This paper presents an online planning framework and a novel benchmark dataset for solving the multi-object relocation problem in partially observable multi-spatial environments. Existing object relocation solutions based on reinforcement learning or manually coded planning methods often lack adaptability to diverse problems. To address these limitations, this paper proposes a Hierarchical Object-Oriented Partially Observable Markov Decision Process (HOO-POMDP) planning approach. This approach consists of (a) an object-oriented POMDP planner that generates subgoals, (b) a set of low-level policies for achieving the subgoals, and (c) an abstraction system that transforms the continuous low-level world into a representation suitable for abstract planning. To enable a rigorous evaluation of the relocation problem, we present MultiRoomR, a comprehensive benchmark featuring diverse multi-spatial environments (partial observability of 10-30%, blocked paths, occluded targets, and 10-20 objects distributed across 2-4 rooms). Experimental results demonstrate that the proposed system effectively handles these complex scenarios despite imperfect perception and achieves promising results on both existing benchmarks and the novel MultiRoomR dataset.
