ABSTRACT: We have recently demonstrated the possibility of learning controllers that are zero-shot transferable to groups of real quadrotors via large-scale, multi-agent, end-to-end reinforcement learning. We train neural networks to control individual drones in a fully decentralized manner. The resulting policies, trained in simulation with realistic quadrotor physics, demonstrate advanced flocking behaviors, perform aggressive maneuvers in tight formations while avoiding collisions with each other, break and re-establish formations to avoid collisions with moving obstacles, and efficiently coordinate in pursuit-evasion tasks. The policies learned in simulation transfer to highly resource-constrained physical quadrotors. Motivated by these results and the observation that neural control of memory-constrained, agile robots requires small yet highly performant and safe models, the talk will conclude with some observations on 1. systematically searching for small networks that can be coaxed onto devices with modest computational capabilities, and 2. how the latent representations these networks encode may be edited for safer flight.

BIOSKETCH: Gaurav S. Sukhatme is Professor of Computer Science and Electrical and Computer Engineering at the University of Southern California (USC) and an Amazon Scholar. He is the inaugural Director of the USC School of Advanced Computing and the Executive Vice Dean of the USC Viterbi School of Engineering. He holds the Donald M. Aldstadt Chair in Advanced Computing and was the Chairman of the USC Computer Science Department from 2012-17. He earned a B.Tech. in Computer Science and Engineering at IIT Bombay and M.S. and Ph.D. degrees in Computer Science from USC. He is the co-director of the USC Robotics Research Laboratory and directs the USC Robotic Embedded Systems Laboratory. His research is in networked robots, learning robots, and field robotics. He is a Fellow of the AAAI, AAAS, and the IEEE, a recipient of the NSF CAREER award, the Okawa Foundation research award, and an Amazon research award. He is one of the founders of the RSS conference and was the program chair of RSS 2005, ICRA 2008, and IROS 2011. He is the Editor-in-Chief of Autonomous Robots (Springer Nature).