Hugging Face released LeRobot v0.6.0, introducing three world model policies — VLA-JEPA, LingBot-VA, and FastWAM — that learn to imagine future states during training, alongside five new vision-language-action models including GR00T N1.7 and MolmoAct2. The update adds a unified reward-model API (Robometer and TOPReward) for evaluating task success, six new simulation benchmarks accessible via a single CLI, and an lerobot-rollout deployment tool supporting DAgger-style human-in-the-loop corrections. Additional technical upgrades include depth camera support, automatic VLM-powered dataset annotation, up to 2x faster data loading, and FSDP multi-GPU training with cloud integration through Hugging Face Jobs.
