ImageNav: Image-Goal Navigation

• ImageNav is an embodied navigation task where an agent uses a goal image to determine actions instead of coordinates or textual instructions.
• It spans various methodologies including discrete and continuous control, fine-grained goal prompting, and explicit geometric registration to ensure accurate goal alignment.
• Research in ImageNav integrates goal-conditioned perception, memory augmentation, and topological mapping to enhance navigation efficiency and success across simulated and real-world settings.

Searching arXiv for recent and foundational ImageNav papers to ground the article and verify relevant citations. Image-goal navigation (ImageNav) is an embodied navigation task in which an agent must act from onboard perception to reach a location specified by a goal image rather than by coordinates, a category label, or a natural-language instruction. In the standard formulation, the agent is given a fixed goal image and must select actions so that its future observation matches that goal; in newer formulations, the goal may instead denote a specific object instance, a free-view photograph taken with a different camera, or an exact 6-DoF target pose required for downstream manipulation (Mezghani et al., 2021, Krantz et al., 2022, Deng et al., 7 Apr 2026). The literature now spans RGB-only and RGB-D settings, discrete and continuous control, explicit and implicit memory, offline sequence modeling, modular planning with renderable 3D scene priors, and real-robot deployment, including drones and quadrupeds (Yan et al., 17 Aug 2025, Zhang et al., 29 Jun 2026).

1. Task formulations and benchmark semantics

A common formalization treats ImageNav as an episodic POMDP in which the policy receives the current observation and a fixed goal image and predicts an action,

$\pi(a_t \mid x_t, x^*),$

or, in decoder-style formulations, a goal-conditioned action sequence model over the history of observations and actions (Mezghani et al., 2021, Pelluri, 2024). In Habitat-based RGB-only work, the observation can be a panoramic RGB state $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$; in single-view settings it is an egocentric RGB image paired with the goal image; in drone settings the policy may receive the channel-wise concatenation $s_t = O_c \oplus O_g$ and output continuous velocities instead of discrete actions (Mezghani et al., 2021, Yan et al., 17 Aug 2025).

ImageNav is distinct from PointGoal and ObjectNav. PointGoal specifies a coordinate target and emphasizes metric localization; ObjectNav specifies a semantic category and accepts any suitable instance; ImageNav specifies a particular view or, in instance-specific variants, a particular object instance depicted by an image (Li et al., 15 Feb 2025, Krantz et al., 2022). This distinction became especially important once the literature recognized that earlier ImageNav definitions were not standardized and often coupled the goal camera to the agent’s camera. “Instance-specific Image Goal Navigation” redefined the task so that the goal image prominently depicts a specific object instance and is captured with camera parameters independent of the agent, while “free-view image-goal navigation” further allows the goal image to be taken by a different camera $A$ at arbitrary pose $\boldsymbol{\rm T}_g \in SE(3)$ while the navigating agent carries camera $B$ (Krantz et al., 2022, Guo et al., 1 Aug 2025).

Success criteria vary accordingly. A common benchmark definition declares success when the agent issues [STOP](https://www.emergentmind.com/topics/self-taught-optimizer-stop) within $1$ m of the goal; other formulations also require view alignment, such as $\alpha_t < 25^\circ$ or $\alpha_s = 30^\circ$ (Mezghani et al., 2021, Yan et al., 17 Aug 2025, Zhang et al., 29 Jun 2026). Instance-specific formulations instead define success relative to valid viewpoints around a target object instance rather than the original goal-camera pose (Krantz et al., 2022). Across these settings, the dominant path-efficiency metric remains

$$\text{SPL} = \frac{1}{N}\sum_{i=1}^{N} S_i \frac{L_i}{\max(L_i, p_i)},$$

with notation varying slightly across papers (Mezghani et al., 2021). More recent work argues that this coarse $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$0 m criterion is sufficient for “finding” but inadequate for “last-meter” tasks such as grasping or inspection, motivating additional pose-precision metrics,

$$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$1

evaluated at the final stop pose (Deng et al., 7 Apr 2026).

2. Goal-conditioned perception and representation learning

A major line of work treats ImageNav primarily as a representation problem: the agent must preserve detailed information in the goal image and condition perception of the current observation on that goal. FGPrompt identifies two limitations of late-fusion policies: loss of fine-grained goal information and weak goal conditioning of the observation encoder. It replaces independent goal/observation encoding with “Fine-grained Goal Prompting,” either through channel-wise early fusion $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$2 or through mid-level FiLM modulation driven by high-resolution goal feature maps. In its ablations, a later-fusion baseline trained for 50M steps on Gibson achieved SR $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$3 and SPL $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$4, whereas FGPrompt-EF reached SR $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$5, SPL $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$6, and FGPrompt-MF reached SR $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$7, SPL $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$8 (Sun et al., 2023).

