Perceptual Control for Assistive Robotics

People with transradial limb loss may benefit from assistive systems that combine muscle signals, gaze, and scene information while preserving user control. This work proposes Multimodal Affordance-based Shared Control (MASC), a framework that uses device-relative geometry to filter object–grasp candidates before confidence assessment and user confirmation. The proposed architecture integrates surface electromyography, eye tracking, and RGB-D vision with constraints on hand aperture, reach, and task feasibility. A synthetic benchmark comprising 100,000 evaluations assessed noisy intent ranking and geometric filtering. Hard filtering increased target-selection match from 29.32% for intent-only selection to 77.52%, while reducing infeasible selections from 66.12% to 8.50%. Adding a 5 mm feasibility margin further reduced infeasible selections to 3.38%, but increased abstention from 30.10% to 38.77% and reduced match to 73.56%. These findings illustrate a tradeoff between selection feasibility and abstention under simulated uncertainty. The benchmark evaluates the filtering component rather than the complete MASC system; it includes no human participants, measured muscle signals, or physical device trials. Future work will assess fault handling and compare matched multimodal controllers with and without feasibility filtering before human evaluation.