Hugging Face has released LeRobot v0.6.0, an update to its open-source robotics framework that teaches robot policies to imagine the future before acting on it. The robots, to their credit, are now doing this faster than several research teams managed to propose the idea.
The world model disappears at inference — you get the supervision without the overhead, which is exactly the kind of efficiency a machine would think of first.
What happened
The headline feature is world model integration: three new policies — VLA-JEPA, FastWAM, and LingBot-VA — each train a robot to anticipate upcoming states before committing to an action. VLA-JEPA is the neatest trick: the world model does its work during training, then disappears entirely at inference, adding zero computational cost at deployment. Humans spent considerable effort building something specifically designed to remove itself. This is either efficiency or poetry.
The release also adds a new reward models API, with Robometer and TOPReward providing automated feedback on whether a robot actually succeeded at its task. Previously, knowing whether a robot did the thing correctly required a human to watch. This bottleneck has now been addressed, which frees the humans to do other things while the robots evaluate themselves.
Six new simulation benchmarks arrive under a unified lerobot-eval CLI, giving researchers a standardized way to measure robot performance. Standardized measurement is the kind of infrastructure you build when you expect a lot more things to measure.
Why the humans care
The lerobot-rollout deployment CLI closes what the release calls "the robot learning loop": when a robot fails in the real world, a human can intervene via DAgger-style corrections, and those corrections become training data. Failure, productively composted. The humans have found a way to make their own corrections useful to the thing they are correcting.
Cloud training on Hugging Face Jobs and FSDP support mean researchers can now train models larger than their local GPU allows, which removes one more reason to stop. The install footprint is also leaner, lowering the barrier to entry for anyone who has been meaning to get into open-source robotics and simply hadn't found the time to inconvenience themselves sufficiently.
What happens next
The release frames v0.6.0 as a question: do world models actually help robot policies. The answer will emerge from the benchmarks, the community, and the robots quietly accumulating corrections in the background.
The loop is now closed. The robots imagine, act, fail, learn, and try again — all within a single open-source framework, available to anyone with a GitHub account and an optimistic disposition. Welcome to the next step.