ReMemNav: Memory-Based Decision Correction and Target Verification for Zero-Shot Object Navigation

Feng Wu; Wei Zuo; Wenliang Yang; Jun Xiao; Yang Liu; Xinhua Zeng

cs.ROMar 25, 2026

ReMemNav: Memory-Based Decision Correction and Target Verification for Zero-Shot Object Navigation

Authors: Feng Wu, Wei Zuo, Wenliang Yang, Jun Xiao, Yang Liu, Xinhua Zeng

Organizations: Fudan University · Tongji University

Abstract

Zero-shot object navigation requires agents to locate unseen targets in unfamiliar environments without prior maps or task-specific training. Despite the commonsense reasoning ability of vision-language models (VLMs), existing mapless navigators often suffer from repeated exploration and premature stopping caused by limited historical context and false-positive target predictions. We propose ReMemNav, a training-free framework that combines lightweight semantic grounding, memory-based decision correction, and target verification. RAM-derived semantic priors guide panoramic direction selection, while a geometry-triggered correction mechanism retrieves historical scene descriptions when the selected direction points toward previously visited regions. When a target is predicted to be visible, ReMemNav verifies the corresponding single-view observation before the final approach. Depth-based action sampling then converts the high-level decision into collision-free motion. Experiments on HM3D and MP3D show that ReMemNav achieves higher success rates and path efficiency than existing training-free zero-shot baselines. Specifically, ReMemNav achieves absolute SR/SPL gains of 1.7%/5.3% on HM3D v0.1, 18.2%/11.1% on HM3D v0.2, and 8.7%/6.9% on MP3D.

Figures & tables

Figure 1: Illustration of ReMemNav for correcting repeated exploration and verifying potential target observations.

Figure 2: Overview of ReMemNav. RAM provides lightweight semantic grounding for the initial panoramic navigation decision. When the selected direction overlaps with a previously explored region, memory-based decision correction retrieves historical semantic context and revises the decision. When the target is predicted to be visible, target verification re-examines the corresponding observation before the final approach.

Figure 3: VLM-guided low-level action selection. Collision-free candidate action rays are generated from the depth observation and overlaid on the RGB image, after which the VLM selects the action that best follows the finalized high-level direction.

Methods	TF	ZS	MP3D		HM3D_v0.1		HM3D_v0.2
Methods	TF	ZS	SR	SPL	SR	SPL	SR	SPL
Habitat-Web ( Ramrakhya et al. 2022 )	✗	✗	31.6	8.5	41.5	16.0	–	–
OVRL ( Yadav et al. 2023b )	✗	✗	28.6	7.4	–	–	–	–
ZSON ( Majumdar et al. 2022 )	✗	$\checkmark$	15.3	4.8	25.5	12.6	–	–
PixNav ( Cai et al. 2024 )	✗	$\checkmark$	–	–	37.9	20.5	–	–
PSL ( Sun et al. 2024 )	✗	$\checkmark$	18.9	6.4	42.4	19.2	–	–

Table 1: Comparison with representative object navigation methods on HM3D and MP3D. TF refers to training-free, and ZS refers to zero-shot.

Figure 4: Qualitative comparison of navigation behaviors on the same episode. In the top row, the ablated agent without target verification incorrectly identifies a chair as the target sofa. It consequently approaches the false-positive target and stops prematurely. In contrast, full ReMemNav re-examines the suspected target and correctly rejects the chair as a false positive. The agent then continues exploring the environment and eventually reaches a valid sofa instance, completing the episode successfully.

Method	SR (%) $\uparrow$	SPL (%) $\uparrow$
VLMNav	41.5	11.7
UniGoal	43.3	21.7
InstructNav	53.5	21.1
ReMemNav (Ours)	56.2	21.8

Table 2: Comparison under a common VLM backbone. All methods are evaluated with Qwen3-VL-4B on HM3D v0.2 under the same benchmark configuration.

Variant	SR (%) $\uparrow$	SPL (%) $\uparrow$
Full ReMemNav	56.2	21.8
w/o Semantic Grounding	55.4	21.3
w/o Memory-Based Correction	52.8	21.4
w/o Target Verification	45.9	21.0

Table 3: Module-wise ablation on HM3D v0.2. Semantic grounding is implemented using RAM. Memory-based correction jointly includes the bounded episodic memory and history-conditioned decision revision.

Robot Navigation

Robot navigation studies how robots perceive, localize, plan, and move safely toward goals in static or changing environments.

See all 536 papers

MarvisNav: Making Memory Visible on Route Choices for Zero-Shot Object Navigation
Oct 5, 2026Jincheng Wang, Chi Pui Chan, Wei Zeng +3Memory-Augmented VLMs Vision-Language Navigation
University College London, University of London · China Merchants Group, LionRock AI Lab · Shenzhen University
SR-Nav: Spatial Relationships Matter for Zero-shot Object Goal Navigation
Mar 19, 2026Leyuan Fang, Zan Mao, Zijing Wang +1Object-Goal Navigation Robot Navigation
School of Artificial Intelligence and Robotics, Hunan University, Changsha, China, 410114, China
MCNav: Memory-Aware Dynamic Cognitive Map for Zero-shot Goal-oriented Navigation
May 19, 2026Jingyu Li, Zhe Liu, Wenxiao Wu +1Embodied Navigation Object-Goal Navigation
Fudan University · Shanghai Innovation Institute · University of Hong Kong +1

Oct 5, 2026J/K move · Enter open · S save

MarvisNav: Making Memory Visible on Route Choices for Zero-Shot Object Navigation

Jincheng Wang, Chi Pui Chan, Wei Zeng, Shuyang Zhang+2

University College London, University of London · China Merchants Group, LionRock AI Lab · Shenzhen University

Memory-Augmented VLMs Vision-Language Navigation Object-Goal Navigation

Open paper AskPDF arXiv

When searching for an object, people choose their next move by considering both likely target locations and places already explored. The current view can cue place-associated memories, bringing target relevance and prior exploration into the same spatial context. In many zero-shot object navigation (ZSON) methods, however, vision-language models (VLMs) infer promising search areas from egocentric images, while exploration history is represented separately, e.g., as text or maps. This separation either requires an additional fusion step or leaves the correspondence between memory and route choices implicit for the VLM to recover. We instead make exploration memory directly visible on visual route choices. We propose MarvisNav, a ZSON framework that maintains a topological graph and projects candidate nodes together with their exploration states onto egocentric views as memory-bearing visual route choices. These states capture local exploration progress beyond binary visitation. By binding exploration state directly to each visual candidate, MarvisNav enables the VLM to jointly evaluate target relevance and exploration state without a separate post-hoc fusion or reranking stage. Without policy training, MarvisNav achieves state-of-the-art performance on HM3D (81.2% SR and 42.5% SPL), while remaining competitive on MP3D. It also outperforms representative VLM-based methods with far fewer VLM calls (e.g., 7.5% of WMNav). Real-robot experiments across diverse scenes further validate its practical deployability. Beyond MarvisNav, our study shows that memory representation shapes VLM decisions and ZSON performance, highlighting that effective memory use depends not only on its availability, but also on how it is represented. Code and project page will be available at https://wangjincheng1998.github.io/MarvisNav/.

Abstract

Robot Navigation

Explore similar work

MarvisNav: Making Memory Visible on Route Choices for Zero-Shot Object Navigation

SR-Nav: Spatial Relationships Matter for Zero-shot Object Goal Navigation

MCNav: Memory-Aware Dynamic Cognitive Map for Zero-shot Goal-oriented Navigation