LeRobot v0.6.0 Released: Robot Learning Based on World Models

Hugging Face's robot learning library LeRobot v0.6.0 has been updated. The core of this release is the introduction of World Models technology, which lets robots predict the future before acting.

Key updates include the following:

• World Models and VLA Models: Supports policies such as VLA-JEPA and LingBot-VA, which learn by predicting the future, as well as the latest VLA (Vision-Language-Action) model families such as GR00T N1.7 and MolmoAct2.
• Reward Models: Improved training efficiency through the Robometer and TOPReward APIs, which judge whether a robot has succeeded.
• Datasets and Evaluation: Achieved depth data support, an automatic language annotation pipeline, and up to 2x faster data loading speed. It also provides integrated benchmarking functionality via lerobot-eval.
• Training and Deployment: Added FSDP support for large-scale model training, cloud training via HF Jobs, and the lerobot-rollout CLI, which turns deployment errors into training data.