arxiv: 2605.13328 · v1 · submitted 2026-05-13 · 💻 cs.RO · cs.AI· cs.CL· cs.CV

Recognition: no theorem link

What Limits Vision-and-Language Navigation ?

Yunheng Wang , Yuetong Fang , Taowen Wang , Lusong Li , Kun Liu , Junzhe Xu , Zizhao Yuan , Yixiao Feng

show 4 more authors

Jiaxi Zhang Wei Lu Zecui Zeng Renjing Xu

Authors on Pith no claims yet

Pith reviewed 2026-05-14 17:55 UTC · model grok-4.3

classification 💻 cs.RO cs.AIcs.CLcs.CV

keywords vision-and-language navigationstereo visionsim-to-real transfertarget-location priorsembodied AIrobot navigationVLN-CE

0 comments

The pith

StereoNav uses target-location priors and stereo vision to achieve robust real-world vision-and-language navigation with fewer parameters and less data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the primary bottleneck in vision-and-language navigation is the lack of robust spatial grounding and cross-domain priors rather than insufficient model scale or training data volume. It introduces StereoNav, which supplies persistent target-location priors to ground the agent when instructions are vague and uses stereo vision to fuse semantics with geometry for depth-aware action prediction. This combination is shown to deliver state-of-the-art success rates on R2R-CE and RxR-CE while requiring substantially smaller models and less training data than scaling-heavy baselines. Real-world robot deployments further demonstrate improved reliability under lighting changes, motion blur, and unstructured settings. A sympathetic reader would therefore see the work as evidence that targeted perceptual mechanisms can close the sim-to-real gap more efficiently than brute-force increases in capacity.

Core claim

StereoNav achieves state-of-the-art egocentric RGB performance on R2R-CE and RxR-CE with SR and SPL scores of 81.1 percent and 68.3 percent, and 67.5 percent and 52.0 percent respectively, by introducing target-location priors that remain invariant across simulation-to-real domain shifts and by leveraging stereo vision to construct a unified semantic-geometric representation that supports precise action prediction despite motion blur and illumination changes.

What carries the argument

Target-location priors, which serve as a persistent visual bridge providing stable, domain-invariant guidance, together with stereo vision that unifies semantics and geometry into a single depth-aware representation for action prediction.

If this is right

Navigation succeeds at higher rates on standard VLN-CE benchmarks while using fewer parameters and less training data than scaling-based methods.
Robots maintain reliable performance in complex unstructured environments where monocular approaches degrade.
Vague instructions are handled more gracefully because the priors supply persistent spatial grounding.
Perceptual robustness for embodied tasks can be obtained without massive increases in model size or data volume.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same priors and stereo mechanism could be applied to other embodied tasks such as object manipulation in changing environments.
Explicit stereo depth may prove more reliable than learned monocular depth for navigation under real-world blur and lighting variation.
Testing whether the priors generalize to dynamic obstacles not present in training would reveal the limits of the invariance assumption.

Load-bearing premise

Target-location priors remain useful and unchanged when moving from simulated training to physical robot execution, and stereo cameras reliably supply depth cues that overcome motion blur and lighting shifts without extra calibration.

What would settle it

A real-world robotic deployment in which StereoNav's success rate falls below that of prior monocular scaling-based methods under strong illumination changes or fast motion would falsify the claim that these priors and stereo cues suffice for robust navigation.

Figures

Figures reproduced from arXiv: 2605.13328 by Jiaxi Zhang, Junzhe Xu, Kun Liu, Lusong Li, Renjing Xu, Taowen Wang, Wei Lu, Yixiao Feng, Yuetong Fang, Yunheng Wang, Zecui Zeng, Zizhao Yuan.

