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

Abstract

Zero-Shot Object-Goal Navigation (ZS-OGN) requires embodied agents to explore and locate target objects without any prior training. To this end, recent methods leverage foundation models. But they typically rely on static priors and lack adaptation, which leads to repeated errors and costly trial and error. In this paper, we propose a self-evolving ZS-OGN framework that enables continuous test-time improvement. Specifically, we build an agentic rule memory by extracting actionable knowledge from past trajectories. Then, we propose a retrieval strategy based on upper confidence bound, selecting effective rules by balancing semantic relevance and historical success. In addition, we introduce a memory-guided preflection module that forecasts potential outcomes before action, reducing inefficient exploration. Extensive experiments show that our method outperforms existing training-free baselines, achieving a 10.1% improvement in success rate with fewer unnecessary steps.

Method

EvolveNav transforms training-free navigation into a continuous self-improvement process. The framework couples real-time spatial exploration with a self-evolving cognitive memory through two tightly integrated components.

Intra-Episode Preflective Exploration

Within an episode, the agent builds a 2D occupancy map and extracts candidate frontiers. Before moving, an LLM retrieves past empirical rules from the rule memory and performs preflection to evaluate the risks of each frontier and select the optimal direction. This shifts navigation from passive correction to proactive risk avoidance.

Inter-Episode Rule Self-Evolution

After each episode, the agent reviews its trajectory and distills key decisions into generalizable rules. A UCB-based algorithm balances exploitation of proven rules with exploration of newly formulated ones, using a momentum-smoothed expected utility μ_r and confidence bound.

Demo Videos

Qualitative navigation rollouts of EvolveNav across diverse indoor scenes. Each clip shows the agent's first-person view.

1 / 12

Quantitative Results

EvolveNav achieves state-of-the-art training-free performance on both HM3D and MP3D benchmarks, surpassing prior training-free methods by 1.9% (HM3D) and 4.5% (MP3D) in Success Rate.

**Comparisons with state-of-the-art methods.** The table compares learning-based and training-free methods on HM3D and MP3D across *Success Rate* (SR) and *Success weighted by Path Length* (SPL). Higher is better (↑). Gray rows are learning-based; the highlighted row is ours.
Method	Venue	Training-free	Instruction Interpolator		HM3D		MP3D
Method	Venue	Training-free	Vision	Language	SR ↑	SPL ↑	SR ↑	SPL ↑
RIM	IROS'23	✗	–	–	57.8	27.2	50.3	17.0
OVG-Nav	RAL'24	✗	–	–	–	–	35.8	12.3
VLFM	ICRA'24	✗	BLIP-2	–	52.5	30.4	36.4	17.5
PIRLNav	CVPR'23	✗	–	–	64.1	27.1	–	–
XGX	ICRA'24	✗	–	–	72.9	35.7	–	–

ZSON	NeurIPS'22	✓	CLIP	–	25.5	12.6	15.3	4.8
L3MVN	IROS'23	✓	–	RoBERTa-large	50.4	23.1	34.9	14.5
PixNav	ICRA'24	✓	LLaMA-Adapter	GPT-4	37.9	20.5	–	–
VLFM	ICRA'24	✓	BLIP-2	–	50.9	23.6	32.5	15.9
SG-Nav	NeurIPS'24	✓	LLaVA-1.6-7B	GPT-4	54.0	24.9	40.2	16.0
OpenFMNav	NAACL-F'24	✓	GPT-4V	GPT-4	54.9	24.4	37.2	15.7
ActPept	RAL'24	✓	GraphSAGE	–	–	–	39.8	17.4
InstructNav	CoRL'24	✓	GPT-4V	GPT-4	58.0	20.9	–	–
ApexNav	RAL'25	✓	BLIP-2	DeepSeek-V3	59.6	33.0	39.2	17.8
MFNP	ICRA'25	✓	Qwen-VLChat-Int4	Qwen2-7B	58.3	26.7	41.1	15.4
MSGNav	CVPR'26	✓	GPT-4o	GPT-4o	63.0	31.4	–	–
ASCENT	RAL'26	✓	BLIP-2	Qwen2.5-7B	65.4	33.5	44.5	15.5
EvolveNav (Ours)	–	✓	BLIP-2	Qwen3-8B	67.3	33.9	49.0	19.1

**Ablation on core modules.** We ablate the Preflection and Memory-Evolving modules on HM3D and MP3D.
#	Module		HM3D		MP3D
#	Preflection	Memory-Evolve	SR ↑	SPL ↑	SR ↑	SPL ↑
1	–	–	64.7	32.5	43.9	15.8
2	✓	–	66.5	33.5	47.4	18.4
3	–	✓	66.7	33.6	48.3	18.7
4	✓	✓	67.3	33.9	49.0	19.1

**Ablation on LLM backbone.** Vision model fixed to BLIP-2. EvolveNav is robust across LLM scales.
Instruction Interpolator		HM3D		MP3D
Vision	Language	SR ↑	SPL ↑	SR ↑	SPL ↑
BLIP-2	Qwen2.5-7B	67.0	33.9	48.8	19.0
	Qwen3-8B	67.3	33.9	49.0	19.1
	Qwen3.5-9B	67.2	34.1	48.9	19.2

Qualitative Results

EvolveNav navigation case study — **Navigation process visualization of EvolveNav.** We show the agent's trajectory alongside its reasoning. During the episode, the agent uses Preflection to evaluate frontiers and proactively avoid unpromising rooms. After the episode, EvolveNav generates a new rule from the history and updates its Rule Bank, autonomously distilling long-term navigation knowledge from short-term trial and error.

Trajectory comparison with ASCENT — **Qualitative comparison.** Start points, paths, and destinations of EvolveNav and ASCENT on several MP3D scenes. ASCENT's trajectory is notably more convoluted, frequently distracted by adjacent small rooms and narrow corridors. In contrast, EvolveNav produces more direct, globally optimal routes.

BibTeX

@misc{evolvenav2026,
  title         = {EvolveNav: Proactive Preflection and Self-Evolving Memory
                   for Zero-Shot Object Goal Navigation},
  author        = {Chai, Qi and Shen, Wenhao and Yao, Nanjie and Xia, Yue and
                   Zhao, Kaiyong and Ma, Jie and Lin, Guosheng and Wang, Hao},
  year          = {2026},
  eprint        = {2606.18235},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}