LeRobot v0.6.0 adds world models, reward APIs, and faster robotics training pipelines

LeRobot v0.6.0, announced by Hugging Face, introduces three world-model policies designed to learn future state prediction during training while avoiding inference overhead. VLA-JEPA uses a compact VLA built on Qwen3-VL-2B to predict future frames in latent space, with the world model removed at inference. LingBot-VA autoregressively predicts future video and actions chunk-by-chunk, optionally saving predicted videos for comparison with real outcomes. FastWAM pairs a ~5B video-generation expert with a compact action expert, learning to generate rollouts during training but denoising actions directly at inference.

The release expands the vision-language-action (VLA) model zoo with six new entries: GR00T N1.7, MolmoAct2, EO-1, Multitask DiT, EVO1, and an upgraded GR00T integration. GR00T N1.7 replaces N1.5 with Cosmos-Reason2-2B and a flow-matching action head, while MolmoAct2 from the Allen Institute for AI is now supported end-to-end with fine-tuning, evaluation, and deployment workflows.

A unified reward models API (lerobot.rewards) debuts with four models—HIL-SERL, SARM, Robometer, and TOPReward—providing success detection and progress estimation for robot policies. Robometer is a pretrained, general-purpose reward model that scores task progress and success from raw observations.

Dataset tools add depth support, an automatic language annotation pipeline, custom video encoding, and up to 2x faster data loading. Benchmarks are unified under lerobot-eval, and a new CLI (lerobot-rollout) supports DAgger-style human-in-the-loop corrections for deployment. Training gains include FSDP support for larger models and cloud training via Hugging Face Jobs.

The update also streamlines installation and reduces codebase complexity, with ready-to-use checkpoints available on the Hugging Face Hub for policies like VLA-JEPA, LingBot-VA, and FastWAM.