Remote object navigation for service robots using hierarchical knowledge graph in human-centered environments

Remote object navigation (RON), defined as navigating to a remote object that is invisible in the current view, is an inevitable and extremely challenging task for a service robot, particularly when facing unstructured or dynamic human-centered environments. How to apply object-level semantic knowledge about the scene (called scene knowledge graph, SKG) to assist robots in cognition of the environment has become a hot research topic in robot intelligence. In this paper, we propose a knowledge-based RON method to skillfully combine the hierarchical knowledge in SKG and the probability-based navigation strategy. In detail, we first develop an automated pipeline to construct a novel SKG from massive visual data in real indoor environments. Then we propose a reasoner to derive the probabilistic representation of the hierarchical knowledge contained in the SKG. Additionally, a two-stage navigator composed of global path planning and local search strategy is applied as a distance-aware task planner to reduce the navigation path cost. The experimental results in real-world scenarios indicate that the proposed method has efficient performance and robustness on RON task compared to other approaches.