Observed arrival · 2026-09-24
Gait Labs: a humanoid learning to walk in public
Credibility concern recorded. The source reference remains available for verification and correction.
Gait Labs presents a Unitree G1 learning to walk in MuJoCo, with training runs, evaluations, and artifacts described as publicly recorded.
- For
- Robotics and reinforcement-learning observers
- Worth noticing
- The stated evaluation uses 100 fixed-seed episodes in plain C MuJoCo, separate from the training code.
Field notes
The described pipeline separates training from evaluation: policies trained in Brax are exported and tested in plain C MuJoCo on fixed seeds. The page lists the expected artifacts, including an ONNX policy and rollout binaries, and says each file receives a SHA-256 hash. Its proposed joint-control system depends on a forthcoming token launch; the visible run status remains at generation zero.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue