Our research tackles sequential decision problems that classical methods alone cannot solve and pure learning methods alone cannot trust. We work in six tightly linked threads, each spanning theory and application.
Our methodological base is grounded in Markov decision processes, approximate dynamic programming, and multi-armed bandit models — and extends outward into deep learning, generative AI, and multi-agent systems.
Safe and multi-objective RL, risk- and longevity-aware agents, inverse RL, duality and occupancy-measure formulations, and general-utility RL objectives for scalable and reliable learning.
Team formation, hierarchical planning, test-time adaptation, and mechanism design for cooperative and self-interested agents.
Hierarchical representation learning, graph neural networks, constrained generative models, discrete diffusion for structured outputs, and out-of-distribution generalization.
Language-conditioned robot learning, multimodal manipulation, human-robot interaction, sim-to-real transfer, and long-horizon autonomy in the real world.
Vision-language-action models, LLM-based decision making, neurosymbolic AI, and interpretable methods for complex sequential decision-making tasks.
Online learning, test-time training, robust time-series analysis and classification, continual learning, and belief revision in dynamic environments.
We partner across disciplines and with industry to turn algorithmic advances into operational impact.
If our research overlaps with your problem, we want to hear about it — whether that's a new algorithmic question, a real-world deployment, or a prospective thesis topic.
Prospective PhD students, collaborators, and industry partners are always invited to reach out.