RNNPB-based Shared Control for Maintaining Sense of Agency in Robotic Teleoperation

The sense of agency (SoA), the feeling of recognizing that the observed movement is caused by oneself, which is important in robot teleoperation, is reduced by shared control, in which the robot and the human cooperate to control the robot. In this study, we developed a system that uses a recurrent neural network with parametric biases (RNNPB) trained on expert operational data to predict the next input from nonexperts and convert it into robot commands in real time. Through an experiment with a pouring task, it was confirmed that the proposed method outputs predicted values that spatially and temporally interpolate the operational inputs, gradually correcting the robot’s movements to align with the experts’ trajectories. The proposed method showed a high SoA comparable to direct control; however, no statistically significant difference in task performance was observed. Future work aims to improve the generality of the model to accommodate a wider variety of input trajectories.