Psychophysics-Based Cognitive Reinforcement Learning to Optimize Human-Robot Interaction in Power-Assisted Object Manipulation

Abstract

This paper introduces a novel method of inclusion of human cognition in the form of weight perception in the dynamics and control of a power assist robotic system (PARS) for object manipulation. A 1-DOF test-bed PARS is developed. The dynamics for human-robot co-manipulation of objects is derived that includes weight perception. Then, an admittance control scheme with position feedback and velocity controller is derived from the weight-perception-based dynamics. In the control model, the mass parameter of the inertial force is considered different from that of the gravitational force. The system is simulated in MATLAB/Simulink for 36 different pairs of inertial and gravitational mass parameters. Human subjects lift an object with the system for each pair of parameters separately. The levels of human-robot interaction (HRI) is psychophysically evaluated by subjects separately using a Likert scale. In each trial, the subject evaluates the system for appropriate level of HRI. Then, a training database is generated following the reinforcement learning approach that includes inputs (pairs of mass parameter values) and corresponding outputs (levels of HRI). Then, the labeled database is used to predict a condition where subjects feel the highest level of HRI. Then, the mass parameters for the best HRI pattern are selected as the mass parameters to be used in the control system. In the testing phase, the best mass parameters are used in the control system, and the subjects evaluate the HRI for the system. The results show that the best mass parameters predicted through the reinforcement learning method produce satisfactory HRI, and in the contrary, the mass parameters deviated from the best mass parameters do not produce satisfactory HRI. The results show that inclusion of weight perception in the dynamics and control and optimization of HRI through reinforcement learning are effective.