LeRobot v0.6.0 is a major release for the open-source robotics framework, introducing world model policies (VLA-JEPA, FastWAM, LingBot-VA) that learn to predict the future during training at zero extra inference cost. The release adds five new VLAs including GR00T N1.7, MolmoAct2, EO-1, EVO1, and Multitask DiT. A new unified reward models API ships with Robometer (a 4B general-purpose reward model trained on 1M+ trajectories) and TOPReward (zero-shot VLM-based reward). Six new simulation benchmarks are unified under the lerobot-eval CLI. Training improvements include FSDP multi-GPU support, cloud training via HF Jobs, and up to 2x faster data loading. Datasets gain depth camera support, automatic language annotation via VLMs, and flexible video encoding. A new lerobot-rollout CLI enables DAgger-style human-in-the-loop corrections for iterative policy improvement.
