Cognitive-State-Adaptive Agency Budgets for Shared Autonomy in Upper-Limb Prosthesis Grasp Selection: Offline Evidence From MeganePro (NinaPro DB10)

Multi-grasp prosthetic hands can execute diverse grasps, but grasp selection remains cognitively demanding, especially when users must monitor sensors, task context, and device state simultaneously. We propose a cognitive-state-adaptive agency budget (CS-AAB) for shared autonomy in upper-limb prosthesis grasp selection. CS-AAB converts user-intent uncertainty and a pupil-derived workload proxy into an episode-level budget τ , then uses set-based constrained shared autonomy to select the autonomy-preferred grasp only when the Agency Loss Index (ALI) remains within that budget. Using the open MeganePro/NinaPro DB10 dataset ( N = 9 ; 3,814 episodes; 10 grasp classes), constrained arbitration improved accuracy while keeping agency loss low. CS-AAB achieved the best mean accuracy (0.585), low ALI (0.027), and significantly higher budgets during high-workload episodes ( p < . 001 ). The framework provides a reproducible, human-factors-grounded benchmark for agency-preserving augmented-cognition prosthetic controllers before online human-subject trials.