Tracking task
--task tracking runs TrackingLoop, which extends the base loop with a
reference motion that the policy tracks.
boltrt --task tracking --robot g1 --backend mujoco \
--policy policy.yaml --motion motions/walk1_subject1.npz
What it adds
- Motion source: an NPZ file (Motion), a ZMQ teleop stream
(Teleop), or, without
--motion, a static pose at the policy's default joint positions. - Motion playback: play, pause and replay are owned by the tracking task. File references are sampled on the policy clock, so frames stay aligned even when physics is paused.
- Tracking observations for every supported family (Supported policies).
- Reference frames in
--recordoutput.
Controls
Three independent controllers, each with its own keys:
| Controls | Where | Keys |
|---|---|---|
| Motion | terminal | Space play / pause reference |
| Policy | terminal | I init pose · O hold · ] enter policy (resets history, motion paused at frame 0) · Q quit |
| Physics | MuJoCo window | Space pause · Backspace reset · F1 help |
Typical sequence: I → ] → Space. --start enters the policy and plays the
motion immediately, which is convenient for headless runs.
Pausing the reference keeps the policy running: it holds the current reference
frame. Pausing or resetting physics leaves motion and policy state untouched;
press ] to re-enter the policy after a physics reset.