target FPS ({fps} Hz). Dataset frames might be dropped and robot control might be unstable. Common causes are: 1) Camera FPS not keeping up 2) Policy inference taking too long 3) CPU starvation"
+            )
+
+        precise_sleep(max(sleep_time_s, 0.0))
 
         timestamp = time.perf_counter() - start_episode_t
 
@@ -445,6 +510,9 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                 root=cfg.dataset.root,
                 batch_encoding_size=cfg.dataset.video_encoding_batch_size,
                 vcodec=cfg.dataset.vcodec,
+                streaming_encoding=cfg.dataset.streaming_encoding,
+                encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
+                encoder_threads=cfg.dataset.encoder_threads,
             )
 
             if hasattr(robot, "cameras") and len(robot.cameras) > 0:
@@ -467,12 +535,16 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                 image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * len(robot.cameras),
                 batch_encoding_size=cfg.dataset.video_encoding_batch_size,
                 vcodec=cfg.dataset.vcodec,
+                streaming_encoding=cfg.dataset.streaming_encoding,
+                encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
+                encoder_threads=cfg.dataset.encoder_threads,
             )
 
         # Load pretrained policy
         policy = None if cfg.policy is None else make_policy(cfg.policy, ds_meta=dataset.meta)
         preprocessor = None
         postprocessor = None
+        interpolator = None
         if cfg.policy is not None:
             preprocessor, postprocessor = make_pre_post_processors(
                 policy_cfg=cfg.policy,
@@ -483,6 +555,10 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                     "rename_observations_processor": {"rename_map": cfg.dataset.rename_map},
                 },
             )
+            # Create interpolator for smoother policy control
+            if cfg.interpolation_multiplier > 1:
+                interpolator = ActionInterpolator(multiplier=cfg.interpolation_multiplier)
+                logging.info(f"Action interpolation enabled: {cfg.interpolation_multiplier}x control rate")
 
         robot.connect()
         if teleop is not None:
@@ -490,6 +566,11 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
 
         listener, events = init_keyboard_listener()
 
+        if not cfg.dataset.streaming_encoding:
+            logging.info(
+                "Streaming encoding is disabled. If you have capable hardware, consider enabling it for way faster episode saving. --dataset.streaming_encoding=true --dataset.encoder_threads=2 # --dataset.vcodec=auto. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding"
+            )
+
         with VideoEncodingManager(dataset):
             recorded_episodes = 0
             while recorded_episodes < cfg.dataset.num_episodes and not events["stop_recording"]:
@@ -509,6 +590,7 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                     control_time_s=cfg.dataset.episode_time_s,
                     single_task=cfg.dataset.single_task,
                     display_data=cfg.display_data,
+                    interpolator=interpolator,
                     display_compressed_images=display_compressed_images,
                 )
 

src/lerobot/scripts/lerobot_replay.py

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

src/lerobot/scripts/lerobot_setup_can.py

@@ -152,6 +152,7 @@ def test_motor(bus, motor_id: int, timeout: float, use_fd: bool):
     )
     try:
         bus.send(disable_msg)
+        bus.recv(timeout=0.1)  # Clear any pending responses
     except Exception:
         print(f"Error sending message to motor 0x{motor_id:02X}")
 

src/lerobot/utils/robot_utils.py

@@ -16,14 +16,14 @@ import platform
 import time
 
 
-def precise_sleep(seconds: float, spin_threshold: float = 0.010, sleep_margin: float = 0.003):
+def precise_sleep(seconds: float, spin_threshold: float = 0.010, sleep_margin: float = 0.005):
     """
     Wait for `seconds` with better precision than time.sleep alone at the expense of more CPU usage.
 
     Parameters:
       - seconds: duration to wait
       - spin_threshold: if remaining <= spin_threshold -> spin; otherwise sleep (seconds). Default 10ms
-      - sleep_margin: when sleeping leave this much time before deadline to avoid oversleep. Default 3ms
+      - sleep_margin: when sleeping leave this much time before deadline to avoid oversleep. Default 5ms
 
     Note:
         The default parameters are chosen to prioritize timing accuracy over CPU usage for the common 30 FPS use case.

