Socially Aware Object Goal Navigation With Heterogeneous Scene Representation Learning

Socially aware Object Goal Navigation (ObjectNav) requires robots to navigate to objects with specific semantic categories while understanding complex human social awareness and semantic co-occurrence relations among objects. Existing solutions usually achieve scene representation by mapping these complicated Human-Robot-Object (HRO) mutual interactions to the same feature space. However, this homogeneous feature modeling may result in a loss of feature specificity. We argue that humans, robots, and objects have different interactive paradigms with each other, which should be represented separately and elaborately. Therefore, this letter proposes a novel Heterogeneous Scene Representation (HSR) method to learn HRO ternary interaction features. In particular, a novel Heterogeneous Graph Attention Network (HGAN) is proposed to exclusively model different interactions and semantic relations so that they maintain their essential properties. Further, a Deep Reinforcement Learning (DRL) based socially aware ObjectNav strategy is proposed by learning HSR-based scene state transition and state value estimation. The feasibility and superiority of our method are verified through sufficient baseline tests and ablation studies. Extensive comparative studies show that our method outperforms existing solutions in challenging domestic crowded scenes.