Experiment Design and Implementation for Physical Human-Robot Interaction Tasks

Abstract

Human-robot interaction has been an active research field for several years with the focus on understanding, designing, and evaluating robotic systems that are useful in helping people to perform certain task. Competitive interaction is one of these tasks that is studied less but has great potentials in entertainment, physical therapy and athletic training areas. The main challenge in this field is to design algorithms that can allow robots to play against human. Recent research in multi-agent reinforcement learning have achieved satisfactory results in generating complex and dexterous robotic behaviors even under non-stationary environment. This progress opens up the research opportunity on applying multi-agent reinforcement learning in human-robot interaction tasks. To facilitate current and future research in such direction, this thesis aims to design and implement a robotic system to evaluate the viability of various multi-agent reinforcement learning algorithms under certain human-robot interaction scenarios. The robotic system consists of the following subsystems: 1. A real-time perception system that captures the dynamical state of certain human body frames(hands, feet, torso, and etc.). 2. Computational server to deploy deep learning models for decision making purpose. 3. A robot with high degree-of-freedom arms and a low latency communication method to connect all subsystems to ensure robot's real-time reactions.