Bipedal Locomotion: Developmental Trajectory Audit and Wrap

September 9, 2026

Part 6. Localizing when the VelStand intervention went wrong during training, and full project cost accounting.

Sep 9: Developmental Trajectory Sweep

Proofload evaluated 7 checkpoints across the VelStand training trajectory ( $1000 \to 1500 \to 2250 \to 3000 \to 4000 \to 5000 \to 5999$ ) across 4 push velocities and 2 frictions: 56 condition cells, 5,600 rollouts in 330 seconds (run 20260909T103415Z-e2cb). The matching baseline sweep is 9 checkpoints, 72 cells, 7,200 rollouts in 490 seconds (run 20260909T003230Z-16d4). Figures below are at $\mu=1.0$ ; the $\mu=0.6$ twins tell the same story.

Findings:

Both recipes pass through the same rigid checkpoint, VelStand one step earlier than the baseline. Baseline 2250 falls 0% at every push up to $0.60$ m/s and 100% at $0.675$ . VelStand 1500 and 2250 do the same with the wall at $0.60$ instead.
Both lose the cliff at the next checkpoint their grid sampled. Baseline 2500 falls $49\%$ at $0.60$ m/s; VelStand 3000 falls $69\%$ . The baseline is at $66\%$ by iteration 3000.
Then they part. The baseline works its fall rate at $0.60$ m/s down from $66\%$ at iteration 3000 to $15\%$ at 5999, which is the compliant slope resolved in Part 3. VelStand goes to $82\%$ at 4000, comes back to $58\%$ at 5000 and ends at $63\%$ . At $0.675$ m/s it never gets below $65\%$ after the cliff.

The intervention did not fail at the end. It failed to recover from losing the rigid solution, and the sweep says when.

Full Project Accounting

Everything in the store, from the pilot through the last sweep:

Segment	Runs	Rollouts	Wall time
First paired pass (confounded)	bd64, 7201	1,600	121 s
Paired rerun	7ab3, e81c	1,600	116 s
Boundary resolution	5abe, a6f1	2,800	202 s
Intervention diff	1090, 21a3	2,800	161 s
Baseline developmental sweep	16d4	7,200	490 s
VelStand developmental sweep	e2cb	5,600	330 s
Everything, including the confounded sweep and a smoke test	12 runs	28,920	1,880 s

Every rollout ran on one Apple Silicon laptop through Proofload’s local backend; nothing was sent to a cluster, and that part cost nothing. The manifests record the platform as macOS on arm64 and do not record the chip. Training ran on Google Colab Pro+ with an NVIDIA L4, about 4 h 45 min for the baseline; I did not record the dollar figure for either training run, so there is no total here.

A standard benchmark would have reported a higher multi-task score for VelStand. The paired grid says its walking boundary moved inward in all 14 cells tested. An engineering instrument made that falsifiable.