SPARC: an efficient way to combine reinforcement learning and supervised autonomy

Abstract

Shortcomings of reinforcement learning for robot control include the sparsity of the environmental reward function, the high number of trials required before reaching an efficient action policy and the reliance on exploration to gather information about the environment, potentially resulting in undesired actions. These limits can be overcome by adding a human in the loop to provide additional information during the learning phase. In this paper, we propose a novel way to combine human inputs and reinforcement by following the Supervised Progressively Autonomous Robot Competencies (SPARC) approach. We compare this method to the principles of Interactive Reinforcement Learning as proposed by Thomaz and Breazeal. Results from a study involving 40 participants show that using SPARC increases the performance of the learning, reduces the time and number of inputs required for teaching and faces fewer errors during the learning process. These results support the use of SPARC as an efficient method to teach a robot to interact with humans.