**Figure 1.** Figure 1: Performance under different backbones and training data scales. (a) Left: The red trend line shows that success rate increases over time with diminishing returns. The inset reports the MLVU-Dev [7] scores of the adopted backbone VLMs, indicating that stronger backbones improve early performance, but their gains gradually saturate. (b) Right: Arrows indicate the performance gains from increased training dat… view at source ↗

**Figure 2.** Figure 2: Impact of visual uncertainty on VLN agents. (a) Top: Visual examples of four common perturbations during embodied navigation. (b) Bottom: Performance degradation of representative open-source VLN methods, where the LLaVA-based and Qwen-based methods correspond to StreamVLN [6] and JanusVLN [18], respectively. Although existing agents perform competitively in the ideal setting, their navigation performance… view at source ↗

**Figure 3.** Figure 3: Impact of instructional under-specification on VLN agents. (a) Top: Representative cases of Directional Ambiguity, where under-specified route or orientation cues permit multiple feasible paths, and Docking Ambiguity, where vague goal descriptions permit multiple plausible stopping targets. (b) Bottom: Distributions of ambiguity scores across representative VLMs, where lower scores indicate stronger ambigu… view at source ↗

**Figure 4.** Figure 4: Overview of StereoNav. StereoNav takes stereo RGB observations, a navigation instruction, and a target-location prior as input. The target prior is rendered as persistent visual guidance, while stereo observations are encoded into unified semantic, structural, and geometric tokens through 2D semantic, 2D structural, and 3D geometry encoders. These tokens are then processed by the MLLM for joint action and … view at source ↗

**Figure 5.** Figure 5: Qualitative examples of StereoNav in real-world. From top to bottom: Outdoor, Office, Lobby, and Gym. The results demonstrate StereoNav’s reliability across diverse scenes. Note that these examples are visualized from a third-person perspective; details regarding the actual sensor inputs used for navigation are provided in Section C.2. 5 Experiment We evaluate StereoNav from multiple perspectives. We first… view at source ↗

**Figure 6.** Figure 6: Robustness and reliability evaluation of StereoNav. (a–b) Robustness under viewpoint oscillation and motion blur, where StereoNav shows smaller performance degradation as perturbation severity increases. (c–d) Reliability in goal stopping under different stopping-error thresholds, where StereoNav achieves higher SR and SPL, especially under strict target-neighborhood constraints. As shown in [PITH_FULL_IM… view at source ↗

**Figure 7.** Figure 7: Prompt template for instruction ambiguity assessment. Given a navigation instruction and sampled observation frames, the evaluator identifies whether Directional Ambiguity or Docking Ambiguity exists and returns binary labels for the two ambiguity types. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

**Figure 8.** Figure 8: Prompt templates for StereoNav model input and real-world instruction preprocessing. (a) Top: The StereoNav model prompt formats observations and instructions as inputs, with the assistant response as the action label. (b) Bottom: The real-world preprocessing prompt converts user commands into the structured instruction and target-location prior required by StereoNav. B.2 Prompt Template We provide the det… view at source ↗

**Figure 9.** Figure 9: Visualization of depth estimation results in a representative virtual environment. StereoNav leverages stereo disparity between the left and right first-person views to produce stable depth estimates, providing reliable geometric cues for robust navigation under approximate goal priors [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗

**Figure 10.** Figure 10: First-person stereo observations in simulation and real-world deployment. (a) Top: First-person left- and right-view observations from a representative virtual environment, where the rendered target-location prior provides persistent visual guidance during navigation. (b) Bottom: First-person left- and right-view observations from a representative real-world Gym scenario, demonstrating the deployment of … view at source ↗

**Figure 11.** Figure 11: Ablation study on fusion weights in Unified Understanding Modeling. The results show that moderate structural guidance and lightweight geometric guidance lead to the best overall performance. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗

