Hugging Face released LeRobot v0.6.0 on July 7, turning its open robotics toolkit into more of a full learning loop than a model-training library.
The release adds world-model policies that try to imagine future states before acting, a reward-model API for judging whether a robot succeeded, six simulation benchmarks under lerobot-eval, and a new lerobot-rollout CLI for deployment runs that can turn failures into more training data.
It also expands the data and training stack. Hugging Face says v0.6.0 brings depth support, VLM-powered dataset annotation, custom video encoding, up to 2x faster data loading, FSDP for larger training jobs, and cloud training through HF Jobs.
The release is about iteration
The useful shift is not a single model name. It is the shape of the workflow.
Robot-learning projects often stall between offline training and messy deployment. A model can look competent in a benchmark, then fail when the camera angle, lighting, object position, grip, or timing changes. LeRobot v0.6.0 tries to make that loop explicit: train a policy, evaluate it in simulation, deploy it with a CLI, capture failure cases, annotate richer data, and feed the next training pass.
The world-model additions point in the same direction. Hugging Face lists VLA-JEPA, FastWAM, and LingBot-VA as policies that use future-state prediction in different ways. The release also adds more vision-language-action models, including GR00T N1.7, MolmoAct2, EO-1, EVO1, and Multitask DiT.
Those are Hugging Face’s release claims, not independent evidence that one policy is best for real robots. The important product signal is that the toolkit now packages more of the surrounding machinery needed to test those claims.
Benchmarks and rollout are now closer together
LeRobot v0.6.0 also makes evaluation more first-class. The six new simulation benchmarks sit behind one evaluation CLI, while reward models such as Robometer and TOPReward give teams another way to score outcomes.
That matters because robotics evaluation needs more than a leaderboard number. A policy can complete the task while taking an unsafe path, damaging an object, relying on brittle state, or failing when the human demonstration style changes. The release gives teams more places to measure those failures before and after deployment.
The lerobot-rollout CLI is especially practical. If deployment failures can be captured as training material, then the gap between “the robot failed” and “the next dataset is better” gets smaller.
