Hugging Face releases LeRobot v0.7.0: steer a running robot by prompt, stream datasets from a 300 TB open hub
The open-source release adds interactive rollouts, four world action models, four new VLAs, remote inference and evaluation confidence intervals, and removes async inference, so existing projects need a migration pass.
Why it matters
LeRobot is Hugging Face's open-source robot-learning library, and this release pushes it from data plumbing toward language-driven control. With an interactive rollout, an operator can change a running robot's subtask, ask it what it sees, or give it a goal to pursue on its own, using a text-generating policy or an external vision-language planner. Training moves to FSDP2 and HSDP for multi-billion-parameter world action models, and datasets can be streamed instead of downloaded. Breaking changes include removed async inference and FSDP checkpoints from 0.6.x that cannot be resumed, so watch how fast the community migrates.
Key facts
- Released 5 Nov 2026 (v0.7.0)
- Hub holds almost 300 TB of open Physical AI data in LeRobot format
- Four world action models: Cosmos 3, Long-WAM, FLUX 3 Action, LaWAM
- Four new VLAs: G0.5, LingBot-VLA 2.0, DM0.5, Wall-OSS-0.5
- Interactive rollouts: /subtask, /vqa and /autosteer
- Async inference removed; FSDP checkpoints from 0.6.x cannot be resumed