read the original abstract

Vision-and-Language Navigation (VLN) is a cornerstone of embodied intelligence. However, current agents often suffer from significant performance degradation when transitioning from simulation to real-world deployment, primarily due to perceptual instability (e.g., lighting variations and motion blur) and under-specified instructions. While existing methods attempt to bridge this gap by scaling up model size and training data, we argue that the bottleneck lies in the lack of robust spatial grounding and cross-domain priors. In this paper, we propose StereoNav, a robust Vision-Language-Action framework designed to enhance real-world navigation consistency. To address the inherent gap between synthetic training and physical execution, we introduce Target-Location Priors as a persistent bridge. These priors provide stable visual guidance that remains invariant across domains, effectively grounding the agent even when instructions are vague. Furthermore, to mitigate visual disturbances like motion blur and illumination shifts, StereoNav leverages stereo vision to construct a unified representation of semantics and geometry, enabling precise action prediction through enhanced depth awareness. Extensive experiments on R2R-CE and RxR-CE demonstrate that StereoNav achieves state-of-the-art egocentric RGB performance, with SR and SPL scores of 81.1% and 68.3%, and 67.5% and 52.0%, respectively, while using significantly fewer parameters and less training data than prior scaling-based approaches. More importantly, real-world robotic deployments confirm that StereoNav substantially improves navigation reliability in complex, unstructured environments. Project page: https://yunheng-wang.github.io/stereonav-public.github.io.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

StereoNav adds stereo depth and target priors to VLN for better sim-to-real transfer, with solid benchmark numbers but thin real-world evidence and missing ablations.

read the letter

StereoNav focuses on the sim-to-real gap in vision-and-language navigation by combining stereo vision for unified semantic and geometric features with Target-Location Priors meant to provide stable guidance across domains. The paper reports strong results on R2R-CE and RxR-CE, reaching 81.1% SR and 68.3% SPL on one and 67.5% SR and 52.0% SPL on the other, while using fewer parameters and less training data than scaling-heavy baselines. They also describe real-robot tests showing improved reliability in unstructured settings. That combination of concrete benchmark scores and actual hardware deployment is more than many VLN papers deliver. The approach is straightforward engineering rather than a new theoretical insight, but it directly targets perceptual issues like blur and lighting shifts plus vague instructions. The soft spots are clear. There are no ablations isolating the priors' contribution, no error bars on the reported scores, and no explicit formulation or fitting procedure for how the priors are built. The real-world section only states qualitative improvement without success rates, SPL values, or direct baseline comparisons on the same robot and environments. This makes it hard to tell whether the gains come from stereo alone, task tuning, or the priors themselves, and it leaves the invariance claim untested. The work is aimed at researchers working on practical embodied navigation who want alternatives to ever-larger models. A reader could pull useful implementation ideas from the stereo unification part, but would need the code or more details to verify the priors. It deserves peer review because the benchmark results are competitive and the real-robot angle is worth checking, even if the current evidence needs tightening on ablations and quantitative hardware metrics.

Referee Report

3 major / 1 minor

Summary. The manuscript introduces StereoNav, a Vision-Language-Action framework for VLN that incorporates Target-Location Priors as a domain-invariant bridge between simulation and real-world execution and uses stereo vision to build a unified semantic-geometric representation. It reports SOTA egocentric RGB results on R2R-CE (SR 81.1%, SPL 68.3%) and RxR-CE (SR 67.5%, SPL 52.0%) with fewer parameters and less training data than prior scaling approaches, and asserts that real-world robotic deployments demonstrate substantially improved navigation reliability in unstructured environments.

Significance. If the Target-Location Priors can be shown to be robustly invariant and the real-world gains are supported by quantitative metrics and ablations, the work would offer a meaningful alternative to pure scaling in embodied navigation by emphasizing geometric priors and stereo depth cues. The reported benchmark numbers and parameter efficiency are notable strengths, but the absence of supporting details on the priors and physical experiments currently prevents a full assessment of impact.

major comments (3)

[Abstract] Abstract and method description: Target-Location Priors are presented as providing 'stable visual guidance that remains invariant across domains' without an explicit mathematical formulation, computation procedure, or fitting details, leaving their contribution to sim-to-real transfer unverified.
[Real-world experiments] Real-world robotic deployments section: the claim of 'substantially improves navigation reliability' rests on qualitative statements only; no quantitative metrics (SR, SPL, failure counts, or comparisons against baselines on the same robot and environments) are supplied.
[Experiments] Experiments section: the reported SR/SPL scores on R2R-CE and RxR-CE lack error bars, ablation tables isolating the priors versus stereo depth, and any description of how the priors are derived or validated for invariance.

minor comments (1)

[Abstract] The project page URL in the abstract contains a redundant '.github.io' suffix that should be corrected.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and will revise the manuscript to incorporate the requested details and analyses.

read point-by-point responses

Referee: [Abstract] Abstract and method description: Target-Location Priors are presented as providing 'stable visual guidance that remains invariant across domains' without an explicit mathematical formulation, computation procedure, or fitting details, leaving their contribution to sim-to-real transfer unverified.

