Human-Aware Reinforcement Learning for Fault Recovery Using Contextual Gaussian Processes

Abstract

This work addresses the iterated nonstationary assistant selection problem, in which over the course of repeated interactions on a mission, an autonomous robot experiencing a fault must select a single human from among a group of assistants to restore it to operation. The assistants in our problem have a level of performance that changes as a function of their experience solving the problem. Our approach uses reinforcement learning via a multi-arm bandit formulation to learn about the capabilities of each potential human assistant and decide which human to task. This study, which is built on our past work, evaluates the potential for a Gaussian-process-based machine learning method to effectively model the complex dynamics associated with human learning and forgetting. Application of our method in simulation shows that our method is capable of tracking performance of human-like dynamics for learning and forgetting. Using a novel selection policy called the proficiency window, it is shown that our technique can outperform baseline selection strategies while providing guarantees on human use. Our work offers an effective potential alternative to dedicated human supervisors, with application to any human–robot system where a set of humans is responsible for overseeing autonomous robot operations.