Visual Navigation Models (VNMs) promise generalizable, robot navigation by learning from large-scale visual demonstrations. Despite growing real-world deployment, existing evaluations rely almost exclusively on success rate, whether the robot reaches its goal, which conceals trajectory quality, collision behavior, and robustness to environmental change. We present a real-world evaluation of five state-of-the-art VNMs (GNM, ViNT, NoMaD, NaviBridger, and CrossFormer) across two robot platforms and five environments spanning indoor and outdoor settings. Beyond success rate, we combine path-based metrics with vision-based goal-recognition scores and assess robustness through controlled image perturbations (motion blur, sunflare). Our analysis uncovers three systematic limitations: (a) even architecturally sophisticated diffusion and transformer-based models exhibit frequent collisions, indicating limited geometric understanding; (b) models fail to discriminate between different locations that are perceptually similar, however some semantics differences are present, causing goal prediction errors in repetitive environments; and (c) performance degrades under distribution shift. We will publicly release our evaluation codebase and dataset to facilitate reproducible benchmarking of VNMs.
Figures & tables
Fig. 1 : Visual Navigation Models promise to deliver generalizable navigation in novel environments. We provide a comprehensive analysis of 5 models with 2 robot embodiments over 5 real-world environments and propose metrics to better capture their performance.
Fig. 2 : A Visual Navigation Model (VNM) takes as input a sequence of k image observations Ot and a goal image og at timestep t . An image and distance encoder encode the observation sequence and the current ( ot ) and goal observation respectively. The backbone and decoder vary by method. The model outputs action A^t , which can take the form of a single next waypoint or a trajectory, and can also include the temporal distance d .
Model
Backbone
Specifics
Output
GNM [ 1 ]
Fully Connected
Shared action space
({xi}t+1t+H,d)
ViNT [ 6 ]
Transformer
Goal Fusion
(xt,d)
NoMaD [ 7 ]
Diffusion
Goal Masking
{xi}t+1t+H
NaviBridger [ 10 ]
Diffusion
Learned action prior
{xi}t+1t+H
CrossFormer [ 11 ]
Transformer
Embodiment specific head
xt
TABLE I : The specific architectures of the Visual Navigation Models (VNMs) considered in this paper. Methods differ in their backbone architecture and output type. Some models introduce specific architecture choices within the VNM pipeline.
Corridor
Office loop
Method
dist.
p.len.
dist.
p.len.
GNM
0.33 ± 0.13
4.52 ± 0.16
9.49 ± 9.43
19.28 ± 10.03
ViNT
0.40 ± 0.09
4.81 ± 0.14
3.53 ± 2.19
32.51 ± 2.57
NoMaD
0.56 ± 0.09
5.25 ± 0.15
7.22 ± 5.80
30.04 ± 9.45
NaviBridger
0.63 ± 0.16
4.46 ± 0.63
17.19 ± 5.29
10.66 ± 5.61
CrossFormer
1.73 ± 0.00
2.34 ± 0.04
-
-
TABLE II: Path length ( p.len ) and Goal Distance ( dist. ) across environments and models (mean ± standard-deviation).
Fig. 3 : Real-world evaluation environments (indoor and outdoor). Blue panels correspond to rover while yellow are quadruped deployments.
Fig. 4: Image quality metrics (LPIPS, PSNR, DSSIM) for goal-predicted cases in the Arena (Top) and Snow (Bottom) environments.
Corridor
Office loop
Arena
Stairs
Snow
Method
n.err.
g.pred.
n.err.
g.pred.
n.err.
g.pred.
n.err.
g.pred.
n.err.
g.pred.
GNM
0.36 ± 0.49
3/3
3.23 ± 5.47
5/10
6.94 ± 1.96
2/3
0.89 ± 0.55
3/3
1.95 ± 1.05
3/3
ViNT
0.43 ± 0.49
3/3
7.34 ± 13.58
9/10
7.43 ± 3.05
2/5
0.94 ± 0.66
3/3
3.82 ± 5.52
3/3
NoMaD
0.49 ± 0.70
0/3
8.53 ± 15.04
8/10
10.59 ± 7.76
5/5
0.98 ± 0.60
3/3
3.95 ± 3.65
3/3
NaviBridger
0.72 ± 0.44
1/3
2.37 ± 2.59
1/10
11.04 ± 6.91
3/5
2.01 ± 1.29
3/3
2.47 ± 1.92
3/3
CrossFormer
0.41 ± 0.49
0/3
-
-
13.21 ± 6.97
5/5
-
-
-
-
TABLE III : Topological Node Error ( n.err. ) and Predicted Goal Counter ( g.pred. ) across environments and models.
Method
Corridor
Office loop
col.
col.
GNM
0/3
5/10
ViNT
0/3
1/10
NoMaD
3/3
2/10
NaviBridger
2/3
9/10
CrossFormer
3/3
-
TABLE IV : Collision ( col. ) results for Corridor and Office loop.
Fig. 5 : Corridor deployment results. Green (left): goal reached without collision in all trials. Red (right): at least one collision occurred. See Table IV for details. (Note: overlapping collisions at the same location are shown only once for clarity).