Authors: We agree that the current presentation lacks sufficient technical detail. In the revision we will add an explicit mathematical definition of the Target-Location Priors (formulated as a persistent 3D semantic-geometric embedding derived from stereo disparity and semantic segmentation), the exact computation pipeline, and the fitting/validation procedure used to demonstrate cross-domain invariance. These additions will directly substantiate their role in sim-to-real transfer. revision: yes
Referee: [Real-world experiments] Real-world robotic deployments section: the claim of 'substantially improves navigation reliability' rests on qualitative statements only; no quantitative metrics (SR, SPL, failure counts, or comparisons against baselines on the same robot and environments) are supplied.

Authors: The referee correctly notes that only qualitative statements are currently provided. We will expand the real-world section with quantitative results (success rate, SPL, failure counts, and head-to-head comparisons against baselines) collected on the same robot and environments. These metrics will be reported alongside the existing qualitative observations. revision: yes
Referee: [Experiments] Experiments section: the reported SR/SPL scores on R2R-CE and RxR-CE lack error bars, ablation tables isolating the priors versus stereo depth, and any description of how the priors are derived or validated for invariance.

Authors: We will revise the experiments section to include error bars computed over multiple random seeds for all SR/SPL numbers. We will also add ablation tables that separately quantify the contribution of Target-Location Priors versus stereo depth cues, together with a detailed derivation of the priors and quantitative validation of their domain invariance. revision: yes

Circularity Check

0 steps flagged

No circularity: claims rest on empirical results and conceptual priors without self-referential reduction

full rationale

The provided manuscript text introduces Target-Location Priors as an asserted invariant bridge and leverages stereo vision for unified semantic-geometric representations, but contains no equations, fitted parameters renamed as predictions, or self-citations that reduce the central performance claims (SR/SPL on R2R-CE/RxR-CE or real-world reliability) to the inputs by construction. The invariance statement is presented as a design choice rather than derived from the same training distribution in a tautological loop, and no uniqueness theorems or ansatzes are smuggled via prior self-work. The derivation chain therefore remains self-contained against external benchmarks, with results framed as experimental outcomes.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the existence of domain-invariant target priors and the sufficiency of stereo geometry to overcome perceptual noise; both are introduced without upstream derivation or external validation in the provided abstract.

free parameters (1)

Target-Location Priors
Persistent visual guidance signals introduced to bridge sim-to-real gap; their exact construction and any fitting procedure are not specified.

axioms (1)

domain assumption Stereo vision supplies reliable depth that remains useful under motion blur and lighting variation
Invoked to justify the unified semantics-and-geometry representation.

invented entities (1)

Target-Location Priors no independent evidence
purpose: Provide stable visual guidance invariant across simulation and real-world domains
New construct introduced to ground the agent when instructions are vague; no independent falsifiable prediction supplied in abstract.

pith-pipeline@v0.9.0 · 5621 in / 1478 out tokens · 38595 ms · 2026-05-14T17:55:13.413265+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

59 extracted references · 40 canonical work pages · 9 internal anchors

[1]

Vision-and-language navigation: A survey of tasks, methods, and future directions

Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Eric Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. InAnnual Meeting of the Association for Computational Linguistics (ACL), pages 7606–7623, 2022

2022
[2]

Vision-and-language navigation today and tomorrow: A survey in the era of foundation models.arXiv preprint arXiv:2407.07035, 2024

Yue Zhang, Ziqiao Ma, Jialu Li, Yanyuan Qiao, Zun Wang, Joyce Chai, Qi Wu, Mohit Bansal, and Parisa Kordjamshidi. Vision-and-language navigation today and tomorrow: A survey in the era of foundation models.arXiv preprint arXiv:2407.07035, 2024

work page arXiv 2024
[3]

