Summary
Hugging Face released LeRobot v0.6.0 on July 7, 2026, adding three world-model policy architectures — VLA-JEPA, FastWAM, and LingBot-VA — plus reward model support, six simulation benchmarks under lerobot-eval, human-correction rollout tooling, and FSDP training with HF Jobs cloud support. The release closes the evaluate-correct-train loop for open embodied AI.
Key Contributions
- VLA-JEPA: Qwen3-VL-2B + JEPA-style world model predicts future frames in latent space during training; world model is dropped at inference so robots benefit from future-prediction supervision at zero runtime cost
- FastWAM: Combines a ~5B-parameter video-generation expert with a smaller action expert in one network; learns full rollouts during training for efficient inference
- LingBot-VA: Autoregressive video-action world-model policy built on the Wan2.2 video-diffusion stack, interleaving future video latent prediction and robot action generation in one sequence
- Reward model support and six new simulation benchmarks (lerobot-eval) enabling closed-loop policy assessment without physical robots
Significance
LeRobot v0.6.0 is the first major open-source robotics toolkit to integrate multiple world-model policy paradigms in one unified framework, making future-imagination-based robot learning accessible to the broader community without proprietary infrastructure.