Home ****Team Research Publications Teaching

We aim to build multi-agent systems for the real world.

Our research focuses on machine learning (especially reinforcement learning), with applications to robotic systems that integrate high-level planning with low-level control, addressing real-world challenges such as generalization, scalability, robustness, safety, and fairness.

Our current research interests include:

Multi-Agent Learning: Learning how agents interact, coordinate, and adapt in cooperative, competitive, and mixed-motive settings.

• Cooperation*:*** Learning how agents coordinate their actions to achieve shared goals. Developing multi-agent reinforcement learning methods for coordinated exploration, decentralized decision-making, and communication.
• Competitive and Mixed-Motive Games: Learning how agents make decisions when their interests conflict or only partially align. Using game-theoretic approaches to address strategic adaptation, incentives, and fairness.
• Data-Driven Learning: Learning multi-agent policies from existing data through offline reinforcement learning and imitation learning. Improving these policies through offline-to-online adaptation to account for interactions with other agents.
• Generalization and Robustness: Developing policies that generalize across teammates, team compositions, and environments, including zero-shot coordination with unseen partners. Maintaining effective coordination under incomplete information, communication failures, and unexpected disruptions.

Learning for Robotic Systems: Developing learning methods for robotic systems, including multi-robot teams, that connect planning and coordination with navigation, manipulation, and control.

• Task Allocation and Multi-Agent Path Planning: Learning policies that assign tasks and coordinate collision-free movement in large robot teams. Emphasizing generalization across team sizes and environments, and adaptation to changing tasks and disruptions.
• Navigation and Information Gathering: Learning to use observations and prior knowledge for navigation, inspection, and monitoring. Extending these capabilities to robot teams that share information and coordinate their actions.
• Robot Learning: Using real-world data to inform simulation and learn policies that transfer back to physical robots, with an emphasis on generalization and recovery from failures.
• Agentic Robotics: Connecting language-based reasoning with perception, planning, and learned robot skills to carry out complex tasks. Using simulation and execution feedback to improve task plans, skill selection, and recovery from failures.

We are not limited to this. We are always open to interesting problems.

We welcome prospective interns and M.S./Ph.D. students interested in (a) multi-agent learning for robotic systems, (b) robot learning, and (c) multi-agent orchestration.

• Awards and Honors
• Projects