source: hugging face blog: lerobot v0.6.0: imagine, evaluate, improve
level: technical
lerobot v0.6.0 introduces policies that imagine future states before acting. vla-jepa predicts future frames in latent space during training but drops the world model at inference, so there is no extra cost. lingbot-va generates future video and actions together, feeding real observations back to stay grounded. fastwam pairs a video-generation expert with an action expert, learning to dream rollouts but skipping the dreaming at inference. these three approaches help test whether world models improve robot policies.
the release adds a unified reward models api with robometer and topreward. robometer is a pretrained model that scores task progress from video and language instructions without task-specific training. topreward uses an off-the-shelf vlm to compute success likelihood from log-probabilities, requiring no reward weights. both can label datasets with per-frame progress curves for reward-aware training and quality checks. six new simulation benchmarks run through a single cli, covering perturbed tasks, bimanual manipulation, kitchen tasks, long-horizon memory exams, and reasoning challenges.
a new deployment cli, lerobot-rollout, supports dagger-style human corrections. operators can take over when a policy fails, record corrections, and feed them back into training. datasets now support depth sensing, custom video encoding, and automatic language annotation via vlms. data loading is up to twice as fast through parallel decoding and persistent workers. training supports fsdp for models larger than a single gpu and can run on cloud jobs with one extra flag. the model zoo grows with gr00t n1.7, molmoact2, eo-1, multitask dit, and evo1.
why it matters: closing the loop from deployment to data collection to retraining makes it practical to improve robot policies continuously.
source: hugging face blog: lerobot v0.6.0: imagine, evaluate, improve