Colloquium Series: Towards Scalable Online Planning Under Uncertainty in Partially Observable Environments

Decision-making under uncertainty in partially observable environments is a central challenge in AI and robotics. It is fundamental to the operation of autonomous systems that must act reliably and safely despite incomplete, imperfect, or ambiguous information. This talk will present recent advances in scalable online planning in such settings, where the problem is commonly formulated as a Partially Observable Markov Decision Process (POMDP) or addressed through belief space planning. These include POMDP planning with hybrid belief representations, which capture real-world complexities such as viewpoint-dependent semantic observations and ambiguous data associations, as well as methods for leveraging past experience to accelerate online POMDP planning in continuous spaces significantly. The second part of the talk will focus on a recently developed framework for simplifying decision-making under uncertainty problems with formal performance guarantees. This framework will be presented from multiple perspectives, showing how it enables fast and adaptive online POMDP planning. Finally, its extension to multi-agent systems will be presented, highlighting how it supports provably consistent decision-making even when agents maintain inconsistent beliefs due to limited data-sharing.