DIFFERENTIABLE SOCIAL PROJECTION WITH DEEP SELF-MODEL IMPLANTS FOR ASSISTIVE HUMAN-ROBOT COMMUNICATION

Abstract

With the proliferation of assistive robotics, it is paramount that both the robot and the human users communicate their intentions to maintain effective collaboration. However, it is unclear if the existing state-of-the-art methods can extract the most relevant information and reason about the best timing to provide the necessary communication. To this problem, we have proposed a framework called MIRROR, which leverages the concepts from the social projection theory. MIRROR uses reinforcement learning to learn the robot’s self-model while “implants” are placed at different parts of the self-model to learn the differences between the human and the robot. In this Thesis, I would like to report on my two contributions: 1) the implementation of the self-model training on a simplified gridworld driving domain, and 2) conducting of an actual human study on a high-dimensional assistive driving domain in CARLA simulator and the analysis of the results.