Abstract

Reinforcement learning (RL) has demonstrated remarkable success in domains such as control and large language models (LLMs), and increasingly powers agents for real-world decision-making. However, whether these agents are trained as control policies or built on foundation models, their practicality hinges on a single obstacle: the gap between simulation and reality, which makes strong benchmark performance a poor predictor of real-world behavior. In this talk, I will trace this simulation-to-reality gap from RL-based control to the emerging era of foundation-model agents and present our recent efforts to address it. Specifically, I will share our preliminary efforts on reframing the foundation-model-agent gap as a classical sim-to-real problem and bridging the sim-to-real gap through uncertainty quantification. These directions aim to move current systems from descriptive and predictive insights toward truly reliable, actionable intelligence.

Bio

Hua Wei is an Assistant Professor in the School of Computing and Augmented Intelligence (SCAI) at Arizona State University (ASU). His research interests lie in data mining and machine learning, with a particular emphasis on spatio-temporal data mining and reinforcement learning. His research has been supported by various federal agencies, including the National Science Foundation (NSF), the Department of Energy (DoE), and the Department of Transportation (DoT). Dr. Wei is a recipient of the NSF CAREER Award, NSF CRII Award, Amazon Research Award, Cisco Faculty Research Award, and Toyota Faculty Research Award. His work has received multiple Best Paper Awards, including honors from ECML-PKDD and ICCPS. His research has been published in leading conferences and journals in machine learning, artificial intelligence, and data mining, such as NeurIPS, ICML, ICLR, AAAI, IJCAI, CVPR, and KDD.