Learning Robotic Contact Juggling
We are happy to share that our project on robotic juggling in proceedings of the 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2021)!
Kazutoshi Tanaka, Masashi Hamaya, Devwrat Joshi, Felix von Drigalski, Ryo Yonetani, Takamitsu Matsubara, and Yoshihisa Ijiri “Learning Robotic Contact Juggling”, accepted to IROS 2021. [Paper]
Overview
Robotic juggling has been a fascinating task for the robotics community and also serving as a benchmark for evaluating agile and dynamic robotic manipulations. In this project, we address contact juggling, a challenging juggling play that a performer stops, rolls, slides, and throws an object over the surface of their body, while keeping the object in and sometimes out of contact with the body, without holding it in their hands.
In particular, we are interested in a method of learning to play contact juggling from trial and error, without laborious physical modeling of complicated robot-ball interactions. Specifically, we seek an approach based on model-based reinforcement learning (MBRL). Nevertheless, modeling the robot-ball interactions is non-trivial, primarily because the robot must control the movement of an object indirectly by changing the position and orientation of the link over which the ball moves.
To enable MBRL for robotic contact juggling, it is effective to model robot-ball interactions into a sequence of simple multiple models. Indeed, contact juggling consists of several distinct behavior primitives depending on whether or not the ball is in contact with the body of the performer.
Based on the above insight, we develop a novel MBRL method named switched multiple model-based reinforcement learning (S-MMRL), which is tailored to learning the state transition dynamics of the robot-ball interactions in robotic contact juggling. S-MMRL discovers the primitives of the robot-ball interactions by fitting a set of simple dynamics models to samples collected from the environment. At the same time, S-MMRL learns a switching model that takes the current state and action as input to determine which dynamics model to use for accurately representing the subsequent dynamics.
To verify the effectiveness of the proposed approach, we have developed a realistic simulation, which simulates a robotic arm playing contact juggling. We confirm that our approach allows the robot to learn to through and catch a ball in a more efficient fashion compared to several existing methods.
At OMRON SINIC X, we will continue fundamental research on computer vision, machine learning, and robotics. If you are interested in working with us as an intern, send us your application at internships@sinicx.com and get in touch!