Homerobot: Open-vocabulary mobile manipulation.arXiv preprint arXiv:2306.11565, 2024

Sriram Yenamandra, Arun Ramachandran, Karmesh Yadav, Austin Wang, Mukul Khanna, Theophile Gervet, Tsung-Yen Yang, Vidhi Jain, Alexander William Clegg, John Turner, Zsolt Kira, Manolis Savva, Angel Chang, Devendra Singh Chaplot, Dhruv Batra, Roozbeh Mottaghi, Yonatan Bisk, and Chris Paxton. Homerobot: Open-vocabulary mobile manipulation.arXiv preprint arXi...

work page arXiv 2024
[4]

Navila: Legged robot vision-language-action model for navigation

An-Chieh Cheng, Yandong Ji, Zhaojing Yang, Xueyan Zou, Jan Kautz, Erdem Biyik, Hongxu Yin, Sifei Liu, and Xiaolong Wang. Navila: Legged robot vision-language-action model for navigation. InRobotics: Science and Systems (RSS), 2025

2025
[5]

Navid: Video-based vlm plans the next step for vision-and-language navigation

Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zhang, and He Wang. Navid: Video-based vlm plans the next step for vision-and-language navigation. InRobotics: Science and Systems (RSS), 2024

2024
[6]

Streamvln: Streaming vision-and-language navigation via slowfast context modeling,

Meng Wei, Chenyang Wan, Xiqian Yu, Tai Wang, Yuqiang Yang, Xiaohan Mao, Chenming Zhu, Wenzhe Cai, Hanqing Wang, Yilun Chen, et al. Streamvln: Streaming vision-and-language navigation via slowfast context modeling.arXiv preprint arXiv:2507.05240, 2025

work page arXiv 2025
[7]

MLVU: Benchmarking Multi-task Long Video Understanding

Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding.arXiv preprint arXiv:2406.04264, 2024

work page internal anchor Pith review arXiv 2024
[8]

Vision-and-language navigation: Interpret- ing visually-grounded navigation instructions in real environments

Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpret- ing visually-grounded navigation instructions in real environments. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3674–3683, 2018

2018
[9]

Matterport3d: Learning from RGB-D data in indoor environments

Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. InInternational Conference on 3D Vision (3DV), 2017

2017
[10]

Beyond the nav- graph: Vision and language navigation in continuous environments

Jacob Krantz, Erik Wijmans, Arjun Majundar, Dhruv Batra, and Stefan Lee. Beyond the nav- graph: Vision and language navigation in continuous environments. InEuropean Conference on Computer Vision (ECCV), pages 104–120, 2020

2020
[11]

Habitat: A platform for embodied ai research

Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A platform for embodied ai research. InIEEE International Conference on Computer Vision (ICCV), pages 9339–9347, 2019

2019
[12]

Available: https://arxiv.org/abs/2512.08186

Meng Wei, Chenyang Wan, Jiaqi Peng, Xiqian Yu, Yuqiang Yang, Delin Feng, Wenzhe Cai, Chenming Zhu, Tai Wang, Jiangmiao Pang, and Xihui Liu. Ground slow, move fast: A dual- system foundation model for generalizable vision-and-language navigation.arXiv preprint arXiv:2512.08186, 2025

work page arXiv 2025
[13]

Prospect: Unified streaming vision-language navigation via semantic–spatial fusion and latent predictive representation.arXiv preprint arXiv:2603.03739, 2026

Zehua Fan, Wenqi Lyu, Wenxuan Song, Linge Zhao, Yifei Yang, Xi Wang, Junjie He, Lida Huang, Haiyan Liu, Bingchuan Sun, Guangjun Bao, Xuanyao Mao, Liang Xu, Yan Wang, and Feng Gao. Prospect: Unified streaming vision-language navigation via semantic–spatial fusion and latent predictive representation.arXiv preprint arXiv:2603.03739, 2026

work page arXiv 2026
[14]

Dygeovln: Infusing dynamic geometry foundation model into vision- language navigation.arXiv preprint arXiv:2603.21269, 2026

