LeRobot v0.6.0: Imagine, Evaluate, Improve

Source: HuggingFace Date: 2026-07-07 URL: https://huggingface.co/blog/lerobot-release-v060

Summary

Hugging Face’s LeRobot v0.6.0 ships under an “Imagine, Evaluate, Improve” framing: world-model policies that predict future states before acting (VLA-JEPA, LingBot-VA, FastWAM), six new simulation benchmarks unified under a lerobot-eval CLI, and a lerobot-rollout CLI for DAgger-style human-in-the-loop correction data collection. It also adds five new vision-language-action models, end-to-end depth-camera support, faster data loading, and a 40%-lighter base install.

Implications

Primarily outside this project’s core threads (open robotics, not LLM/agent tooling), but touches model capability clocks (open) tangentially — “world models that imagine before acting” is the same predict-then-act pattern showing up in language-agent planning research, and the standardized eval CLI (six benchmarks, one entrypoint) mirrors the harness-consolidation pattern seen in LLM eval tooling. Logged for completeness; unlikely to need follow-up unless a VLA policy family crosses over into general agent planning work.