REGNav argues that many ImageNav failures arise before fine-grained matching becomes useful: when the current observation and goal image belong to different rooms, direct similarity is weak and the agent tends to wander. It therefore pre-trains a “room expert” on unlabeled room images to predict whether $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$9 and $s_t = O_c \oplus O_g$0 are taken in the same room, then injects that relation into the policy. The paper reports that explicit fusion of the room relation substantially outperforms implicit fusion; on Gibson, explicit fusion attains SPL $s_t = O_c \oplus O_g$1, SR $s_t = O_c \oplus O_g$2, while implicit fusion attains SPL $s_t = O_c \oplus O_g$3, SR $s_t = O_c \oplus O_g$4 (Li et al., 15 Feb 2025).

RFSG pursues a similar objective with a lightweight architecture. Its refined feature fusion module combines spatial-channel attention with a weight decoupling module,

$s_t = O_c \oplus O_g$5

and supplements this with parameter-free self-distillation and an image scene graph. In a Gibson ablation at 100M steps, adding refined feature fusion, self-distillation, and scene graph enhancement raised SPL from $s_t = O_c \oplus O_g$6 to $s_t = O_c \oplus O_g$7 and SR from $s_t = O_c \oplus O_g$8 to $s_t = O_c \oplus O_g$9; at 500M steps the full model reported SPL $A$0, SR $A$1, and runtime up to $A$2 frames per second on an RTX3080 (Feng et al., 14 Mar 2025).

Continuous-control settings expose an additional representation bottleneck. SIGN treats drone ImageNav as end-to-end visual RL with a ResNet9 backbone, GRU memory, and auxiliary tasks that improve encoder quality: future transition prediction in latent space and RandomShift consistency. Its training objective augments PPO with auxiliary losses, and its ablations report that each auxiliary task improves early training speed and final performance, while combining both yields the best sample efficiency and performance for continuous ImageNav (Yan et al., 17 Aug 2025). This suggests that, across both discrete and continuous settings, ImageNav performance is tightly coupled to how explicitly the architecture preserves goal detail and conditions perception on goal-relevant structure.

3. Memory, topology, and structured scene reasoning

Another major family of methods treats long-horizon ImageNav as a memory problem. “Memory-Augmented Reinforcement Learning for Image-Goal Navigation” builds an RGB-only policy around a self-supervised reachability network, an episodic memory of previously visited states, and attention over that memory. The reachability encoder $A$3 is trained to predict whether two observations are within $A$4 random-walk steps, yielding a topometric embedding used both for novelty-based memory writing and attention-based policy conditioning. On Gibson, this method reports overall SPL $A$5, Success $A$6, exceeding NTS-D’s SPL $A$7, Success $A$8, while using RGB alone and no pre-exploration of test scenes (Mezghani et al., 2021).

TSGM makes the memory explicitly topological and semantic. It incrementally builds a graph

$A$9

whose image nodes represent visited places and object nodes represent landmarks detected from panoramic RGB-D observations. A Cross Graph Mixer performs goal-conditioned message passing between image-image, object-object, and image-object neighborhoods, and a memory attention module reads out contextual current and goal features for a recurrent policy. On Gibson ImageNav, TSGM reports $\boldsymbol{\rm T}_g \in SE(3)$0 Success and $\boldsymbol{\rm T}_g \in SE(3)$1 SPL, compared with VGM’s $\boldsymbol{\rm T}_g \in SE(3)$2 Success and $\boldsymbol{\rm T}_g \in SE(3)$3 SPL, and it shows especially strong gains on curved paths and hard episodes (Kim et al., 2022).

MemoNav further decomposes memory into short-term memory (STM), long-term memory (LTM), and working memory (WM). STM is the topological map’s node feature set $\boldsymbol{\rm T}_g \in SE(3)$4; a forgetting module discards the bottom $\boldsymbol{\rm T}_g \in SE(3)$5 fraction of nodes according to goal-decoder attention scores; a global node serves as LTM; and GATv2 fuses retained STM and LTM into WM,

$\boldsymbol{\rm T}_g \in SE(3)$6

The paper’s central claim is that previous memory methods use all historical observations without considering the goal-relevant fraction. In multi-goal Gibson experiments, MemoNav improves over VGM from $\boldsymbol{\rm T}_g \in SE(3)$7 to $\boldsymbol{\rm T}_g \in SE(3)$8 on 2-goal PR/PPL, from $\boldsymbol{\rm T}_g \in SE(3)$9 to $B$0 on 3-goal PR/PPL, and from $B$1 to $B$2 on 4-goal PR/PPL (Li et al., 2024). The broader implication is that explicit memory is valuable, but indiscriminate memory is not.

