Bipedal Locomotion: High-Frequency Telemetry and Mechanism

September 9, 2026

Part 4. Telemetry investigation of why 2250 collapses into a cliff while 5999 still walks away from more than half its rollouts at $0.700$ m/s.

Sep 9: Instrumenting the Recovery Window

Logged joint angles, actuator target rates, contact forces and base orientation across the post-impact recovery window ($t \in [2.3, 3.5]$ s) at $v_y = 0.650$ m/s. Fifty discordant seeds for the aggregate roll figures, plus a single fully traced rollout at seed 0 for the step-by-step comparison below. Control rate is 50 Hz throughout, same as the sweeps.

Telemetry Findings

Numbers 1 to 3 and the roll ceiling are the seed-0 trace; the mean maximum roll is over the fifty seeds.

1. Action Differential: Following impact, 5999 exhibits a $2.3\times$ surge in actuator rate of change ($1.123$ vs $0.482$). 2250 barely reacts, holding nominal walking posture.
2. Hip Abduction: 5999 actively abducts its right hip outward to $-0.165$ rad. 2250’s hip stays near zero ($+0.002$ rad).
3. Stance Width Expansion: Coincident with hip abduction, 5999 expands its maximum lateral stance width to $9.70$ cm against 2250’s $7.29$ cm, a $+33\%$ difference in support polygon.
4. Roll Dynamics: 5999 arrests roll at a maximum tilt of $11.8^\circ$ and re-centers within $0.36$ s. 2250 passes $23.2^\circ$ and does not recover, striking the floor at $t=4.16$ s. Across the fifty seeds, mean maximum roll is $70.3^\circ$ for 2250 against $38.5^\circ$ for 5999.

The measurements are consistent with 5999 responding more aggressively to the disturbance, widening its base of support, and damping trunk rotation, where 2250 makes much smaller corrective changes and fails to recover. That reading is a hypothesis the telemetry does not contradict, not a causal result of this sweep. The intervention trial in the next part is the direct test.