tests/datasets/test_datasets.py

@@ -31,7 +31,6 @@ from lerobot.configs.train import TrainPipelineConfig
 from lerobot.datasets.factory import make_dataset
 from lerobot.datasets.image_writer import image_array_to_pil_image
 from lerobot.datasets.lerobot_dataset import (
-    VALID_VIDEO_CODECS,
     LeRobotDataset,
     MultiLeRobotDataset,
     _encode_video_worker,
@@ -45,6 +44,7 @@ from lerobot.datasets.utils import (
     hf_transform_to_torch,
     hw_to_dataset_features,
 )
+from lerobot.datasets.video_utils import VALID_VIDEO_CODECS
 from lerobot.envs.factory import make_env_config
 from lerobot.policies.factory import make_policy_config
 from lerobot.robots import make_robot_from_config
@@ -393,7 +393,7 @@ def test_tmp_mixed_deletion(tmp_path, empty_lerobot_dataset_factory):
         vid_key: {"dtype": "video", "shape": DUMMY_HWC, "names": ["height", "width", "channels"]},
     }
     ds_mixed = empty_lerobot_dataset_factory(
-        root=tmp_path / "mixed", features=features_mixed, batch_encoding_size=2
+        root=tmp_path / "mixed", features=features_mixed, batch_encoding_size=2, streaming_encoding=False
     )
     ds_mixed.add_frame(
         {
@@ -1450,7 +1450,10 @@ def test_valid_video_codecs_constant():
     assert "h264" in VALID_VIDEO_CODECS
     assert "hevc" in VALID_VIDEO_CODECS
     assert "libsvtav1" in VALID_VIDEO_CODECS
-    assert len(VALID_VIDEO_CODECS) == 3
+    assert "auto" in VALID_VIDEO_CODECS
+    assert "h264_videotoolbox" in VALID_VIDEO_CODECS
+    assert "h264_nvenc" in VALID_VIDEO_CODECS
+    assert len(VALID_VIDEO_CODECS) == 10
 
 
 def test_delta_timestamps_with_episodes_filter(tmp_path, empty_lerobot_dataset_factory):

tests/datasets/test_streaming_video_encoder.py

@@ -0,0 +1,730 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for streaming video encoding and hardware-accelerated encoding."""
+
+import queue
+import threading
+from unittest.mock import patch
+
+import av
+import numpy as np
+import pytest
+
+from lerobot.datasets.video_utils import (
+    VALID_VIDEO_CODECS,
+    StreamingVideoEncoder,
+    _CameraEncoderThread,
+    _get_codec_options,
+    detect_available_hw_encoders,
+    resolve_vcodec,
+)
+from lerobot.utils.constants import OBS_IMAGES
+
+# ─── _get_codec_options tests ───
+
+
+class TestGetCodecOptions:
+    def test_libsvtav1_defaults(self):
+        opts = _get_codec_options("libsvtav1")
+        assert opts["g"] == "2"
+        assert opts["crf"] == "30"
+        assert opts["preset"] == "12"
+
+    def test_libsvtav1_custom_preset(self):
+        opts = _get_codec_options("libsvtav1", preset=8)
+        assert opts["preset"] == "8"
+
+    def test_h264_options(self):
+        opts = _get_codec_options("h264", g=10, crf=23)
+        assert opts["g"] == "10"
+        assert opts["crf"] == "23"
+        assert "preset" not in opts
+
+    def test_videotoolbox_options(self):
+        opts = _get_codec_options("h264_videotoolbox", g=2, crf=30)
+        assert opts["g"] == "2"
+        # CRF 30 maps to quality = max(1, min(100, 100 - 30*2)) = 40
+        assert opts["q:v"] == "40"
+        assert "crf" not in opts
+
+    def test_nvenc_options(self):
+        opts = _get_codec_options("h264_nvenc", g=2, crf=25)
+        assert opts["rc"] == "constqp"
+        assert opts["qp"] == "25"
+        assert "crf" not in opts
+        # NVENC doesn't support g
+        assert "g" not in opts
+
+    def test_vaapi_options(self):
+        opts = _get_codec_options("h264_vaapi", crf=28)
+        assert opts["qp"] == "28"
+
+    def test_qsv_options(self):
+        opts = _get_codec_options("h264_qsv", crf=25)
+        assert opts["global_quality"] == "25"
+
+    def test_no_g_no_crf(self):
+        opts = _get_codec_options("h264", g=None, crf=None)
+        assert "g" not in opts
+        assert "crf" not in opts
+
+
+# ─── HW encoder detection tests ───
+
+
+class TestHWEncoderDetection:
+    def test_detect_available_hw_encoders_returns_list(self):
+        result = detect_available_hw_encoders()
+        assert isinstance(result, list)
+
+    def test_detect_available_hw_encoders_only_valid(self):
+        from lerobot.datasets.video_utils import HW_ENCODERS
+
+        result = detect_available_hw_encoders()
+        for encoder in result:
+            assert encoder in HW_ENCODERS
+
+    def test_resolve_vcodec_passthrough(self):
+        assert resolve_vcodec("libsvtav1") == "libsvtav1"
+        assert resolve_vcodec("h264") == "h264"
+
+    def test_resolve_vcodec_auto_fallback(self):
+        """When no HW encoders are available, auto should fall back to libsvtav1."""
+        with patch("lerobot.datasets.video_utils.detect_available_hw_encoders", return_value=[]):
+            assert resolve_vcodec("auto") == "libsvtav1"
+
+    def test_resolve_vcodec_auto_picks_hw(self):
+        """When a HW encoder is available, auto should pick it."""
+        with patch(
+            "lerobot.datasets.video_utils.detect_available_hw_encoders",
+            return_value=["h264_videotoolbox"],
+        ):
+            assert resolve_vcodec("auto") == "h264_videotoolbox"
+
+    def test_resolve_vcodec_auto_returns_valid(self):
+        """Test that resolve_vcodec('auto') returns a known valid codec."""
+        result = resolve_vcodec("auto")
+        assert result in VALID_VIDEO_CODECS
+
+    def test_hw_encoder_names_accepted_in_validation(self):
+        """Test that HW encoder names pass validation in VALID_VIDEO_CODECS."""
+        assert "auto" in VALID_VIDEO_CODECS
+        assert "h264_videotoolbox" in VALID_VIDEO_CODECS
+        assert "h264_nvenc" in VALID_VIDEO_CODECS
+
+    def test_resolve_vcodec_invalid_raises(self):
+        """Test that resolve_vcodec raises ValueError for invalid codecs."""
+        with pytest.raises(ValueError, match="Invalid vcodec"):
+            resolve_vcodec("not_a_real_codec")
+
+
+# ─── _CameraEncoderThread tests ───
+
+
+class TestCameraEncoderThread:
+    def test_encodes_valid_mp4(self, tmp_path):
+        """Test that the encoder thread creates a valid MP4 file with correct frame count."""
+        num_frames = 30
+        height, width = 64, 96
+        fps = 30
+        video_path = tmp_path / "test_output" / "test.mp4"
+
+        frame_queue: queue.Queue = queue.Queue(maxsize=60)
+        result_queue: queue.Queue = queue.Queue(maxsize=1)
+        stop_event = threading.Event()
+
+        encoder_thread = _CameraEncoderThread(
+            video_path=video_path,
+            fps=fps,
+            vcodec="libsvtav1",
+            pix_fmt="yuv420p",
+            g=2,
+            crf=30,
+            preset=13,
+            frame_queue=frame_queue,
+            result_queue=result_queue,
+            stop_event=stop_event,
+        )
+        encoder_thread.start()
+
+        # Feed frames (HWC uint8)
+        for _ in range(num_frames):
+            frame = np.random.randint(0, 255, (height, width, 3), dtype=np.uint8)
+            frame_queue.put(frame)