4. Policy classes, action spaces, and control regimes

ImageNav policies now span discrete action classification, autoregressive sequence models, trajectory generators, and continuous control. The standard embodied-AI setting remains discrete—MOVE_FORWARD, TURN_LEFT, TURN_RIGHT, and STOP—and “Transformers for Image-Goal Navigation” recasts this regime as offline sequence modeling. Its decoder-only Transformer prepends a goal token $B$3, interleaves observation and action embeddings, and learns

$B$4

by goal-conditioned behavior cloning from OVRL-v2 trajectories. On Gibson-curved, it outperforms BC baselines but remains far below the RL expert: overall $B$5 Success/SPL versus OVRL-v2’s $B$6 (Pelluri, 2024). The result is methodologically important because it shows that ImageNav can be posed as autoregressive action generation conditioned on goal and history, not only as online RL.

RoamFlow extends the generative view from step-wise actions to trajectories. Its policy predicts an action sequence $B$7 with MeanFlow, modeling the average velocity field

$B$8

and using a two-stage training pipeline: imitation from Hybrid A* trajectories followed by PPO fine-tuning in trajectory space. In Habitat, RoamFlow reports Gibson SR $B$9, SPL $1$0, CR $1$1, inference time $1$2 ms, and MP3D SR $1$3, SPL $1$4, CR $1$5, inference time $1$6 ms; on a Unitree Go2 quadruped it reports SR $1$7, $1$8 collisions per run, and $1$9 ms inference time (Zhang et al., 29 Jun 2026).

Continuous ImageNav changes the control problem more fundamentally. SIGN emphasizes that prior ImageNav is dominated by ground robots, simulated environments, and discrete control, whereas drones are inherently unstable and require high-frequency continuous control. Its policy outputs

$\alpha_t < 25^\circ$0

with direct velocity control, while PX4 tracks the commands at about $\alpha_t < 25^\circ$1 Hz. The task remains goal-conditioned through $\alpha_t < 25^\circ$2, but the policy operates without explicit global maps or GPS, and stop is implemented through velocity thresholds rather than as a discrete action (Yan et al., 17 Aug 2025). This makes clear that ImageNav is not intrinsically tied to low-frequency discrete action abstractions.

5. From semantic matching to geometric localization

A persistent debate in ImageNav is whether the task is best treated as semantic retrieval, policy learning, or explicit geometric localization. IEVE addresses the instance-specific case by inserting a dedicated verification stage between exploration and exploitation. It uses class-conditioned instance segmentation, local feature extraction with DISK, matching with LightGlue, and a switching rule

$\alpha_t < 25^\circ$3

where $\alpha_t < 25^\circ$4 is the number of matched keypoints and $\alpha_t < 25^\circ$5 is the distance to a potential target. On HM3D-SEM, IEVE reports success $\alpha_t < 25^\circ$6 with Mask R-CNN and $\alpha_t < 25^\circ$7 with InternImage, exceeding the prior Mod-IIN baseline at $\alpha_t < 25^\circ$8 (Lei et al., 2024). The paper’s core claim is behavioral rather than architectural: for ambiguous instance goals, agents should “get closer to confirm.”

BEINGS turns ImageNav into an optimal-control problem. It assumes a pre-built 3D Gaussian Splatting scene prior and defines the HJB cost

$\alpha_t < 25^\circ$9

with Bayesian updates over a belief distribution $\alpha_s = 30^\circ$0 on goal location. Model predictive control evaluates rollouts by rendering future observations from the 3DGS map and minimizing image dissimilarity plus a Bayesian exploration cost. In blimp experiments, BEINGS reports SR $\alpha_s = 30^\circ$1 on easy tasks, $\alpha_s = 30^\circ$2 on medium tasks, and $\alpha_s = 30^\circ$3 on hard tasks, with corresponding step counts $\alpha_s = 30^\circ$4, $\alpha_s = 30^\circ$5, and $\alpha_s = 30^\circ$6 (Meng et al., 2024).

