Image: huggingface.co
LeRobot release closing the robot-learning loop with new policies, unified evaluation benchmarks and a leaner install.
Key specs
| Highlights | 6 new sim benchmarks via lerobot-eval; LeLab browser UI |
|---|
Technical details
LeRobot v0.6.0 (July 2026) closes the robot-learning loop with world-model policies (VLA-JEPA, FastWAM, LingBot-VA), new VLAs (GR00T N1.7, MolmoAct2, EO-1, EVO1, Multitask DiT) and a reward-models API (Robometer, TOPReward). Six new simulation benchmarks (LIBERO-plus, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, VLABench) run under one lerobot-eval CLI alongside LIBERO, Meta-World and IsaacLab-Arena. The new lerobot-rollout CLI adds deployment strategies including DAgger-style human corrections, and training gains FSDP and HF Jobs cloud runs. Datasets add depth, VLM-based language annotation and up to ~2x faster loading; LeLab provides a browser UI. Breaking changes include a leaner base install and GR00T N1.5 being replaced by N1.7.
| Architecture | Open-source PyTorch robot-learning library (datasets, policies, reward models, simulation evaluation, deployment CLIs). |
|---|---|
| Inputs outputs | LeRobot datasets (RGB, now depth, language annotations) in; trained policies, evaluations and rollouts out. |
| Embodiments supported | SO-100/101, Koch, OpenArm, reBot, Unitree G1 and others via plugin system. |
| License weights | Per-model (policy checkpoints on the Hub carry their own licenses). |
| License code | Apache-2.0 |
| Hardware requirements | PyTorch 2.7-2.11 (CUDA 12.8 wheels for Linux uv installs); FSDP multi-GPU and HF Jobs cloud training supported. |
| How to try | github.com/…/lerobot |
| Sources |
