EvolveNav: Proactive Preflection and Self-Evolving Memory for Zero-Shot Object Goal Navigation

A new self-evolving framework for Zero-Shot Object-Goal Navigation enables embodied agents to improve continuously during testing, according to research posted on arXiv. The method, called EvolveNav, builds a memory of past actions and forecasts outcomes to reduce inefficient exploration. Zero-Shot Object-Goal Navigation requires agents to locate target objects in unfamiliar environments without any prior training [1][2]. Existing approaches often use foundation models but depend on static priors that do not adapt, causing repeated errors and costly trial-and-error [1][2]. The EvolveNav framework addresses this by constructing an agentic rule memory that extracts actionable knowledge from past trajectories [1][2]. A retrieval strategy based on upper confidence bound selects effective rules by balancing semantic relevance and historical success [1][2]. The system also includes a memory-guided preflection module that forecasts potential outcomes before an action is taken, which the authors say reduces inefficient exploration [1][2]. In experiments, EvolveNav outperformed existing zero-shot baselines, achieving a 10.1% improvement in success rate while taking fewer unnecessary steps [1][2]. The paper was submitted to arXiv on 16 June 2026 [1]. arXiv is an open-access repository of electronic preprints that, as of late 2024, receives about 24,000 submissions per month and has surpassed two million articles [6]. The repository hosts papers across mathematics, physics, computer science, and other fields, and its content is moderated but not peer-reviewed [6]. The arXivLabs framework, which appears on the article record page, allows community collaborators to develop and share experimental tools that add functionality for readers and authors [4][5]. Current arXivLabs projects include the Bibliographic Explorer, which displays citation-tree information, and the CORE Recommender, which surfaces relevant open-access papers from a global network of repositories [4][5]. arXiv has stated that third-party collaborators receive only minimal and anonymized user data, and any other use is prohibited without written consent [4]. The arXivLabs program is currently pausing new proposals while the development team focuses on modernizing arXiv's infrastructure and moving systems to the cloud [3].