Physics-Informed Learning and Control for Mixed Autonomy Systems: Application to Traffic Smoothing

About the talk

Mixed-autonomy systems, where automated and human agents coexist, are already emerging in real-world cyber-physical systems. A key challenge in these systems is how to leverage a small number of automated agents to influence overall system behavior under nonlinear dynamics, behavioral uncertainty, and partial observability. This talk presents a model-informed co-design framework that integrates physics-informed learning, control design, and real-world experimentation for mixed-autonomy systems. Using traffic flow smoothing as a specific case, it will be shown how Neural Finite Volume Methods enable accurate and data-efficient modeling of traffic dynamics, and how kernel-based and imitation-learning control strategies allow a few automated vehicles to dissipate stop-and-go waves while maintaining throughput. These methods are validated through the largest scientific traffic field experiment to date, involving 100 automated vehicles on public highways, demonstrating measurable improvements in traffic stability and energy efficiency.

Speaker