Bi-Level Reinforcement Learning for Multi-Robot Systems

I aim to build safe and collaborative multi-robot systems. My research so far has focused on bi-level optimization. In the constrained zero-sum case, I solved the reach-avoid problem, where one robot must reach a target area defended by another. In the unconstrained general-sum case, I am currently working on improving actor-critic reinforcement learning (RL) algorithms with low-rank approximations of the inverse-Hessian vector product to capture the dependency between actor and critic. I hope to extend my research to heterogeneous teams of robots by augmenting RL with classic control algorithms through differentiable programming and through continual multi-agent RL to organically learn diverse policies. These directions aim to advance scalable, safe, and collaborative AI for dynamic real-world environments.