AnyImageNav: Any-View Geometry for Precise Last-Meter Image-Goal Navigation

Yijie Deng; Shuaihang Yuan; Yi Fang

cs.ROApr 7, 2026

AnyImageNav: Any-View Geometry for Precise Last-Meter Image-Goal Navigation

Authors: Yijie Deng, Shuaihang Yuan, Yi Fang

Organizations: NYUAD Center for Artificial Intelligence and Robotics (CAIR), Abu Dhabi, UAE · New York University Abu Dhabi, Electrical Engineering, Abu Dhabi 129188, UAE · New York University, Electrical & Computer Engineering Dept., Brooklyn, NY 11201, USA. · Embodied AI and Robotics (AIR) Lab, NYU Abu Dhabi, UAE.

Abstract

Image Goal Navigation (ImageNav) is evaluated by a coarse success criterion, the agent must stop within 1m of the target, which is sufficient for finding objects but falls short for downstream tasks such as grasping that require precise positioning. We introduce AnyImageNav, a training-free system that pushes ImageNav toward this more demanding setting. Our key insight is that the goal image can be treated as a geometric query: any photo of an object, a hallway, or a room corner can be registered to the agent's observations via dense pixel-level correspondences, enabling recovery of the exact 6-DoF camera pose. Our method realizes this through a semantic-to-geometric cascade: a semantic relevance signal guides exploration and acts as a proximity gate, invoking a 3D multi-view foundation model only when the current view is highly relevant to the goal image; the model then self-certifies its registration in a loop for an accurate recovered pose. Our method sets state-of-the-art navigation success rates on Gibson (93.1%) and HM3D (82.6%), and achieves pose recovery that prior methods do not provide: a position error of 0.27m and heading error of 3.41 degrees on Gibson, and 0.21m / 1.23 degrees on HM3D, a 5-10x improvement over adapted baselines.Our project page: https://yijie21.github.io/ain/

Figures & tables

Figure 1: AnyImageNav overview. A BEV relevance map identifies highly relevant regions and guides the agent toward the goal vicinity ( Pre-Locate ). Once a highly relevant region is detected, any-view geometry registers the goal image with previous observations to recover a precise target pose ( Last-Meter Localization ). The agent then navigates to the pose that closely matches the goal viewpoint ( Navigate to Goal ).

Figure 2: AnyImageNav pipeline. (a) Frontier Exploration: The dense relevance map between the current observation and the goal image $I_{g}$ is projected onto a BEV map, scoring candidate frontiers and yielding the proximity score $\mathcal{S}_{\text{relev}}$ . (b) Geometric Verification: When $\mathcal{S}_{\text{relev}}>\theta$ , a short memory window and $I_{g}$ are passed to a 3D multi-view foundation model; the model’s internal correspondence confidence $\mathcal{S}_{\text{conf}}$ gates the transition to localization. (c) Last-Meter Localization: Upon acceptance, the full long-memory cache is registered with $I_{g}$ in a single forward pass; the recovered goal pose is aligned to the agent’s map via Sim(3)+orientation correction and continuously refined as confidence improves.

Figure 3: BEV Relevance Map Construction. A dense pixel-level relevance map between the current observation and the goal image is calculated per step. Then a frustum-aware projection then maps the relevance map onto the top-down BEV grid: a 2D cone mask is combined with depth-contoured truncation to prevent relevance from bleeding through obstacles, and element-wise multiplied with a confidence cone that attenuates rays toward the field-of-view boundary. The resulting per-step relevance cone is accumulated into a persistent global BEV relevance map.

Figure 4: Navigation examples on HM3D (top) and Gibson (bottom). In both episodes, the BEV relevance map directs the agent toward goal-relevant regions well before the target is reached (shown at $t=70$ ). At $t=111$ (HM3D) and $t=135$ (Gibson), the target pose is accurately recovered while the agent is still approaching, and the agent subsequently navigates to the estimated goal position with high precision.

Table 5

Table 6

Image Matching

Image matching finds corresponding points, regions, or objects across images despite changes in viewpoint, scale, lighting, or sensor.

See all 83 papers

Think before Go: Hierarchical Reasoning for Image-goal Navigation
Apr 19, 2026Pengna Li, Kangyi Wu, Shaoqing Xu +5Visual Navigation Robot Policy Learning
National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University · University of Macau · School of Automation, Beijing Institute of Technology +1
G $^2$ WAM: World Action Models for Efficient and Geometry Grounded Image-Goal Navigation
Oct 12, 2026Ke Gui, Chenghao Shi, Ziyue Wang +3Scene Coordinate Regression Image-Goal Navigation
College of Intelligence Science and Technology, National University of Defense Technology, China · National Key Laboratory of Equipment State Sensing and Smart Support, National University of Defense Technology, China · School of Future Technology, National University of Defense Technology, China
StopWatch: Goal Grounding for Object-goal Navigation
Oct 16, 2026Xinhao Liu, Yunsheng Tian, Yuheng Qiu +4Object-Goal Navigation Visual Grounding for Robotics
Amazon FAR · New York University · UC Berkeley +1

Apr 19, 2026J/K move · Enter open · S save

Think before Go: Hierarchical Reasoning for Image-goal Navigation

Pengna Li, Kangyi Wu, Shaoqing Xu, Fang Li+4

National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University · University of Macau · School of Automation, Beijing Institute of Technology+1

Visual Navigation Robot Policy Learning Robot Navigation Image-Goal Navigation

Open paper AskPDF arXiv

Image-goal navigation steers an agent to a target location specified by an image in unseen environments. Existing methods primarily handle this task by learning an end-to-end navigation policy, which compares the similarities of target and observation images and directly predicts the actions. However, when the target is distant or lies in another room, such methods fail to extract informative visual cues, leading the agent to wander around. Motivated by the human cognitive principle that deliberate, high-level reasoning guides fast, reactive execution in complex tasks, we propose Hierarchical Reasoning Navigation (HRNav), a framework that decomposes image-goal navigation into high-level planning and low-level execution. In high-level planning, a vision-language model is trained on a self-collected dataset to generate a short-horizon plan, such as whether the agent should walk through the door or down the hallway. This downgrades the difficulty of the long-horizon task, making it more amenable to the execution part. In low-level execution, an online reinforcement learning policy is utilized to decide actions conditioned on the short-horizon plan. We also devise a novel Wandering Suppression Penalty (WSP) to further reduce the wandering problem. Together, these components form a hierarchical framework for Image-Goal Navigation. Extensive experiments in both simulation and real-world environments demonstrate the superiority of our method.

Abstract

Image Matching

Explore similar work

Think before Go: Hierarchical Reasoning for Image-goal Navigation

G2^22WAM: World Action Models for Efficient and Geometry Grounded Image-Goal Navigation

StopWatch: Goal Grounding for Object-goal Navigation

G $^2$ WAM: World Action Models for Efficient and Geometry Grounded Image-Goal Navigation