Skip to main content

Motion

--motion path/to/clip.npz plays a retargeted motion clip.

Format​

Clips follow the any4hdmi layout: an NPZ per clip below a dataset directory with a manifest.json. Clips store qpos (and optionally velocities); the loader resamples to the policy rate and computes body poses and velocities with MuJoCo forward kinematics on the robot model. Velocities are interpolated from source-frame finite differences, matching the training pipeline.

  • The motion must be retargeted for the same robot: G1 motions cannot drive an RP1 policy, and serial and parallel-ankle models are distinct robots.
  • Joints are matched by name. The floating base is found by the model's freejoint name or canonical root_* fields.
  • Motion data is stored as float16 by default for compatibility with existing datasets; all observation math runs in float32 or higher. Set motion.storage_dtype: float32 in the policy YAML to keep full precision.

Without a file​

Without --motion, the reference is a static pose at the policy's default joint positions, which is useful for checking that a policy stands.