+
+        # Send sentinel
+        frame_queue.put(None)
+        encoder_thread.join(timeout=60)
+        assert not encoder_thread.is_alive()
+
+        # Check result
+        status, data = result_queue.get(timeout=5)
+        assert status == "ok"
+        assert data is not None  # Stats should be returned
+        assert "mean" in data
+        assert "std" in data
+        assert "min" in data
+        assert "max" in data
+        assert "count" in data
+
+        # Verify the MP4 file is valid
+        assert video_path.exists()
+        with av.open(str(video_path)) as container:
+            stream = container.streams.video[0]
+            # The frame count should match
+            total_frames = sum(1 for _ in container.decode(stream))
+        assert total_frames == num_frames
+
+    def test_handles_chw_input(self, tmp_path):
+        """Test that CHW format input is handled correctly."""
+        num_frames = 5
+        fps = 30
+        video_path = tmp_path / "test_chw" / "test.mp4"
+
+        frame_queue: queue.Queue = queue.Queue(maxsize=60)
+        result_queue: queue.Queue = queue.Queue(maxsize=1)
+        stop_event = threading.Event()
+
+        encoder_thread = _CameraEncoderThread(
+            video_path=video_path,
+            fps=fps,
+            vcodec="libsvtav1",
+            pix_fmt="yuv420p",
+            g=2,
+            crf=30,
+            preset=13,
+            frame_queue=frame_queue,
+            result_queue=result_queue,
+            stop_event=stop_event,
+        )
+        encoder_thread.start()
+
+        # Feed CHW frames
+        for _ in range(num_frames):
+            frame = np.random.randint(0, 255, (3, 64, 96), dtype=np.uint8)
+            frame_queue.put(frame)
+
+        frame_queue.put(None)
+        encoder_thread.join(timeout=60)
+
+        status, _ = result_queue.get(timeout=5)
+        assert status == "ok"
+        assert video_path.exists()
+
+    def test_stop_event_cancellation(self, tmp_path):
+        """Test that setting the stop event causes the thread to exit."""
+        fps = 30
+        video_path = tmp_path / "test_cancel" / "test.mp4"
+
+        frame_queue: queue.Queue = queue.Queue(maxsize=60)
+        result_queue: queue.Queue = queue.Queue(maxsize=1)
+        stop_event = threading.Event()
+
+        encoder_thread = _CameraEncoderThread(
+            video_path=video_path,
+            fps=fps,
+            vcodec="libsvtav1",
+            pix_fmt="yuv420p",
+            g=2,
+            crf=30,
+            preset=13,
+            frame_queue=frame_queue,
+            result_queue=result_queue,
+            stop_event=stop_event,
+        )
+        encoder_thread.start()
+
+        # Feed a few frames
+        for _ in range(3):
+            frame = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+            frame_queue.put(frame)
+
+        # Signal stop instead of sending sentinel
+        stop_event.set()
+        encoder_thread.join(timeout=10)
+        assert not encoder_thread.is_alive()
+
+
+# ─── StreamingVideoEncoder tests ───
+
+
+class TestStreamingVideoEncoder:
+    def test_single_camera_episode(self, tmp_path):
+        """Test encoding a single camera episode."""
+        encoder = StreamingVideoEncoder(fps=30, vcodec="libsvtav1", pix_fmt="yuv420p", g=2, crf=30, preset=13)
+
+        video_keys = [f"{OBS_IMAGES}.laptop"]
+        encoder.start_episode(video_keys, tmp_path)
+
+        num_frames = 20
+        for _ in range(num_frames):
+            frame = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+            encoder.feed_frame(f"{OBS_IMAGES}.laptop", frame)
+
+        results = encoder.finish_episode()
+        assert f"{OBS_IMAGES}.laptop" in results
+
+        mp4_path, stats = results[f"{OBS_IMAGES}.laptop"]
+        assert mp4_path.exists()
+        assert stats is not None
+
+        # Verify frame count
+        with av.open(str(mp4_path)) as container:
+            stream = container.streams.video[0]
+            total_frames = sum(1 for _ in container.decode(stream))
+        assert total_frames == num_frames
+
+        encoder.close()
+
+    def test_multi_camera_episode(self, tmp_path):
+        """Test encoding multiple cameras simultaneously."""
+        encoder = StreamingVideoEncoder(fps=30, vcodec="libsvtav1", pix_fmt="yuv420p", g=2, crf=30)
+
+        video_keys = [f"{OBS_IMAGES}.laptop", f"{OBS_IMAGES}.phone"]
+        encoder.start_episode(video_keys, tmp_path)
+
+        num_frames = 15
+        for _ in range(num_frames):
+            frame0 = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+            frame1 = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+            encoder.feed_frame(video_keys[0], frame0)
+            encoder.feed_frame(video_keys[1], frame1)
+
+        results = encoder.finish_episode()
+
+        for key in video_keys:
+            assert key in results
+            mp4_path, stats = results[key]
+            assert mp4_path.exists()
+            assert stats is not None
+
+        encoder.close()
+
+    def test_sequential_episodes(self, tmp_path):
+        """Test that multiple sequential episodes work correctly."""
+        encoder = StreamingVideoEncoder(fps=30, vcodec="libsvtav1", pix_fmt="yuv420p", g=2, crf=30)
