Building Safer Encounters Between Humans and Autonomous Systems - Profile Jingqi Li

Oftentimes, coordination between people is not spoken out loud. For example, envision two people moving a table through a doorway. As they maneuver, each person slows down and adjusts positions naturally by feeling each other’s movements. Moments like these leave Peter O’Donnell, Jr. Postdoctoral Fellow Jingqi Li wondering whether a robot could ever do the same: read intentions the way people do, noticing obstacles and signaling its own plans back without being told.

The question of how machines coordinate as naturally as people do sits at the heart of Li's work as a second-year Fellow in the Center for Autonomy at the Oden Institute for Computational Engineering and Sciences, where he works with David Fridovich-Keil, Oden Institute core faculty member and assistant professor of aerospace engineering & engineering mechanics.

Li was trained in optimization, control theory and machine learning, and he was drawn to research by problems where mathematics could explain how complex systems behave. While earning his Ph.D. at the University of California, Berkeley, which he completed in 2025, he became especially interested in autonomous systems that make decisions while interacting with others. This led him toward game theory, optimization, and deep reinforcement learning as tools for understanding strategic interactions.

Today, his research asks how autonomous systems should act when the other decision makers around them have goals, information, or future plans that are uncertain. Self-driving cars offer the most intuitive example. "A self-driving car has to anticipate whether another driver will yield, merge, or behave unexpectedly while still making safe and efficient decisions itself," Li said. Researchers call these decision makers agents, a category that can include a self-driving car, a robot, a generative language model, or a human driver. He develops algorithms that help agents anticipate one another, learn from those interactions, and act safely in response.

The hard part, he says, is making the strategic reasoning computable when many agents are interacting, each learning as it goes and each acting with an incomplete picture of the others. Researchers still need to understand what information agents really need about one another, how accurately they have to predict each other's behavior, and how safety guarantees can hold up as interactions grow more complex. "I am especially interested in uncovering the mathematical structure that makes these otherwise difficult problems tractable," he said.

The search for structure is something Li says his collaborators push him toward. Getting a method to work is only the first step; the next is understanding the underlying framework of the problem. "David has been particularly influential in encouraging me to look for simple mathematical representations that expose what is essential in a complicated strategic interaction," Li said.

The payoff could reach well beyond the road. As autonomous systems increasingly share space and resources with people and with one another, safety depends not just on each machine behaving correctly, but on how they respond to each other. The same challenge appears in robotics, transportation networks, advanced air mobility (an emerging area of aviation that includes delivery drones and air taxis), and infrastructure systems, where better strategic reasoning could make autonomy safer, more efficient, and more trustworthy.

The interdisciplinary nature of Li’s work is part of what brought him to the Oden Institute, which he describes as a place that brings together mathematics, computation, engineering, and real-world applications. "My research sits exactly at that intersection," he said. What surprised him once he arrived was how easily conversations moved between fields. 

"A question that begins in robotics can quickly turn into a discussion about optimization, machine learning, numerical methods, or even a completely different application area," he said. "That kind of environment is very valuable to me because many of the problems I focus on do not fit neatly within one field."

For the rest of his fellowship, Li plans to dig deeper into the fundamental questions behind how autonomous systems interact, while making sure the methods he develops work on physical robots, not just on paper. Longer term, he hopes to lead his own academic research group spanning game theory, optimization, machine learning, and robotics. One direction he is particularly excited about is interaction among multiple AI agents. "Even if each AI agent is well designed individually, their interactions could produce unexpected or harmful behavior that no one intended," he said. "I want to understand what fundamental rules and safeguards these systems should follow so that their collective behavior remains predictable, beneficial, and under human control."

Moving furniture isn’t the only everyday example that turns into research material for Li. Teaching prompts the same kind of thinking. A classroom, he notes, is shaped by information asymmetry: instructors know what they plan to teach, while students may not know what to expect. At the same time, instructors have limited knowledge of each student’s prior understanding and learning needs. Effective teaching therefore requires anticipating how different students learn, adapting instruction accordingly, and encouraging students to learn from one another. “Everyday experiences like these often become starting points for how I think about information, communication, and coordination between intelligent agents,” he said.

Outside the lab, Li has been settling into Austin, which he says has a very different pace and atmosphere from the Bay Area while still offering a strong university and tech community. He has been gradually exploring the city's neighborhoods and coffee shops, finding places to work, read, or simply spend time away from campus.

For anyone considering a similar path, Li recommends cultivating both depth and breadth: knowing one area well enough to understand its foundations, while staying open to the ideas that emerge when tools from different fields meet. He also suggests paying close attention to results that don't match expectations. "When a hardware experiment or simulation does not behave as expected," he said, "that is often where the most interesting question begins."

More information on the Peter O'Donnell, Jr. Postdoctoral Fellowship, including application deadlines, can be found here.