Title: Vectorized Online POMDP Planning

Abstract: For autonomous robots, the ability to plan under conditions of partial observability is critical. The Partially Observable Markov Decision Process (POMDP) offers a robust framework for addressing these challenges by accounting for the randomness inherent in actions and the constraints imposed by noisy, limited observations. While modern hardware capabilities suggest that massive parallelization could significantly enhance POMDP solving, achieving this has proven difficult. Traditional solvers typically alternate between numerical optimization of actions and value estimation, a process that generates dependencies and synchronization hurdles among parallel processes, often negating the advantages of parallel execution.

To address these issues, we introduce the Vectorized Online POMDP Planner (VOPP), a new parallel online solver. VOPP utilizes a recent POMDP formulation that analytically resolves a portion of the optimization task, thereby restricting numerical computations to the estimation of expectations alone. By representing all planning-related data structures as collections of tensors and executing all planning steps as fully vectorized operations on this representation, VOPP achieves massive parallelism without the dependencies or synchronization bottlenecks that plague concurrent processes. Our experimental findings demonstrate that VOPP computes near-optimal solutions with at least 20 times greater efficiency than the current leading parallel online solver. Furthermore, VOPP surpasses state-of-the-art sequential online solvers while operating with a planning budget that is 1,000 times smaller.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC