Unitree A1 · teacher run 1 of 2 · 1500 iterations

The motion-library prior is finally the thing driving the robot

APT-RL §1a wired into the parkour stack. This checkpoint is the first one whose torque actually comes from the pretrained decoder rather than from the PD controller quietly cancelling it.

model_1499 · 16 envs · parkour curriculum · 400 frames

The audit that mattered more than the reward curve

Aggregate return has now failed five separate times on this task to distinguish a working policy from a broken one. So before reading anything into the terrain level, the question was: who is supplying the joint torque?

model_1499, 16 envs × 250 steps	mean	p95
Decoder torque \|τ_dec\|	2.690 N·m	3.842
Total applied \|τ\|	2.982 N·m	5.124
Decoder share of applied torque	90.2 %	—
Latent spread (library fit: 0.538)	0.996	—
Gait selection	11.2 % trot	88.8 % bound

Which retracts a comparison I made earlier

I had flagged this run's terrain level as disappointing against an earlier run that reached 3.54. That comparison was measuring nothing.

Earlier run · iter 906

312 %

Decoder share above 100 % means the PD was spending its authority cancelling the decoder, which was being driven roughly 30× outside the latent distribution it was fitted on. Terrain 3.54 — reached by discarding the prior and reverting to plain position control.

This run · iter 1499

90.2 %

Decoder driving, latent within ~2× of the fitted spread, auxiliary channel acting as a correction rather than a replacement. Terrain 1.888 — a lower number describing a fundamentally different policy.

Why a flat-ground prior doesn't sink the parkour policy

The library is 2D sagittal trajectory optimisation on flat ground. The terrain is not. The paper names this tension itself, in explaining why its AMP baseline fails:

combining the AMP discriminator with additional hand-crafted style rewards can overly constrain the policy when the training environment requires motions outside the reference dataset… For obstacle traversal, however, the policy must adapt these motion patterns beyond the range covered by the flat-terrain data. arXiv 2607.13579, supplement p24

The prior sits in the action space, not the reward. AMP charges a penalty for looking unlike the data; this architecture charges nothing. The decoder shapes what is cheap to reach and never what is permitted, so deviation is free.
The auxiliary action is the escape hatch, and they measure it firing. On the log jump, decoder torque dominates the approach and auxiliary torque takes over the jump itself — a motion absent from the pretraining set. The sharpest case is yaw: the library is 2D, so it contains no hip-abduction torque at all, and the auxiliary channel supplies that entire degree of freedom.
Gait choice routes around the mismatch. Bound's paired-leg flight phases suit obstacle crossing; trot's envelope does not. The selector picks per half-second — which makes this run's 88.8 % bound preference an unforced match to the paper, though lopsided enough to keep an eye on.
The curriculum front-loads the in-distribution regime. Two flat levels precede the obstacle sequence, banking basic walking exactly where the prior is most valid.

None of which makes the prior optional. Deleting τ_dec collapses their success rate from 94.6 % to 2.9 %. It is load-bearing across the easy majority of the task and permissive across the hard minority — and that asymmetry is the whole design.

Run 2 — identical policy against a randomly-initialised decoder, the control that says whether any of this is the library rather than the architecture — is in progress. At last check it sat at terrain 0.47 with longer episodes than run 1, the signature of a policy surviving by not going anywhere.