What Is LeRobot?
LeRobot is HuggingFace's robot learning library: a standardized dataset format, a growing collection of pre-trained models, and training scripts for ACT, diffusion policy, and TD-MPC2 policies — all MIT licensed with a focus on real robot deployment rather than simulation-only research. The key differentiator from prior frameworks is standardization: a dataset collected with LeRobot on any supported robot can be trained on by any LeRobot training script, without custom data loaders or format conversion.
As of early 2025, LeRobot has an active Discord community with over 8,000 members, weekly research updates from the core team, and growing model zoo contributions from research groups worldwide. It has become the de facto standard for open-source robot learning data sharing.
Installation and Hardware Support
- Installation: pip install lerobot. Requires Python 3.10+, PyTorch 2.1+. Optional hardware dependencies installed separately (e.g., pip install lerobot[so100] for SO-100 arm support).
- SO-100 arm ($300): The primary low-cost supported platform. 5-DOF leader+follower teleoperation arm from Koch Robotics. Best starting point for teams new to physical robot learning — low hardware risk, well-documented configuration files.
- ALOHA bimanual: Stanford ALOHA and ALOHA 2 configurations included. 14-DOF total (7 per arm), intended for bimanual manipulation research.
- OpenArm: Supported via community adapter. SVRC OpenArm 7-DOF configuration available through SVRC's LeRobot integration package.
Data Collection Workflow
Data collection produces a standardized HDF5 + MP4 dataset automatically:
- Step 1 — Configure robot: Edit the YAML configuration for your robot (e.g., lerobot/configs/robot/so100.yaml) with your serial port and camera indices.
- Step 2 — Record dataset: python lerobot/scripts/record_dataset.py --robot-path lerobot/configs/robot/so100.yaml --num-episodes 50 --dataset-id my_task_v1. The script handles frame synchronization across cameras, HDF5 episode writing, and real-time episode thumbnail generation.
- Step 3 — Review episodes: python lerobot/scripts/visualize_dataset.py --dataset-id my_task_v1 opens an interactive episode viewer. Flag failed episodes for exclusion before training.
- Step 4 — Push to hub (optional): python lerobot/scripts/push_dataset_to_hub.py --dataset-id my_task_v1 uploads to HuggingFace Hub for sharing or to SVRC's data platform for managed storage.
Policy Training
| Policy | Command | Training Time (200 demos, 1× GPU) | Best For |
|---|---|---|---|
| ACT (Action Chunking with Transformers) | --policy act | ~4 hours | Fast-moving tasks, reliable baseline |
| Diffusion Policy | --policy diffusion | ~12 hours | Precision tasks, multimodal behavior |
| TD-MPC2 | --policy tdmpc | ~8 hours | Tasks with clear reward structure |
ACT training command: python lerobot/scripts/train.py --policy act --dataset-id my_task_v1 --training.num_epochs 100. Monitor eval/success_rate in the TensorBoard log. Target >70% before deploying. Diffusion policy requires more compute but typically achieves better performance on precision tasks with multimodal behavior (e.g., tasks where there are multiple valid approaches to grasping an object).
Model Zoo and Zero-Shot Testing
LeRobot maintains a pre-trained model zoo on HuggingFace Hub. Available as of early 2025: ACT and diffusion policies trained on SO-100 for pick-place, stacking, and insertion tasks; ALOHA policies for bimanual wiping and assembly. These can be used for zero-shot testing on your hardware — a useful sanity check that your robot is calibrated correctly before collecting your own data.
To run a pre-trained policy: python lerobot/scripts/eval.py --pretrained-policy-name-or-path lerobot/act_so100_pick_place --robot-path lerobot/configs/robot/so100.yaml.
Community and Resources
- Discord: 8,000+ members, active #help and #showcase channels. Most questions answered within hours.
- HuggingFace Hub: hf.co/lerobot — growing collection of datasets and models. Filter by robot type or task.
- Weekly updates: The core HuggingFace team posts research updates on HuggingFace blog every 1–2 weeks.
- SVRC integration: SVRC's data platform accepts LeRobot-format datasets for managed storage, annotation, and training pipeline export. Our data collection service delivers datasets in LeRobot format by default.