Hou, Junyi (Zhejiang University), Zhu, Wenxin (Zhejiang University), Ren, Qinyuan (Zhejiang University)

Semantically-Aware Exploration in Zero-Shot Object Goal Navigation Via Bayesian Memory

Scheduled for presentation during the Invited Session "S20b-Foundation Model-Enabled Scene Understanding, Reasoning, and Decision-Making for Autonomous Driving and ITS" (TH-EA-T20), Thursday, November 20, 2025, 13:50−14:10, Surfers Paradise 2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Navigating in unfamiliar environments to an open-vocabulary target object which has never been seen before poses a significant challenge. This work proposes an efficient exploration strategy for Zero-Shot Object Goal Navigation (ZS-OGN) tasks which exploits spatial object-target relationships derived from temporal observations to guide exploration. The exploration strategy includes a fine-grained scene understanding module that enables the agent to extract comprehensive semantic relationships from open-vocabulary objects, followed by a robust Bayesian spatial memory that integrates temporal scene understanding results to maintain robust environmental representation in unfamiliar environments, thus enabling semantically-aware navigation decision making. By comprehensively scene understanding and robust memory constructing, the agent can navigate to regions most likely to contain the target object based on object relationships, thereby enhancing exploration efficiency. Evaluated on the HM3D and MP3D datasets, this strategy significantly outperforms baselines in both success rate and exploration efficiency. Comprehensive ablation studies further validate that scene understanding and spatial memory synergistically improve exploration efficiency.