Leveraging Unknown Objects to Construct Labeled-Unlabeled Meta-Relationships for Zero-Shot Object Navigation

Yanwei Zheng, Changrui Li, Chuanlin Lan, Yaling Li, Xiao Zhang, Yifei Zou, Dongxiao Yu, Zhipeng Cai·May 24, 2024

Summary

This paper contributes to the field of zero-shot object navigation (ZSON) by introducing a novel approach that addresses the limitations of prior works. The method incorporates unseen objects (unknown objects) into training through the Label-wise Meta-Correlation Module (LWMCM), which leverages relationships between labeled and unlabeled objects. The Target Feature Generator (TFG) and Unlabeled Object Identifier (UOI) work together to generate and recognize target representations, while the Meta Contrastive Feature Modifier (MCFM) adjusts features based on observed and unobserved objects. The proposed technique improves navigation performance in AI2THOR and RoboTHOR environments by utilizing both labeled and unlabeled data, resulting in enhanced ability to navigate towards both seen and unseen targets. Experiments demonstrate state-of-the-art performance and highlight the benefits of meta-learning and the integration of unknown objects for better generalization. The study also suggests that while the approach enhances unseen object navigation, it can have a negative impact on known object performance, indicating a need for further research on balancing the two.

Advanced features