Xiangchen Liu, Hanghan Zheng, Jeil Jeong, Minsung Yoon, Lin Zhao, Zhide Zhong, Haoang Li, and Sung-Eui Yoon. Dygeovln: Infusing dynamic geometry foundation model into vision- language navigation.arXiv preprint arXiv:2603.21269, 2026. 10

work page arXiv 2026
[15]

Internvla-n1: An open dual-system navigation foundation model with learned latent plans

InternNav Team. Internvla-n1: An open dual-system navigation foundation model with learned latent plans. 2025

2025
[16]

Embodied navigation foundation model.arXiv preprint arXiv:2509.12129, 2025

Jiazhao Zhang, Anqi Li, Yunpeng Qi, Minghan Li, Jiahang Liu, Shaoan Wang, Haoran Liu, Gengze Zhou, Yuze Wu, Xingxing Li, et al. Embodied navigation foundation model.arXiv preprint arXiv:2509.12129, 2025

work page arXiv 2025
[17]

Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks

Jiazhao Zhang, Kunyu Wang, Shaoan Wang, Minghan Li, Haoran Liu, Songlin Wei, Zhongyuan Wang, Zhizheng Zhang, and He Wang. Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks. InRobotics: Science and Systems (RSS), 2024

2024
[18]

Janusvln: Decoupling semantics and spatiality with dual implicit memory for vision-language navigation

Shuang Zeng, Dekang Qi, Xinyuan Chang, Feng Xiong, Shichao Xie, Xiaolong Wu, Shiyi Liang, Mu Xu, and Xing Wei. Janusvln: Decoupling semantics and spatiality with dual implicit memory for vision-language navigation. InInternational Conference on Learning Representations (ICLR), 2026

2026
[19]

Navforesee: A unified vision-language world model for hierarchical planning and dual-horizon navigation prediction.arXiv preprint arXiv:2512.01550, 2026

Fei Liu, Shichao Xie, Minghua Luo, Zedong Chu, Junjun Hu, Xiaolong Wu, and Mu Xu. Navforesee: A unified vision-language world model for hierarchical planning and dual-horizon navigation prediction.arXiv preprint arXiv:2512.01550, 2026

work page arXiv 2026
[20]

AstraNav-World: World Model for Foresight Control and Consistency

Junjun Hu, Jintao Chen, Haochen Bai, Minghua Luo, Shichao Xie, Ziyi Chen, Fei Liu, Zedong Chu, Xinda Xue, Botao Ren, Xiaolong Wu, Mu Xu, and Shanghang Zhang. Astranav-world: World model for foresight control and consistency.arXiv preprint arXiv:2512.21714, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[21]

Span-nav: Generalized spatial awareness for versatile vision-language navigation.arXiv preprint arXiv:2603.09163, 2026

Jiahang Liu, Tianyu Xu, Jiawei Chen, Lu Yue, Jiazhao Zhang, Zhiyong Wang, Minghan Li, Qisheng Zhao, Anqi Li, Qi Su, Zhizheng Zhang, and He Wang. Span-nav: Generalized spatial awareness for versatile vision-language navigation.arXiv preprint arXiv:2603.09163, 2026

work page arXiv 2026
[22]

Qiaolin Xia, Xiujun Li, Chunyuan Li, Yonatan Bisk, Zhifang Sui, Jianfeng Gao, Yejin Choi, and Noah A. Smith. Multi-view learning for vision-and-language navigation.arXiv preprint arXiv:2003.00857, 2020

work page arXiv 2003
[23]

Why only text: Empowering vision-and-language navigation with multi-modal prompts.arXiv preprint arXiv:2406.02208, 2024

Haodong Hong, Sen Wang, Zi Huang, Qi Wu, and Jiajun Liu. Why only text: Empowering vision-and-language navigation with multi-modal prompts.arXiv preprint arXiv:2406.02208, 2024

work page arXiv 2024
[24]

Diagnosing the environment bias in vision-and- language navigation.arXiv preprint arXiv:2005.03086, 2020

