This seminar will be presented by Dr. Michael Everett. Michael Everett received the S.B., S.M., and Ph.D. degrees in mechanical engineering from the Massachusetts Institute of Technology (MIT), in 2015, 2017, and 2020, respectively. He is currently a Postdoctoral Associate with the Department of Aeronautics and Astronautics at MIT.
Autonomous robots have the potential to transform our everyday lives, yet most of today's autonomous robots struggle in the real world. This talk will first describe our work toward a new generation of robots that learn to handle the highly dynamic and uncertain nature of human environments.
In particular, I will highlight the importance of obtaining accurate cost-to-go models, which we show can be learned from self-play or aerial imagery for a variety of applications, from navigation among pedestrians to last-mile delivery. The talk will then dive into the challenges of certifying the safety and robustness properties of machines that learn. I will describe our work that uses convex relaxations and set partitioning to simplify the analysis of highly nonlinear neural networks used across AI. These analysis tools led to the first framework for deep reinforcement learning that is certifiably robust to adversarial attacks and noisy sensor data. The tools also enable reachability analysis -- the calculation of all states that a system could reach in the future -- for systems that employ neural networks in the feedback loop, which provides another notion of safety for learning machines that interact with uncertain environments.
Finally, I will discuss my long-term vision to spark a new era of autonomy defined by robots that are resilient, dependable, and ready to support humans throughout the real world.