LeRobot 0.6 closes more of the learning loop

Hugging Face released LeRobot 0.6.0 with policies that predict future states, a reward-model interface, additional simulation benchmarks, and deployment tooling that turns corrections into training data. The release also expands depth sensing, annotation, and cloud training.

Context

The interesting change is the connection between failure and the next experiment. Recording what went wrong only helps if the data can feed an improved policy and a repeatable evaluation. Bringing those stages together can reduce friction, while real hardware remains the final test of whether the improvement transfers.