IGL-Nav retains the renderable-geometry idea but makes it incremental and free-view compatible. It builds a feed-forward 3DGS representation online from posed RGB-D observations, then localizes the goal by coarse matching in a discretized $\alpha_s = 30^\circ$7D pose space $\alpha_s = 30^\circ$8, showing that the alignment problem can be reformulated as efficient 3D convolution over voxelized scene and goal embeddings. Fine localization then uses differentiable rendering and matching-constrained optimization. On the standard NRNS benchmark, IGL-Nav reports straight overall SR $\alpha_s = 30^\circ$9, SPL $$\text{SPL} = \frac{1}{N}\sum_{i=1}^{N} S_i \frac{L_i}{\max(L_i, p_i)},$$0, and curved overall SR $$\text{SPL} = \frac{1}{N}\sum_{i=1}^{N} S_i \frac{L_i}{\max(L_i, p_i)},$$1, SPL $$\text{SPL} = \frac{1}{N}\sum_{i=1}^{N} S_i \frac{L_i}{\max(L_i, p_i)},$$2; on its free-view benchmark, supervised wide-FOV evaluation reaches SR $$\text{SPL} = \frac{1}{N}\sum_{i=1}^{N} S_i \frac{L_i}{\max(L_i, p_i)},$$3, SPL $$\text{SPL} = \frac{1}{N}\sum_{i=1}^{N} S_i \frac{L_i}{\max(L_i, p_i)},$$4 (Guo et al., 1 Aug 2025).

AnyImageNav makes the geometric turn explicit. It argues that the usual $$\text{SPL} = \frac{1}{N}\sum_{i=1}^{N} S_i \frac{L_i}{\max(L_i, p_i)},$$5 m success criterion hides a last-meter problem: many prior methods can say that the agent is “near,” but they do not recover the exact 6-DoF pose of the goal image. AnyImageNav treats the goal image as a geometric query, uses a semantic relevance map from DINOv2 to guide exploration, then invokes VGGT and Pi3 only when proximity is high, recovering the goal pose through dense correspondences and Sim(3) alignment. It reports Gibson SR $$\text{SPL} = \frac{1}{N}\sum_{i=1}^{N} S_i \frac{L_i}{\max(L_i, p_i)},$$6, SPL $$\text{SPL} = \frac{1}{N}\sum_{i=1}^{N} S_i \frac{L_i}{\max(L_i, p_i)},$$7, $$\text{SPL} = \frac{1}{N}\sum_{i=1}^{N} S_i \frac{L_i}{\max(L_i, p_i)},$$8 m, $$\text{SPL} = \frac{1}{N}\sum_{i=1}^{N} S_i \frac{L_i}{\max(L_i, p_i)},$$9, and HM3D SR $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$00, SPL $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$01, $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$02 m, $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$03 (Deng et al., 7 Apr 2026). This marks a conceptual shift from “visually similar enough” to “geometrically registered.”

6. Evaluation regimes, embodiments, and open problems

The experimental ecosystem around ImageNav is now broad but still centered on Habitat and a small set of scene corpora. Gibson remains the most common training benchmark, often with 72 training scenes and 14 test scenes, while MP3D and HM3D are standard for cross-domain evaluation; HM3D-SEM supports instance-specific navigation; and newer work adds straight/curved path taxonomies, multi-goal episodes, and free-view splits (Mezghani et al., 2021, Krantz et al., 2022, Guo et al., 1 Aug 2025). This diversity has improved coverage, but it also means that reported numbers are only partially comparable unless the goal definition, camera assumptions, embodiment, and success metric are aligned.

Embodiments now extend well beyond the original ground-robot setting. TSGM demonstrates a mobile robot system with a Ricoh Theta V and depth estimated from OmniDepth (Kim et al., 2022). BEINGS is evaluated on a miniature underactuated blimp with OptiTrack localization and a 3DGS prior (Meng et al., 2024). SIGN targets a quadrotor with RGB, IMU/VIO, PX4, and a RealSense D435i safety module, explicitly addressing continuous velocity control and obstacle avoidance in cluttered indoor environments (Yan et al., 17 Aug 2025). RoamFlow reports onboard operation on a Unitree Go2 quadruped with Jetson Orin NX and RealSense D435i (Zhang et al., 29 Jun 2026). The practical consequence is that ImageNav is no longer synonymous with simulated wheeled agents issuing symbolic actions.

Several limitations recur across otherwise different paradigms. RL-based methods still face substantial sample-efficiency demands—500M steps in memory-augmented RL and 500M-step training regimes in multiple later works—and offline sequence models remain vulnerable to covariate shift and short context windows (Mezghani et al., 2021, Pelluri, 2024). Map-less approaches can get stuck in visually aliased or maze-like environments, and explicit global mapping is often absent by design (Yan et al., 17 Aug 2025). Geometry-heavy pipelines reduce ambiguity but introduce other dependencies: static-scene assumptions, threshold tuning, depth or odometry requirements, and significant computational overhead (Deng et al., 7 Apr 2026, Guo et al., 1 Aug 2025). This suggests that the most durable research direction is not a single paradigm but tighter integration: goal-conditioned perception, selective memory, verification before commitment, explicit safety when embodiment demands it, and geometric localization when the task requires more than being merely within $$x_t, x^* \in \mathbb{R}^{v \times 3 \times 128 \times 128}$$04 m of the goal.