Object-aware navigation for remote embodied visual referring expression

In the Remote Embodied Visual Referring Expression (REVERIE) task, an agent needs to navigate through an unseen environment to identify a referred object following high-level instructions. Despite recent efforts of vision-and-language navigation (VLN), previous methods commonly rely on detailed navigational instructions, which might not be available in practice. To address this issue, we present a method that strengthens vision-and-language (V&L) navigators with object-awareness. By combining object-aware textual grounding and visual grounding operations, our technique helps the navigator recognize the relationship between instructions and the contents of captured images. As a generic method, the proposed solution can be seamlessly integrated into other V&L navigators with different frameworks (for example, Seq2Seq or BERT). In order to alleviate the problem of data scarcity, we synthesize augmented data based on a simple yet effective prompt template that retains object information and destination information. Experimental results on REVERIE and R2R datasets demonstrate the proposed methods’ applicability and performance improvement across different domains.