Efficient and resilient coordination among autonomous agents plays a central role in domains such as energy management, robotic swarms, and transportation systems. As these systems scale in size and complexity, achieving high performance while respecting safety constraints becomes increasingly challenging. The first part of the talk focuses on safety-critical learning and introduces a new control-theoretic framework for derivative-free optimization under unknown constraints. Designing safe zeroth-order (ZO) algorithms remains difficult because the gradient is generally estimated from zeroth order information and is very noisy. Leveraging tools from feedback linearization, we develop a family of ZO algorithms capable of guaranteeing constraint satisfaction using only noisy, sample-based gradient estimates. The second part of the talk turns to scalability. I will discuss how exploiting spatial structure in large-scale networked systems enables efficient and near-optimal decentralized control. Under mild assumptions, the optimal controller itself inherits a compatible spatial structure, leading to theoretical guarantees for distributed strategies and offering practical insights for coordinating large populations of agents.

Bio: Runyu (Cathy) Zhang is a postdoc researcher at MIT Laboratory for Information & Decision Systems. Before joining MIT, she obtained her Ph.D. degree at Harvard University, School of Engineering and Applied Sciences. Her research interests lie broadly in learning-based control, reinforcement learning, game theory and optimization, with particular focus on multi-agent systems. She has won the MIT Postdoctoral Fellowship for Engineering Excellence, was selected as the EECS rising star in 2024, and was a finalist of the Two Sigma Diversity PhD Fellowship in 2022.