Difference-Oriented Network for Enhanced Zero-Shot Object Goal Visual Navigation

Object-goal visual navigation, where an intelligent agent locates a target object using only egocentric RGB observations, presents significant challenges in generalizing across unseen environments and object categories. Existing methods often struggle to learn domain-independent visual representations, which hinders their ability to perform effectively in novel scenarios. To address these limitations, this letter introduces the Difference-Oriented Network (DONet), a new framework that leverages object-level differences between the target object and all other observed objects to enhance zero-shot visual navigation. DONet incorporates a novel difference-oriented contrastive learning mechanism to extract fine-grained relational information and employs separation and aggregation operations to process object features individually, enabling a holistic understanding of scene information. Trained end-to-end with deep reinforcement learning, DONet is evaluated in the AI2-THOR simulation environment. Experimental results demonstrate that DONet achieves superior performance in terms of success rate (SR) and success weighted by path length (SPL) compared to state-of-the-art methods in zero-shot tasks involving unseen objects. These results highlight DONet’s strong generalization capabilities, making it a promising approach for real-world object-goal visual navigation in unfamiliar environments with unseen objects.