Agency-Constrained Shared Autonomy for Prosthetic Grasp Selection: A Reproducible Offline Benchmark for Interactive AI in Assistive Robotics

Abstract

Upper-limb prostheses increasingly combine user-intent decoding with contextual autonomy, improving robustness but risking reduced sense of agency when assistance overrides user control. We present agency-constrained shared autonomy (ACSA), which treats control authority as an explicit, auditable interaction budget. ACSA computes an Agency Loss Index (ALI) from the user-intent distribution and restricts autonomy to an agency-feasible envelope. SetACSA then selects the autonomy-preferred grasp within that envelope. Using the open MeganePro/NinaPro DB10 dataset, we created an offline replay benchmark comprising 3,814 grasp episodes from five able-bodied participants and four transradial amputees. SetACSA improved accuracy from 0.532 ± 0.021 for user-only control to 0.580 ± 0.014 while maintaining low agency loss (ALI = 0.012 ± 0.001). Confidence-weighted blending achieved comparable accuracy (0.560 ± 0.017) but substantially greater agency loss (ALI = 0.087 ± 0.004). These results support evaluating prosthetic shared autonomy through a performance–agency frontier rather than accuracy alone.