Bipedal Locomotion: The VelStand Intervention Trial

September 9, 2026

Part 5. Using the telemetry diagnosis to engineer an intervention: retraining with standing and recovery rewards to expand the boundary.

Sep 9: The Intervention Recipe

Hypothesis: If compliance and self-righting expand the boundary, retraining on Mjlab-VelStand-Flat-MicroDuck (VelStand) should produce a policy with 2250’s interior stability and 5999’s recovery extension.

The upstream repo ships that task as a second recipe. It adds to the walking objective:

• Ground-contact recovery and getting back upright.
• Stumble catching.

Trained velstand_5999.pt on Colab with the same seed and iteration count as the baseline, then benchmarked it against baseline model_5999.pt across 14 operating cells ($0.600 \to 0.750$ m/s, $\mu \in \{0.6, 1.0\}$, 100 paired rollouts per cell, 2,800 total). Runs 20260909T102454Z-1090 (baseline) and 20260909T102639Z-21a3 (VelStand).

Sep 9: Catching the Intervention Failing

VelStand was significantly worse in all 14 evaluated operating cells and safer in none. The largest $q$ across the 14 tests was $0.0054$, the worst-powered cell could detect a 20 pp change, and an omnibus permutation test over the whole grid rejects “no cell changed” at $p = 0.00025$.

• At $0.600$ m/s, failure rate skyrocketed from $12\%$ to $62\%$, in both friction rows.
• At $0.650$ m/s, baseline survives $>75\%$; VelStand falls in $69\%\text{--}71\%$.
• At the far edge of the grid, $0.750$ m/s, VelStand reads $81\%$ and $88\%$ against the baseline’s $62\%$ and $61\%$.

Baseline checkpoint 5999 (left) and VelStand checkpoint 5999 (right) under the same 0.650 m/s push, same control step and angle on both sides. One rollout each.

The telemetry that comes with it: post-impact corrective action rate differential dropped from $0.943$ to $0.321$. A story about the ground-recovery rewards interfering with the stepping reflexes needed to stay upright while walking is a plausible reading of that number, and it is not a result of this sweep. I do not have a measured mechanism.

What the sweep does establish is that the retrain moved the boundary inward by 19 to 57 pp across the whole tested range, and that a score on the VelStand task alone would not have shown it.