What Should Be Personalized in Ankle Exoskeleton Assistance? Decision Robustness under Biomechanical Uncertainty

Personalizing ankle-exoskeleton assistance relies on either hour-scale human-in-the-loop (HIL) optimization or musculoskeletal simulations driven by joint kinematics and net joint torques. Because net torque does not uniquely identify muscle states, we asked which features of effort-optimal ankle assistance can be inferred from torque-level gait data. We analyzed 20 young adults at three treadmill speeds in the public Fukuchi et al. dataset, using ensembles of 128 three-muscle ankle models consistent with each measured gait cycle. Uniform muscle-strength scaling is equivalent to rescaling antagonist co-contraction, and normalized regret is invariant under this identity; in a development-set probe, strength-ignoring regret was at most 0.05% of baseline effort. Optimal peak timing varied by a median 1% of the gait cycle across ensembles, and the predicted distribution was consistent with reported HIL timing; the muscle-cost-weighted matched filter provides an approximation to optimal timing. At the primary 0.5 N·m/kg device cap, every ensemble member selected the cap in all 60 cases. Full-ensemble saturation persisted through 0.6 N·m/kg and occurred in 92% of cases at 0.8 N·m/kg. Without a cap, model optima exceeded published HIL magnitudes at the matched speed. A capped profile with at most 1% worst-case regret existed in 57/60 cases. The paired leave-one-subject-out penalty for sharing a profile had median values of 0%, 5%, and 13% of baseline effort at slow, comfortable, and fast speeds. These findings support using torque-level data to guide assistance timing and profile shape within this model class, but do not identify the human-preferred magnitude.