Title:
Bridging Model-based Control and Data-driven Learning: Differentiable Optimization for Autonomous Vehicles and Beyond
Abstract:
Recent advances in machine learning have enabled end-to-end frameworks that integrate perception, planning, and control for autonomous vehicles and robotics. However, many learning-based methods lack explicit incorporation of control-theoretic structure, resulting in poor data efficiency and limited interpretability. In contrast, model-based optimization and control provide efficient and certifiable solutions by leveraging structured dynamics and task specifications, but often at the cost of reduced adaptability and autonomy. This report explores how the two paradigms can be unified by embedding optimization-based planning and decision-making as inductive biases into learning architectures via differentiable optimization.
The talk begins with applications in energy-efficient and human-like autonomous driving, where optimal control offers interpretable and efficient behaviors. With this foundation, I will introduce an inverse optimal control-induced learning strategy for human-inspired vehicular cruise control of CAVs. Finally, I will present a unified, end-to-end trainable framework—differentiable optimization-integrated network—that enjoy both the flexibility of learning and the rigor of optimization, advancing robustness and safety in next-generation autonomous systems.
Biography:
Zifei Nie earned his Ph.D. in Mechanical and Systems Engineering from Kyushu University, Japan in March 2025. He is currently a Postdoctoral Research Fellow at Kyushu University and a Q-Energy Innovator Fellow at the International Institute for Carbon-Neutral Energy Research (I²CNER).
His research focuses on real-time optimization with limited resources, safe learning via differentiable MPC for eco-driving of CAVs, and human-guided differentiable optimization-integrated learning for End-to-End Planning. He has published multiple papers in journals such as Applied Energy, IEEE-TITS, Energy, and Transportation Research Part C, and received the 2024 IEEE Best Paper Award.
During his Ph.D., he led and participated in JST-funded decarbonization projects and received multiple honors, including Kyushu University’s Outstanding Graduate Student Award and the Award of Support Program for Young Researchers.
报告时间：2025年6月16日（星期一)上午9：40-10：40
报告地点：正新楼3楼人工智能学院报告厅
主办单位：吉林大学人工智能学院
