Lightweight Data-Driven Admittance Control for Massage Robots via Offline Temporal Difference Learning
Abstract
Safe and effective physical human-robot interaction (pHRI) is crucial for massage robots operating in unstructured environments. Traditional fixed-parameter admittance control struggles to adapt to heterogeneous human soft tissues, while online parameter identification methods can be affected by computational latency and sensor noise. To address these limitations, we propose a lightweight, data-driven, adaptive admittance control framework. This framework uses a quadratic action-value approximation and utilizes model-free Temporal Difference (TD) learning on offline interaction data. By doing so, it extracts a scalar state-feedback gain subject to parameter constraints. This scalar gain is then deployed in a 125 Hz low-level control loop as a fixed state-feedback law to modulate the normal contact force. Furthermore, incorporating low-level digital filtering and kinematic hard constraints, we derive numerical bounds on admittance velocity and virtual kinetic energy. Progressive experiments on a 6-DOF manipulator interacting with a silicone phantom and four human participants cover complete leave-one-subject-out validation and a 10/20/30 N by 10/30/60 kg force–mass grid. Across 1,152 human massage trials (288 per controller), the proposed framework achieved a mean absolute error of 4.94 N and an In-Zone Ratio of 56.34%, compared with 5.86 N and 47.80% for fixed-parameter admittance. These correspond to a 15.71% reduction in average error and an 8.54-percentage-point increase in tolerance occupancy. The improvements occurred in the participant-averaged results of all four held-out subjects. The framework connects offline interaction learning with lightweight online force regulation, requiring no online policy update.