Fall 2024 GRASP SFI - Dylan Shell, Texas A&M University

“Robot situatedness and information requirements for tasks” ABSTRACT This talk describes investigations into the nature of robot–environment interaction and “niche fit” through the lens of state (or memory) minimization. The idea is that by limiting what a robot can store, much like so-called bottleneck methods, one hopes to uncover the information needed to perform specific tasks. We study a setting in which robots are able to exploit structural regularity within the environment. Doing so alters the minimization problem from classical reduction problems (i.e., those of Myhill–Nerode or bisimulation relations) in an important, fundamental way: it changes computational complexity class. The later part of the talk will try to explore interpretations and intuitions behind the (multiple) extra sources of this complexity. Touching briefly upon intriguing ways in which nondeterminism and casualty manifest themselves, the final part of the talk will describe how we are now approaching sensors within this theoretical framework too. Presenter Dylan Shell is a computer scientist at Texas A&M University who works in the areas of robotics and AI. Broadly, his research aims to analyze and synthesize complex, intelligent behavior in systems that interact with the physical world. He has an interest in algorithmic and formal foundations of planning problems, and extremely simple (or minimal) robots. He has published papers on topics from multi-robot task allocation, biologically inspired multiple robot systems, estimation of group-level swarm properties, rigid-body simulation and contact models, and robotic theatre. His work has been funded by DARPA, DoD (ONR, ARL), NSF as well as Ford and 3M; he has been the recipient of teaching, service and reviewing, and research awards. Dylan serves as the President of the Robotics Science and Systems Foundation.