LeRobot v0.6.0 Ships World Models, Reward Models, and New VLAs
LeRobot v0.6.0 adds world model policies, new VLAs, a reward models API, six benchmarks, and faster dataset loading for robot learning.
Hugging Face's open-source robotics library LeRobot has released v0.6.0, aimed at closing the full robot learning loop. The update introduces world model policies—VLA-JEPA, LingBot-VA, and FastWAM—that learn to predict future states during training, most offering world-model supervision at little to no extra inference cost. It also expands the VLA lineup with GR00T N1.7, MolmoAct2, EO-1, Multitask DiT, and the compact EVO1, several of which can be fine-tuned or run on a single consumer GPU.
On the evaluation side, a new unified reward models API (lerobot.rewards) adds Robometer, a general-purpose model that scores task progress from raw video and language instructions with no task-specific training, and TOPReward, which turns any capable vision-language model into a zero-shot reward function. Together with six new simulation benchmarks unified under a single lerobot-eval CLI, this gives teams a standard way to measure whether a policy is actually improving.
Deployment gets its own lerobot-rollout CLI supporting DAgger-style human-in-the-loop corrections that convert failures into training data, alongside FSDP training and cloud training via Hugging Face Jobs for scaling beyond a single GPU.
Dataset tooling also improves: configurable video encoding, end-to-end depth sensing support, automatic VLM-based language annotation, and up to 2x faster data loading, all on top of a lighter installation footprint. For robotics engineers, v0.6.0 consolidates policy training, evaluation, and deployment into a more cohesive workflow.