everyone
since 04 Oct 2024">EveryoneRevisionsBibTeXCC BY 4.0
Efficiently searching for target objects in intricate environments poses a significant challenge for mobile robots, due to perception errors, limited field of view (FOV), and visual occlusion. These factors cause the problem to be partially observed. Therefore, we formulate the object-search task as a high-dimensional Partially Observable Markov Decision Process (POMDP) with hybrid (continuous and discrete) action spaces. We propose a novel sampling-based online POMDP solver named Neural Process Filtered $k$-Center Clustering Tree (NPF-$k$CT). The optimal action is selected using Monte Carlo Tree Search (MCTS) in conjunction with a neural process network to filter out ineffective primitive actions (i.e., basic robot operations), alongside $k$-center clustering hypersphere discretization to efficiently refine high-dimensional continuous sub-action spaces. Adhering to the hierarchical optimistic optimization (HOO) concept, we leverage an upper-confidence bound (UCB) on the action value function within the hypersphere with estimated diameters to guide the MCTS expansion. We extensively tested our approach in Gazebo simulations using Fetch and Stretch robots across diverse target-finding scenarios. Comparative results show higher success rates and faster target detection than baseline methods, with no additional computational cost. We also validated our method on a physical robot in an office environment. Project page: \url{https://sites.google.com/view/npfkct}.