Transfer Learning exploiting Demonstrations in a Human-Robot Interactive Game

Abstract

Enhancing Human-Robot Collaboration requires robots that are not only socially aware but also proficient in adapting and learning from interactions to perform interdependent tasks effectively. This thesis extends recent research by focusing on the dynamics of Human-Robot Collaboration where humans collaborate with a Deep Reinforcement Learning agent to achieve a common goal. The performance in such collaborations depends on the Deep Reinforcement Learning agent’s ability to adapt and learn from its human partner and vice versa. Our study implements an alternative Transfer Learning (TL) approach, Learning from Demonstrations, specifically through Deep Q-Learning from Demonstrations (DQfD), aimed at encouraging more efficient human-robot teamwork. In contrast to the foundational work that utilized Probabilistic Policy Reuse, our approach, coupled with adjustments to the Soft Actor Critic algorithm’s settings, seeks to enhance adaptability and learning outcomes. We conducted experiments involving 24 participants to evaluate the impact of these changes. Our findings suggest that the direct transfer of expertise with Learning from Demonstrations, complemented by specific SAC algorithm settings can significantly influence the collaborative performance.