DP-Nav: Dynamic Exploration Driven by Semantic Region Potential for Zero-shot Visual Navigation

ICLR 2026 Conference Submission22676 Authors

Keywords: visual navigation, large language model, path optimization

TL;DR: we propose DP-Nav, a novel dynamic navigation framework driven by the potential of semantic regions.

Abstract: Visual navigation requires the agent to autonomously navigate to a specified goal based on sequential visual perception. A key challenge is to achieve target localization and optimize the path simultaneously. However, most existing frontier-based methods rely on static navigation policies, which update the target frontiers at fixed time intervals to guide the agent's exploration. These approaches cannot dynamically assess potential regions encountered during navigation, thereby preventing timely policy adjustments. Moreover, the presence of multiple frontiers within the same region often leads to repeated exploration of identical regions, further exacerbating path redundancy and inefficiency. To address the above limitations, we propose DP-Nav, a novel dynamic navigation framework driven by the potential of semantic regions. Our approach first identifies distinct semantic regions from sequential visual perception and treats an independent semantic region as a policy unit. Furthermore, we introduce a Scoring-Screening Mechanism (SSM) that evaluates and filters these semantic regions based on their potential utility. Then SSM assigns exploration priorities to different regions, selecting the semantic region with the highest potential value for the agent's subsequent exploration. More significantly, we design a Dynamic Policy Trigger (DPT) module that enables on-demand activation of the SSM, allowing the agent to dynamically adapt its exploration policy in response to environmental changes and perceptual feedback, thereby addressing the rigidity of static policies. Extensive experiments on Object Goal Navigation, Text Goal Navigation, and Instance Image Goal Navigation across Gibson, HM3D, and MP3D datasets demonstrate that DP-Nav achieves SOTA performance and improves path efficiency by about $7\%\sim17\%$.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 22676