Skip to main content
Reference motions are uniform-rate, both-arm joint trajectories committed to the repository (almond_axol/tuning/motions/), so the exact same motion can be replayed on any robot — today or on future units — and the tracking metrics compared 1:1 with tune.motion. motion.build postprocesses a flight-recorder capture into one, from either recorder:
  • Teleop (axol teleop --teleop.record PREFIX): the final guarded command stream (_cmd stage), clipped to the engaged span.
  • Gravity comp (axol gravity-comp --record PREFIX): the measured arm-joint positions of the hand-guided session (_gc stage). There is no engage state, so the still lead-in/lead-out is trimmed instead — start and end the demonstration with the arms briefly at rest.
Either way, the stream is resampled onto a uniform grid, zero-phase low-pass smoothed (keeping the operator’s intent, dropping hand tremor and network jitter), and projected waypoint-by-waypoint through the same collision-aware solver the teleop return-to-rest uses — the stored motion is joint-limit- and self-collision-safe by construction.
Commit the generated .npz so every robot can replay the identical motion. Every build also saves a before/after tuning-run artifact (kind=build, ~/.almond/diagnostics/tuning/): the clipped raw command recording and the built motion side by side per joint, with per-joint change RMS/max and peak velocities. The diagnostics dashboard charts it like any other tuning run (zoom / fullscreen), so the effect of the smoothing pass and the collision projection is directly visible before the motion is ever replayed.