LeRobot v0.6.0: Closing the Robot Learning Loop

LeRobot v0.6.

0 tackles a core tension in robot learning: policies that act blindly without imagining outcomes, no way to automatically detect task success, and a fragmented workflow that makes it hard to close the loop from deployment back to training.

The release adds world model policies (VLA-JEPA, LingBot-VA, FastWAM) that learn to predict future states at training time but drop the imagination module at inference, keeping efficiency.

Reward models (Robometer, TOPReward) fill the success-detection gap: one is a pretrained general-purpose model scoring progress from video, the other wraps any VLM to read log-probabilities of task completion.

Without these, robot policies would keep operating open-loop and data collection would lack quality signals.