A physics-based Juggling Simulation using Reinforcement Learning

Abstract

Juggling is a physical skill which consists in keeping one or several objects in continuous motion in the air by tossing and catching it. Jugglers need a high dexterity to control their throws and catches which require speed, accuracy and synchronization. Also, the more balls we juggle with, the more those qualities have to be strong to achieve this performance. This thesis follows a previous project made by Lee et al.[1] where they performed juggling to demonstrate their method. In this work, we want to generalize the juggling skill and create a real time simulation by using machine learning. A reason to choose this skill is that Studying the ability to toss and catch balls and rings provides insight into human coordination, robotics and mathematics as written in the article Science of Juggling[2]. That is why juggling can be a good challenge for realistic physical based simulation to improve our knowledge on these fields, but also to help jugglers to evaluate the feasibility of their tricks. In order to do it, we have to understand all the different notations used in juggling and to apply the mathematical theory of juggling to reproduce it. In this thesis, we find an approach to learn juggling. We first break the need of synchronization of both hands by dividing our character in two. Then we divide the juggling into two subtasks catching and throwing a ball, where we present a deep reinforcement learning method for both of them. Finally, we use these tasks sequentially on both sides of the body to recreate the all juggling process. As a result, our character learns to catch all balls randomly thrown to him and to throw it at the velocity wanted. After combination of both subtasks, our juggler is able to react accurately and with enough speed and power to juggle up to 6 balls, even with external forces applied on it.