Learning Shared Control by Demonstration for Personalized Wheelchair Assistance

An emerging research problem in assistive robotics is the design of methodologies that allow robots to provide personalized assistance to users. For this purpose, we present a method to learn shared control policies from demonstrations offered by a human assistant. We train a Gaussian process (GP) regression model to continuously regulate the level of assistance between the user and the robot, given the user's previous and current actions and the state of the environment. The assistance policy is learned after only a single human demonstration, i.e. in one-shot. Our technique is evaluated in a one-of-a-kind experimental study, where the machine-learned shared control policy is compared to human assistance. Our analyses show that our technique is successful in emulating human shared control, by matching the location and amount of offered assistance on different trajectories. We observed that the effort requirement of the users were comparable between human-robot and human-human settings. Under the learned policy, the jerkiness of the user's joystick movements dropped significantly, despite a significant increase in the jerkiness of the robot assistant's commands. In terms of performance, even though the robotic assistance increased task completion time, the average distance to obstacles stayed in similar ranges to human assistance.