Image: huggingface.co
Major open-source LeRobot release with chunked Dataset v3.0 and integration of NVIDIA GR00T N models.
Key specs
| Highlights | Dataset v3.0; GR00T N integration |
|---|
Technical details
LeRobot v0.4.0 (October 2025) introduced LeRobotDataset v3.0 with chunked episodes, video streaming and unified Parquet metadata to support OXE-scale (>400 GB) datasets, plus lerobot-edit-dataset tools to merge, split and edit datasets. It added LIBERO and Meta-World simulation support, a modular Processor pipeline for data flow between models and robots, and one-command multi-GPU training with Accelerate. New policies include π0/π0.5 and NVIDIA GR00T N1.5, and a plugin system makes third-party robots, cameras and teleoperators pip-installable. It also launched the Hugging Face Robot Learning Course.
| Architecture | Open-source PyTorch robot-learning library; v0.4.0 adds LeRobotDataset v3.0, a Processor pipeline (PolicyProcessorPipeline / RobotProcessorPipeline) and a hardware plugin system. |
|---|---|
| Inputs outputs | Robot datasets (v3.0 chunked episodes, streaming video, Parquet metadata) in; trained policies and robot control out. |
| Embodiments supported | SO-100/101 and other LeRobot arms, Pollen Reachy 2, phone teleoperation; third-party hardware via plugins. |
| License weights | Per-model (Hub checkpoints carry their own licenses). |
| License code | Apache-2.0 |
| Hardware requirements | Multi-GPU training via Accelerate; gymnasium ≥1.0.0 and mujoco ≥3.0.0 for simulation. |
| How to try | github.com/…/lerobot |
| Sources |
