Seminar: Coupling Online Learning with Physical Models to Enable Autonomy in Uncertain Environments - Jan. 29
John Bird
Postdoctoral Researcher, RECUV
Friday, Jan. 29 | 12:30 P.M. | Zoom Webinar
Abstract : As autonomous systems become ubiquitous they will be tasked in critical roles while at the same time they will more frequently encounter situations which were not anticipated by their designers. To operate safely and accomplish the missions these systems are tasked with they will need to reason about their environments and interactions between themselves and the environment. Learning systems offer one route to manage complexity in the environment and feedback between actions and performance, but in practice a suitable model for pre-training may not be available and opportunities for online training may be very limited. Physical principles and characteristics of the environment and agent can be used to accelerate online learning by focusing on environmental states and agent actions which are physically realistic and likely to produce large rewards.
This seminar will discuss approaches to learning about the environment and filtering learned models through an agent or system’s self-knowledge to enable intelligent decision-making in complex and uncertain environments. In particular, I will focus on small aircraft systems managing uncertainty in the atmosphere. While weather is governed by physical principles it also exhibits structural uncertainty and stochasticity, making atmospheric flight an excellent laboratory to study decision-making under uncertainty. I will describe how an autonomous system can couple environmental modeling with active learning to identify high value actions, refine its knowledge of the environment relevant to these actions, and make intelligent decisions to efficiently accomplish a mission.
Bio: John Bird is a Postdoctoral Associate in the Research and Engineering Center for Unmanned Vehicles (RECUV) in the Smead Aerospace Engineering Sciences Department at the University of Colorado. His research interests include field robotics; online learning methods; autonomous scientific platforms; and decision-making under uncertainty, especially to improve safety and performance of aircraft flying in challenging weather conditions. John received his B.S. in Aerospace Engineering from Wichita State University in 2011 and his Ph.D. in Aerospace Engineering from Pennsylvania State University in 2019.