+        video_keys = [f"{OBS_IMAGES}.cam"]
+
+        for ep in range(3):
+            encoder.start_episode(video_keys, tmp_path)
+            num_frames = 10 + ep * 5
+            for _ in range(num_frames):
+                frame = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+                encoder.feed_frame(f"{OBS_IMAGES}.cam", frame)
+            results = encoder.finish_episode()
+
+            mp4_path, stats = results[f"{OBS_IMAGES}.cam"]
+            assert mp4_path.exists()
+
+            with av.open(str(mp4_path)) as container:
+                stream = container.streams.video[0]
+                total_frames = sum(1 for _ in container.decode(stream))
+            assert total_frames == num_frames
+
+        encoder.close()
+
+    def test_cancel_episode(self, tmp_path):
+        """Test that canceling an episode cleans up properly."""
+        encoder = StreamingVideoEncoder(fps=30, vcodec="libsvtav1", pix_fmt="yuv420p", g=2, crf=30)
+        video_keys = [f"{OBS_IMAGES}.cam"]
+
+        encoder.start_episode(video_keys, tmp_path)
+
+        for _ in range(5):
+            frame = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+            encoder.feed_frame(f"{OBS_IMAGES}.cam", frame)
+
+        encoder.cancel_episode()
+
+        # Should be able to start a new episode after cancel
+        encoder.start_episode(video_keys, tmp_path)
+        for _ in range(5):
+            frame = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+            encoder.feed_frame(f"{OBS_IMAGES}.cam", frame)
+        results = encoder.finish_episode()
+
+        assert f"{OBS_IMAGES}.cam" in results
+        encoder.close()
+
+    def test_feed_without_start_raises(self, tmp_path):
+        """Test that feeding frames without starting an episode raises."""
+        encoder = StreamingVideoEncoder(fps=30, vcodec="libsvtav1", pix_fmt="yuv420p")
+        with pytest.raises(RuntimeError, match="No active episode"):
+            encoder.feed_frame("cam", np.zeros((64, 96, 3), dtype=np.uint8))
+        encoder.close()
+
+    def test_finish_without_start_raises(self, tmp_path):
+        """Test that finishing without starting raises."""
+        encoder = StreamingVideoEncoder(fps=30, vcodec="libsvtav1", pix_fmt="yuv420p")
+        with pytest.raises(RuntimeError, match="No active episode"):
+            encoder.finish_episode()
+        encoder.close()
+
+    def test_close_is_idempotent(self, tmp_path):
+        """Test that close() can be called multiple times safely."""
+        encoder = StreamingVideoEncoder(fps=30, vcodec="libsvtav1", pix_fmt="yuv420p")
+        encoder.close()
+        encoder.close()  # Should not raise
+
+    def test_video_duration_matches_frame_count(self, tmp_path):
+        """Test that encoded video duration matches num_frames / fps."""
+        encoder = StreamingVideoEncoder(fps=30, vcodec="libsvtav1", pix_fmt="yuv420p", g=2, crf=30, preset=13)
+        video_keys = [f"{OBS_IMAGES}.cam"]
+        encoder.start_episode(video_keys, tmp_path)
+
+        num_frames = 90  # 3 seconds at 30fps
+        for _ in range(num_frames):
+            frame = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+            encoder.feed_frame(f"{OBS_IMAGES}.cam", frame)
+
+        results = encoder.finish_episode()
+        mp4_path, _ = results[f"{OBS_IMAGES}.cam"]
+
+        expected_duration = num_frames / 30.0  # 3.0 seconds
+
+        with av.open(str(mp4_path)) as container:
+            stream = container.streams.video[0]
+            total_frames = sum(1 for _ in container.decode(stream))
+            if stream.duration is not None:
+                actual_duration = float(stream.duration * stream.time_base)
+            else:
+                actual_duration = float(container.duration / av.time_base)
+
+        assert total_frames == num_frames
+        # Allow small tolerance for duration due to codec framing
+        assert abs(actual_duration - expected_duration) < 0.5, (
+            f"Video duration {actual_duration:.2f}s != expected {expected_duration:.2f}s"
+        )
+
+        encoder.close()
+
+    def test_multi_camera_start_episode_called_once(self, tmp_path):
+        """Test that with multiple cameras, no frames are lost due to double start_episode."""
+        encoder = StreamingVideoEncoder(fps=30, vcodec="libsvtav1", pix_fmt="yuv420p", g=2, crf=30)
+
+        video_keys = [f"{OBS_IMAGES}.cam1", f"{OBS_IMAGES}.cam2"]
+        encoder.start_episode(video_keys, tmp_path)
+
+        num_frames = 30
+        for _ in range(num_frames):
+            frame0 = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+            frame1 = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+            encoder.feed_frame(video_keys[0], frame0)
+            encoder.feed_frame(video_keys[1], frame1)
+
+        results = encoder.finish_episode()
+
+        # Both cameras should have all frames
+        for key in video_keys:
+            mp4_path, stats = results[key]
+            assert mp4_path.exists()
+            with av.open(str(mp4_path)) as container:
+                stream = container.streams.video[0]
+                total_frames = sum(1 for _ in container.decode(stream))
+            assert total_frames == num_frames, (
