Learning and Control for Interactions in Mixed Human-Robot Environments
Speaker: Wilko Schwarting, MIT
Date: March 24, 2021, 11 a.m.
Host: Ludovic Righetti
Autonomous robots are on the verge of transforming our homes, factories, and roads. But to reap the tremendous benefits that robots offer to society, we must ensure that they can interact with humans seamlessly and safely.
Towards this goal, my work focuses on giving robots the ability to acquire complex skills from interaction, to perform robustly in uncertain environments, and to communicate intent through naturalistic behavior.
In this talk, I will first present intelligent agents that learn how to reason about human behavior and people’s intentions. By incorporating social behavior, we can measure people’s willingness to cooperate when our interests are not aligned and negotiate through game-theoretic interactions.
Second, I will show how in stochastic environments with partial observability, autonomous agents can leverage information gain and reason about others' beliefs by combining game-theoretic and belief-space planning. I will present fast, scalable, and calibrated uncertainty estimation of neural networks and discuss how we can leverage uncertainty estimates in learned models for safe operation and efficient exploration in RL.
Third, to learn complex behaviors, I will present reinforcement learning agents that learn competitive visual control policies through self-play in imagination. They learn complex skills from competition by imagining multi-agent interaction sequences in the compact latent space of a learned world model.
Wilko is the AI lead at iSee AI. He received his PhD in Electrical Engineering and Computer Science from MIT CSAIL in 2021, advised by Daniela Rus and Sertac Karaman. Prior to that, he obtained his BSc and MSc in Robotics, Systems and Control from ETH Zurich in 2014 and 2016 respectively. His research interests include interactions in mixed human-robot environments, learning complex behaviors that enable high-level reasoning, and planning under uncertainty arising from perception and prediction while incorporating others' beliefs. His work has been recognized through Best Paper Awards at ICRA & ITSC and extensive coverage in the press (Forbes, Wired, Daily Mail).