Assist-as-Needed Control for Upper Limb Rehabilitation Using a Learned Passive Interaction Force Reference

Upper-limb rehabilitation robots should provide adaptive assistance while avoiding excessive robotic intervention when users can contribute to movement. A key challenge is that measured interaction forces contain both voluntary and passive mechanical components, making direct force based participation assessment unreliable. This study proposes an Assist-as-Needed control framework based on a learned passive interaction force reference for a four degree of freedom upper-limb rehabilitation robot. Passive experiments with five healthy participants were conducted during eating and drinking tasks, and interaction forces were recorded at the upper-arm and forearm interfaces. The results revealed clear task, interface, and directional dependence in passive interaction. A Gated Recurrent Unit (GRU) model was then developed to estimate the expected passive interaction force from robot kinematics, movement phase, and task information, achieving a maximum (R²) of approximately 0.92. The deviation from the estimated passive reference was used to construct a control-oriented participation estimate, which was combined with trajectory tracking performance to regulate adaptive impedance. Controller-level simulation showed that robotic assistance decreased with increasing prescribed interaction contribution and increased again when tracking performance deteriorated. These results support the feasibility of using a learned passive interaction reference for Assist-as-Needed regulation and provide a basis for subsequent active human validation of the participation estimator.