+                f"Camera {key}: expected {num_frames} frames, got {total_frames}"
+            )
+
+        encoder.close()
+
+    def test_encoder_threads_passed_to_thread(self, tmp_path):
+        """Test that encoder_threads is stored and passed through to encoder threads."""
+        encoder = StreamingVideoEncoder(
+            fps=30, vcodec="libsvtav1", pix_fmt="yuv420p", g=2, crf=30, encoder_threads=2
+        )
+        assert encoder.encoder_threads == 2
+
+        video_keys = [f"{OBS_IMAGES}.cam"]
+        encoder.start_episode(video_keys, tmp_path)
+
+        # Verify the thread received the encoder_threads value
+        thread = encoder._threads[f"{OBS_IMAGES}.cam"]
+        assert thread.encoder_threads == 2
+
+        # Feed some frames and finish to ensure it works end-to-end
+        num_frames = 10
+        for _ in range(num_frames):
+            frame = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+            encoder.feed_frame(f"{OBS_IMAGES}.cam", frame)
+
+        results = encoder.finish_episode()
+        mp4_path, stats = results[f"{OBS_IMAGES}.cam"]
+        assert mp4_path.exists()
+        assert stats is not None
+
+        with av.open(str(mp4_path)) as container:
+            stream = container.streams.video[0]
+            total_frames = sum(1 for _ in container.decode(stream))
+        assert total_frames == num_frames
+
+        encoder.close()
+
+    def test_encoder_threads_none_by_default(self, tmp_path):
+        """Test that encoder_threads defaults to None (codec auto-detect)."""
+        encoder = StreamingVideoEncoder(fps=30, vcodec="libsvtav1", pix_fmt="yuv420p")
+        assert encoder.encoder_threads is None
+        encoder.close()
+
+    def test_graceful_frame_dropping(self, tmp_path):
+        """Test that full queue drops frames instead of crashing."""
+        encoder = StreamingVideoEncoder(
+            fps=30, vcodec="libsvtav1", pix_fmt="yuv420p", g=2, crf=30, preset=13, queue_maxsize=1
+        )
+        video_keys = [f"{OBS_IMAGES}.cam"]
+        encoder.start_episode(video_keys, tmp_path)
+
+        # Feed many frames quickly - with queue_maxsize=1, some will be dropped
+        num_frames = 50
+        for _ in range(num_frames):
+            frame = np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8)
+            encoder.feed_frame(f"{OBS_IMAGES}.cam", frame)
+
+        # Should not raise - frames are dropped gracefully
+        results = encoder.finish_episode()
+        assert f"{OBS_IMAGES}.cam" in results
+
+        mp4_path, _ = results[f"{OBS_IMAGES}.cam"]
+        assert mp4_path.exists()
+
+        # Some frames should have been dropped (queue was tiny)
+        dropped = encoder._dropped_frames.get(f"{OBS_IMAGES}.cam", 0)
+        # We can't guarantee drops but can verify no crash occurred
+        assert dropped >= 0
+
+        encoder.close()
+
+
+# ─── Integration tests with LeRobotDataset ───
+
+
+class TestStreamingEncoderIntegration:
+    def test_add_frame_save_episode_streaming(self, tmp_path):
+        """Full integration test: add_frame -> save_episode with streaming encoding."""
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+        features = {
+            "observation.images.cam": {
+                "dtype": "video",
+                "shape": (64, 96, 3),
+                "names": ["height", "width", "channels"],
+            },
+            "action": {"dtype": "float32", "shape": (6,), "names": ["j1", "j2", "j3", "j4", "j5", "j6"]},
+        }
+
+        dataset = LeRobotDataset.create(
+            repo_id="test/streaming",
+            fps=30,
+            features=features,
+            root=tmp_path / "streaming_test",
+            use_videos=True,
+            streaming_encoding=True,
+        )
+
+        assert dataset._streaming_encoder is not None
+
+        num_frames = 20
+        for _ in range(num_frames):
+            frame = {
+                "observation.images.cam": np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8),
+                "action": np.random.randn(6).astype(np.float32),
+                "task": "test task",
+            }
+            dataset.add_frame(frame)
+
+        dataset.save_episode()
+
+        # Verify dataset metadata
+        assert dataset.meta.total_episodes == 1
+        assert dataset.meta.total_frames == num_frames
+
+        # Verify stats exist for the video key
+        assert dataset.meta.stats is not None
+        assert "observation.images.cam" in dataset.meta.stats
+        assert "action" in dataset.meta.stats
+
+        dataset.finalize()
+
+    def test_streaming_disabled_creates_pngs(self, tmp_path):
+        """Test that disabling streaming encoding falls back to PNG path."""
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+        features = {
+            "observation.images.cam": {
+                "dtype": "video",
+                "shape": (64, 96, 3),
+                "names": ["height", "width", "channels"],
+            },
+            "action": {"dtype": "float32", "shape": (6,), "names": ["j1", "j2", "j3", "j4", "j5", "j6"]},
+        }
+
+        dataset = LeRobotDataset.create(
+            repo_id="test/no_streaming",
+            fps=30,
+            features=features,
+            root=tmp_path / "no_streaming_test",
