Robotics Colloquium: Guest Speaker - Josiah Hanna
Title: Toward Deploying Reinforcement Learning with Confidence in Real-time and Dynamic Robotic Tasks
Abstract: Recent years have seen a surge of interest in reinforcement learning (RL) as a powerful method for enabling AI agents to learn how to act so as to achieve the goals set by their designers. In robotics, RL should be a natural choice for developing high-performing decision-making policies, and yet several challenges prevent its application, especially in complex, real-time, and dynamic environments. In the first part of this talk, I will describe my lab’s work on enabling RL in the domain of robot soccer, which exhibits a number of challenges that are absent from current RL benchmarks. This work will motivate a deeper dive into recent work that addresses the challenge of knowing when a learned behavior is performant enough to be deployed on a physical system. I will describe recent work from my group on predicting the performance of RL using past data. By building upon a framework of state abstraction from the reinforcement learning literature, we have developed new methods for offline policy evaluation that accurately estimate policy performance before a learned policy is deployed. I will describe our recent results and how these methods are a promising path toward scaling offline evaluation methods to real-world robotics settings.
Biography: Josiah Hanna is an assistant professor in the University of Wisconsin -- Madison Computer Sciences Department. He received his Ph.D. in the Computer Science Department at the University of Texas at Austin. Prior to attending UT Austin, he completed his B.S. in computer science and mathematics at the University of Kentucky. Before joining UW – Madison, he was a post-doc at the University of Edinburgh and also spent time at FiveAI working on autonomous driving. His research interests lie in artificial intelligence and machine learning, seeking to develop algorithms that allow autonomous agents such as robots to learn (efficiently) from their experience. In particular, he studies reinforcement learning and methods to make reinforcement learning more broadly applicable to real-world domains.
Assistant professor University of Wisconsin -Madison
Computer Sciences Department.
Computer Sciences Department.
Start date
Friday, Oct. 4, 2024, 2:30 p.m.
End date
Friday, Oct. 4, 2024, 3:30 p.m.
Location
In-person: Lind Hall L125