A quantitative fall threshold for graded unilateral motor-command attenuation in a simulated feedforward humanoid locomotion policy
Year: 2026
Type: article
Abstract: Learned locomotion policies for humanoid robots are usually evaluated with all actuators and command channels intact, and how much one-sided degradation of their motor commands they tolerate without retraining is rarely quantified. We addressed this question entirely in physics simulation, using eight independently trained feedforward locomotion policies for the Unitree G1 humanoid model. We attenuated the output-layer weights and biases for all 13 actuated joints on one side of the body by a common factor α, which scales the commanded deviation from the default posture on that side by exactly α while leaving torque limits and joint stiffness unchanged. No retraining was performed. Falls, defined as a tilt of the pelvis (the robot’s root body) exceeding 70° within 5 s, rose steeply as α decreased from about 0.75 to 0.55. In single-rollout sweeps at 0.5 m/s, per-policy thresholds ranged from 65% to 75% remaining output for right-side attenuation and from 60% to 70% for left-side attenuation, and the right side was more fragile in 6 of 8 policies, with none in the reverse direction (exact sign test, p = 0.031). Repeated trials under 10 randomized conditions per policy showed substantial within-policy variability and gave mean 50% fall thresholds of 68.4% (95% CI 66.6%–70.1%) for right-side and 64.6% (62.6%–66.6%) for left-side attenuation (right higher in 7 of 8 policies; Wilcoxon p = 0.039). Across four walking speeds the group-level right-side threshold was never lower than the left, but the per-policy asymmetry reached significance only at 0.5 m/s. At fall onset, joint angles had diverged progressively from the default posture over the preceding ≈0.6 s rather than converging to it. These results provide a reproducible, variability-aware measure of the unilateral command-degradation margin of a learned humanoid locomotion policy. The perturbation was motivated by acute hemiparesis, the sudden weakness of one side of the body after a stroke, before any recovery or relearning, but it is a computational analogy and not a model of human stroke.
Source: Frontiers in Robotics and AI
Author: Kenshi Obaru
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,353
OpenAlex ID: W7225799913