+            use_videos=True,
+            streaming_encoding=False,
+        )
+
+        assert dataset._streaming_encoder is None
+
+        num_frames = 5
+        for _ in range(num_frames):
+            frame = {
+                "observation.images.cam": np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8),
+                "action": np.random.randn(6).astype(np.float32),
+                "task": "test task",
+            }
+            dataset.add_frame(frame)
+
+        # With streaming disabled, PNG files should be written
+        images_dir = dataset.root / "images"
+        assert images_dir.exists()
+
+        dataset.save_episode()
+        dataset.finalize()
+
+    def test_multi_episode_streaming(self, tmp_path):
+        """Test recording multiple episodes with streaming encoding."""
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+        features = {
+            "observation.images.cam": {
+                "dtype": "video",
+                "shape": (64, 96, 3),
+                "names": ["height", "width", "channels"],
+            },
+            "action": {"dtype": "float32", "shape": (2,), "names": ["j1", "j2"]},
+        }
+
+        dataset = LeRobotDataset.create(
+            repo_id="test/multi_ep",
+            fps=30,
+            features=features,
+            root=tmp_path / "multi_ep_test",
+            use_videos=True,
+            streaming_encoding=True,
+        )
+
+        for ep in range(3):
+            num_frames = 10 + ep * 5
+            for _ in range(num_frames):
+                frame = {
+                    "observation.images.cam": np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8),
+                    "action": np.random.randn(2).astype(np.float32),
+                    "task": f"task_{ep}",
+                }
+                dataset.add_frame(frame)
+            dataset.save_episode()
+
+        assert dataset.meta.total_episodes == 3
+        assert dataset.meta.total_frames == 10 + 15 + 20
+
+        dataset.finalize()
+
+    def test_clear_episode_buffer_cancels_streaming(self, tmp_path):
+        """Test that clearing episode buffer cancels streaming encoding."""
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+        features = {
+            "observation.images.cam": {
+                "dtype": "video",
+                "shape": (64, 96, 3),
+                "names": ["height", "width", "channels"],
+            },
+            "action": {"dtype": "float32", "shape": (2,), "names": ["j1", "j2"]},
+        }
+
+        dataset = LeRobotDataset.create(
+            repo_id="test/cancel",
+            fps=30,
+            features=features,
+            root=tmp_path / "cancel_test",
+            use_videos=True,
+            streaming_encoding=True,
+        )
+
+        # Add some frames
+        for _ in range(5):
+            frame = {
+                "observation.images.cam": np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8),
+                "action": np.random.randn(2).astype(np.float32),
+                "task": "task",
+            }
+            dataset.add_frame(frame)
+
+        # Cancel and re-record
+        dataset.clear_episode_buffer()
+
+        # Record a new episode
+        for _ in range(10):
+            frame = {
+                "observation.images.cam": np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8),
+                "action": np.random.randn(2).astype(np.float32),
+                "task": "task",
+            }
+            dataset.add_frame(frame)
+        dataset.save_episode()
+
+        assert dataset.meta.total_episodes == 1
+        assert dataset.meta.total_frames == 10
+
+        dataset.finalize()
+
+    def test_multi_camera_add_frame_streaming(self, tmp_path):
+        """Test that start_episode is called once with multiple video keys."""
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+        features = {
+            "observation.images.cam1": {
+                "dtype": "video",
+                "shape": (64, 96, 3),
+                "names": ["height", "width", "channels"],
+            },
+            "observation.images.cam2": {
+                "dtype": "video",
+                "shape": (64, 96, 3),
+                "names": ["height", "width", "channels"],
+            },
+            "action": {"dtype": "float32", "shape": (2,), "names": ["j1", "j2"]},
+        }
+
+        dataset = LeRobotDataset.create(
+            repo_id="test/multi_cam",
+            fps=30,
+            features=features,
+            root=tmp_path / "multi_cam_test",
+            use_videos=True,
+            streaming_encoding=True,
+        )
+
+        num_frames = 15
+        for _ in range(num_frames):
+            frame = {
+                "observation.images.cam1": np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8),
+                "observation.images.cam2": np.random.randint(0, 255, (64, 96, 3), dtype=np.uint8),
+                "action": np.random.randn(2).astype(np.float32),
+                "task": "test task",
+            }
+            dataset.add_frame(frame)
+
+        dataset.save_episode()
+
+        assert dataset.meta.total_episodes == 1
+        assert dataset.meta.total_frames == num_frames
+
+        dataset.finalize()

