Searching from Superior to Inferior: A Hierarchical Relationship for Visual Navigation

Target-driven visual navigation is a challenging task for home service robots, where the robots lack prior information about the environment map and obtain environment information solely through the built-in camera. Home scenarios are characterized by complex spatial layouts, diverse room types, and concealed target locations. Due to the limited information available from the robot's built-in camera, navigating to unseen objects becomes more challenging. In this paper, we propose a superior-inferior relationship visual navigation model (SIR) to learn the dependency relationship among objects. By proposing a new reward function, SIR provides spatial relationship information among objects. In this way, the visible object can be regarded as a reference for navigating to unseen objects. Additionally, our method to some extent solves the problem of sparse rewards in navigation tasks. The robot not only receives positive rewards when it successfully finds the target, but also receives partial rewards when it detects superior objects. It also enhances the interpretability of the visual navigation deep reinforcement learning network. Experimental results in the AI2Thor environment demonstrate that our SIR achieve a 6.5% gain on success rate and gain 13.5% improvements on long episodes environment than the baseline method.