报告时间:2025年2月21日(星期五)上午10:30
报告地点:浙江大学紫金港校区蒙民伟楼402会议室
报告题目:Reinforced Continual Evolution of Decision-Making Systems
报告人:Wenhui Huang 博士
主持人:陈昊 研究员
报告摘要:Reinforcement learning (RL) has emerged as a transformative framework, driving breakthroughs in decision-making tasks across diverse fields, including gaming, robotic control, and autonomous driving. This success stems from RLs ability to achieve robust performance even in scenarios where data distributions are insufficiently representative of the broader sample space. However, RLs inherent trial-and-error learning paradigm is widely regarded as inefficient, particularly in real-world environments characterized by significant uncertainty. When integrated with foundation models, such as large language models (LLMs) and visionlanguage models (VLMs), RL holds significant promise for generalizing across long-tail distributions. Yet, this integration often amplifies existing inefficiencies and impracticalities, particularly due to catastrophic forgetting, presenting critical challenges for real-world deployment. This talk presents how to construct a continually evolving decisionmaking system grounded in foundation models, focusing on three pivotal aspects: enhancing exploration efficiency, improving generalization through foundation models, and enabling continual learning on top of well-generalized decision policies. The presentation will start with fundamental theoretical insights, transition to real-world applications, and conclude with a forward-looking discussion on collective continual evolution and input-output modality alignment in decision-making systems.
报告人介绍:Wenhui Huang (Member of IEEE, ITSS, and RAS) is currently a Ph.D. candidate and research associate at the School of Mechanical and Aerospace Engineering, Nanyang Technological University (NTU), Singapore. This March, he will join the computational robotics group in Harvard university as a visiting Ph.D. scholar. He was one of the pioneering team leaders representing Singapore in its historic first participation in a real-world super formula autonomous racing challenge. His research focuses on autonomous driving and racing, human-in-the-loop AI, foundation models, and lifelong reinforcement learning. He received many awards and honors, selectively including distinguished doctoral student in ITS field from IEEE ITSS, first place in occupancy flow prediction at the Waymo Open Dataset Challenges (CVPR 2024), Best Paper Runner-Up at the 26th IEEE International Conference on Intelligent Transportation Systems (ITSC 2023), and Silver medalist at the Intelligent Algorithm Final of the 2022 Alibaba Global “Future Vehicle” Challenge. Additionally, he serves as a reviewer for over 20 journals and conferences, including IEEE T-PAMI, Engineering, IEEE T-ITS, IEEE TNNLS, IEEE TIV, IEEE TVT, IEEE TASE, IEEE IV, IEEE ITSC, IEEE ICRA, and IEEE IROS.