AAI-CCI-MHI Seminar on CPS: David Snyder

Speaker: David Snyder, PhD Candidate

Talk Title: How Do We Know Good Policies When We See Them?

Abstract: The recent and ongoing revolution in state-of-the-art robotic architectures has dramatically altered the empirical and theoretical landscape of robotics research. Unprecedented empirical success rates on long-horizon tasks with nonlinear dynamics and high-dimensional sensory inputs have been achieved through the adoption of large models, trained on large datasets, requiring large compute budgets. However, these substantial successes contravene many principles of generalization and robustness in the theories of statistical learning and control. Equally pressing, significant empirical gaps remain to human-level performance, and the observed presence of unpredictable yet preventable failure modes has driven research attention to rigorous methods for uncertainty quantification, robust control, and the understanding of fundamental limits of robotic procedures that are implemented in the real world. In this talk, I will discuss several works that investigate the "Evaluator Problem" for robotic tasks in varied uncertainty regimes, with the goal of providing practical, computable, and non-vacuous guarantees that complement the increasingly complex robotic decision-making procedures used in practice. In each instance, the core insight is to treat large models as input-output filters and design structure in the output space that is amenable to analytical methods. I will conclude with a brief discussion of two promising future directions: modeling uncertainty regimes which interpolate stochastic and adversarial scenarios, and nonparametric questions of data efficiency in robotic applications. 
 

Biography: David Snyder is a final-year PhD candidate in the Department of Mechanical and Aerospace Engineering at Princeton University, where he works in the Intelligent Robot Motion (IRoM) Lab advised by Prof. Anirudha Majumdar. Previously, he received a B.S. in Aerospace Engineering and Economics from the University of Maryland. Snyder's work focuses on constructing efficient, non-vacuous, and practical guarantees for the generalization and safety of robotic systems operating in unstructured environments, with particular emphasis on augmenting the roboticist's capacity to adaptively evaluate the behavior of complex systems and reduce the epistemic risk of static design decisions. During his PhD he is fortunate to have collaborated with multiple external research groups including Google DeepMind and the Toyota Research Institute. Additionally, he is a grateful recipient of the NSF GRFP. 

Host: Stephen Tu

Location: Hughes Aircraft Electrical Engineering Center (EEB) - 132

Audiences: Everyone Is Invited

Contact: Ariana Perez

This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor.

Add to Google Calendar