报告人
Jiachen Hu
Peking University
时间
2025年11月11日 星期二
下午 14:00-15:00
地点
602会议室
Abstract
Embodied intelligence is widely regarded as a key pathway toward general artificial intelligence, where agents progressively acquire perception, planning, and decision-making capabilities through interaction with the physical world. However, training agents directly in real environments is often hindered by data collection challenges, high costs, and safety risks. Sim-to-real transfer—training policies in simulation and deploying them in the real world—offers an effective route toward embodied intelligence. Despite the notable successes of approaches such as Domain Randomization and Robust Adversarial Training in robotic control and related tasks, their theoretical foundations remain underdeveloped.
This talk introduces theoretical models and algorithmic analysis for sim-to-real transfer. In discrete state spaces, we model the simulator as a family of parameterized Markov Decision Processes; in continuous state spaces, we adopt Linear Quadratic Gaussian (LQG) systems. In both settings, we derive rigorous upper and lower bounds for the sim-to-real gap. Our results show that, under appropriate conditions, it is possible for an agent to learn near-optimal policies even without access to real-world samples by effectively leveraging historical information. Collectively, these works establish a unified theoretical framework for sim-to-real transfer, deepen our understanding of its efficacy and boundary conditions, and provide a solid foundation for both the theoretical study and practical deployment of embodied intelligence.
Biography
Jiachen Hu received his Ph.D. degree in 2025 from the School of Computer Science at Peking University, advised by Prof. Liwei Wang and Yuqing Kong. In 2023, he was a visiting scholar at Princeton University, collaborating with Prof. Chi Jin. He is currently a research scientist working on recommendation foundation models at ByteDance. His research interests include fundamental theory and algorithms for reinforcement learning and online learning, as well as applications of reinforcement learning in economics, quantum computing, and mathematics. He obtained his B.S. degree in 2020 from the Turing Class of the School of EECS, Peking University. He has published 10+ papers at top venues in machine learning and economics, including ICML, ICLR, and FC. He is a recipient of the Peking University Presidential Scholarship for outstanding M.S./Ph.D. students and multiple awards for excellent research achievements. He also actively serves as a reviewer for premier conferences such as ICML, ICLR, NeurIPS, and AAAI.