Fig. 6 : All Stairs trajectories for GNM , ViNT , NoMaD and NaviBridger (see Table III ) with the reference trajectory .
Fig. 7 : All Snow trajectories for GNM , ViNT , NoMaD and NaviBridger (see Table II and III ) with the reference trajectory .
Fig. 8 : One Arena trajectory for GNM , ViNT , NoMaD , NaviBridger , and CrossFormer (see Table II and III ) with the reference trajectory .
Fig. 9 : Last three Office loop trajectories for GNM , ViNT , NoMaD and NaviBridger (see Table II , III and IV ) with the reference trajectory .
Precision
Generalization
Model
dist.
p. len.
n. err.
g. pred.
GNM
0.62 ± 0.32
4.68 ± 0.35
0.34 ± 0.53
1/3
ViNT
0.30 ± 0.35
4.70 ± 0.34
0.50 ± 0.67
2/3
GNM
0.40 ± 0.15
4.15 ± 0.04
0.44 ± 0.59
2/3
ViNT
0.21 ± 0.04
4.33 ± 0.03
0.39 ± 0.52
3/3
TABLE V : Navigation metrics of Corridor environment with perturbations: BLUR (gray) vs SUNFLARE (yellow).
Vision Navigation Foundation Models (VNMs) promise end-to-end learned navigation policies capable of zero-shot deployment across diverse embodiments and environments. To maintain generality, many vision-based navigation models predict normalized actions. However, this normalization introduces a critical deployment vulnerability: applying different scaling factors to the same normalized trajectory alters its physical geometry, which degrades navigation performance and increases collision risks. We address this vulnerability by conditioning the model on normalized action histories alongside image observations, providing explicit context on the relationship between the model's predictions and the robot's actual physical displacement. Furthermore, current VNMs often struggle in visually repetitive environments that lack distinct features. To resolve this issue, we integrate a DINOv3 encoder, whose richer representations enable our model to capture both spatial and geometric dimensions between observations. VISTA generalizes robustly to out-of-distribution environments, achieving 100% goal prediction accuracy in zero-shot, real-world deployment in Outdoor, Forest and Office settings, and an average of 95% checkpoints crossed, demonstrating consistent path following in unseen environments.
Maeva Guerrier, Koki Kobayashi, Simon Roy +2
Polytechnique Montreal, MILA · Institute of Science Tokyo · Polytechnique Montreal +2
Navigation is a fundamental capability of autonomous systems, yet most existing approaches rely on highly structured models and strong prior assumptions, limiting their robustness in open and uncertain real-world environments. Vision-and-Language Navigation (VLN) offers a promising direction by enabling robots to integrate natural language understanding with visual perception in a data-driven manner. Although VLN has attracted increasing research attention, systematic methodological taxonomy and real-world validation remain limited. This survey presents a comprehensive review of VLN research. Specifically, state-of-the-art methods are organized along two orthogonal dimensions: action paradigms, including hierarchical and monolithic frameworks, and model paradigms, including discriminative and generative approaches. A critical analysis of their respective strengths and limitations is provided. Additionally, we conduct a systematic real-world evaluation of representative VLN system configurations on a physical robotic platform. Experiments across ten diverse real-world scenes show a substantial performance gap between simulation and real-world deployment under the tested configurations: a representative monolithic RGB-only method achieves 61% success in simulation but drops to 22% in real-world deployment, while a hierarchical framework achieves a higher real-world success rate of 51%, suggesting stronger robustness in our evaluation setting. Finally, we highlight key challenges in perception, decision-making, and control that must be addressed in future research.
Liuyi Wang, Kai Sheng, Zongtao He +8
College of Electronic and Information Engineering, Tongji University, Shanghai, China
Vision-Language Models (VLMs) have shown remarkable progress in Vision-Language Navigation (VLN), offering new possibilities for navigation decision-making that could benefit both robotic platforms and human users. However, real-world navigation is inherently conditioned by the agent's mobility constraints. For example, a sweeping robot cannot traverse stairs, while a quadruped can. We introduce Capability-Conditioned Navigation (CapNav), a benchmark designed to evaluate how well VLMs can navigate complex indoor spaces given an agent's specific physical and operational capabilities. CapNav defines five representative human and robot agents, each described with physical dimensions, mobility capabilities, and environmental interaction abilities. CapNav provides 45 real-world indoor scenes, 473 navigation tasks, and 2365 QA pairs to test if VLMs can traverse indoor environments based on agent capabilities. We evaluate 13 modern VLMs and find that current VLM's navigation performance drops sharply as mobility constraints tighten, and that even state-of-the-art models struggle with obstacle types that require reasoning on spatial dimensions. We conclude by discussing the implications for capability-aware navigation and the opportunities for advancing embodied spatial reasoning in future VLMs. The benchmark is available at https://github.com/makeabilitylab/CapNav
Xia Su, Ruiqi Chen, Benlin Liu +4
University of Washington Seattle, WA, USA · University of California, Santa Cruz Santa Cruz, CA, USA