tests/optim/test_schedulers.py

@@ -11,6 +11,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import torch
+from packaging.version import Version
 from torch.optim.lr_scheduler import LambdaLR
 
 from lerobot.optim.schedulers import (
@@ -38,6 +40,10 @@ def test_diffuser_scheduler(optimizer):
         "last_epoch": 1,
         "lr_lambdas": [None],
     }
+
+    if Version(torch.__version__) >= Version("2.8"):
+        expected_state_dict["_is_initial"] = False
+
     assert scheduler.state_dict() == expected_state_dict
 
 
@@ -56,6 +62,10 @@ def test_vqbet_scheduler(optimizer):
         "last_epoch": 1,
         "lr_lambdas": [None],
     }
+
+    if Version(torch.__version__) >= Version("2.8"):
+        expected_state_dict["_is_initial"] = False
+
     assert scheduler.state_dict() == expected_state_dict
 
 
@@ -76,6 +86,10 @@ def test_cosine_decay_with_warmup_scheduler(optimizer):
         "last_epoch": 1,
         "lr_lambdas": [None],
     }
+
+    if Version(torch.__version__) >= Version("2.8"):
+        expected_state_dict["_is_initial"] = False
+
     assert scheduler.state_dict() == expected_state_dict
 
 

tests/robots/test_reachy2.py

@@ -142,6 +142,7 @@ def _make_reachy2_camera_mock(*args, **kwargs):
     cam.connect = MagicMock()
     cam.disconnect = MagicMock()
     cam.async_read = MagicMock(side_effect=lambda: np.zeros((height, width, 3), dtype=np.uint8))
+    cam.read_latest = MagicMock(side_effect=lambda: np.zeros((height, width, 3), dtype=np.uint8))
     return cam
 
 

tests/scripts/test_edit_dataset_parsing.py

@@ -21,6 +21,7 @@ from lerobot.scripts.lerobot_edit_dataset import (
     ConvertImageToVideoConfig,
     DeleteEpisodesConfig,
     EditDatasetConfig,
+    InfoConfig,
     MergeConfig,
     ModifyTasksConfig,
     OperationConfig,
@@ -46,6 +47,7 @@ class TestOperationTypeParsing:
             ("remove_feature", RemoveFeatureConfig),
             ("modify_tasks", ModifyTasksConfig),
             ("convert_image_to_video", ConvertImageToVideoConfig),
+            ("info", InfoConfig),
         ],
     )
     def test_operation_type_resolves_correct_class(self, type_name, expected_cls):
@@ -63,6 +65,7 @@ class TestOperationTypeParsing:
             ("remove_feature", RemoveFeatureConfig),
             ("modify_tasks", ModifyTasksConfig),
             ("convert_image_to_video", ConvertImageToVideoConfig),
+            ("info", InfoConfig),
         ],
     )
     def test_get_choice_name_roundtrips(self, type_name, expected_cls):