Yubo Zhang, Hao Tan, and Mohit Bansal. Diagnosing the environment bias in vision-and- language navigation.arXiv preprint arXiv:2005.03086, 2020

work page arXiv 2005
[25]

Rethinking the embodied gap in vision-and-language navigation: A holistic study of physical and visual disparities

Liuyi Wang, Xinyuan Xia, Hui Zhao, Hanqing Wang, Tai Wang, Yilun Chen, Chengju Liu, Qijun Chen, and Jiangmiao Pang. Rethinking the embodied gap in vision-and-language navigation: A holistic study of physical and visual disparities. InIEEE International Conference on Computer Vision (ICCV), pages 9455–9465, 2025

2025
[26]

Qwen2.5-VL Technical Report

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, et al. Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[27]

Llama-vid: An image is worth 2 tokens in large language models

Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. InEuropean Conference on Computer Vision (ECCV), pages 323–340, 2024

2024
[28]

LLaVA-Video: Video Instruction Tuning With Synthetic Data

Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Llava- video: Video instruction tuning with synthetic data.arXiv preprint arXiv:2410.02713, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[29]

Vila: On pre-training for visual language models

Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models.arXiv preprint arXiv:2312.07533, 2023

work page arXiv 2023
[30]

Instruction Tuning with GPT-4

Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4.arXiv preprint arXiv:2304.03277, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[31]

Correctnav: Self-correction flywheel empowers vision-language-action navigation model.arXiv preprint arXiv:2508.10416, 2025

Zhuoyuan Yu, Yuxing Long, Zihan Yang, Chengyan Zeng, Hongwei Fan, Jiyao Zhang, and Hao Dong. Correctnav: Self-correction flywheel empowers vision-language-action navigation model.arXiv preprint arXiv:2508.10416, 2025. 11

work page arXiv 2025
[32]

Provable benefits of un- supervised pre-training and transfer learning via single-index models.arXiv preprint arXiv:2502.16849, 2025

Taj Jones-McCormick, Aukosh Jagannath, and Subhabrata Sen. Provable benefits of un- supervised pre-training and transfer learning via single-index models.arXiv preprint arXiv:2502.16849, 2025

work page arXiv 2025
[33]

A survey on the robustness of computer vision models against common corruptions.arXiv preprint arXiv:2305.06024, 2024

Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. A survey on the robustness of computer vision models against common corruptions.arXiv preprint arXiv:2305.06024, 2024

work page arXiv 2024
[34]

Robustnav: Towards benchmarking robustness in embodied navigation

Prithvijit Chattopadhyay, Judy Hoffman, Roozbeh Mottaghi, and Ani Kembhavi. Robustnav: Towards benchmarking robustness in embodied navigation. InIEEE International Conference in Computer Vision (ICCV), pages 15691–15700, 2021

2021
[35]

Nomad: Goal masked diffusion policies for navigation and exploration.arXiv preprint arXiv:2310.07896, 2023

Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for navigation and exploration.arXiv preprint arXiv:2310.07896, 2023

work page arXiv 2023
[36]

Navdp: Learning sim-to-real navigation diffusion policy with privileged information guidance.arXiv preprint arXiv:2505.08712, 2025

Wenzhe Cai, Jiaqi Peng, Yuqiang Yang, Yujian Zhang, Meng Wei, Hanqing Wang, Yilun Chen, Tai Wang, and Jiangmiao Pang. Navdp: Learning sim-to-real navigation diffusion policy with privileged information guidance.arXiv preprint arXiv:2505.08712, 2025

work page arXiv 2025
[37]

Dynam3d: Dynamic layered 3d tokens empower vlm for vision-and-language navigation.arXiv preprint arXiv:2505.11383, 2025

Zihan Wang, Seungjun Lee, and Gim Hee Lee. Dynam3d: Dynamic layered 3d tokens empower vlm for vision-and-language navigation.arXiv preprint arXiv:2505.11383, 2025

work page arXiv 2025
[38]

Agentvln: Towards agentic vision-and-language navigation.arXiv preprint arXiv:2603.17670, 2026

Zihao Xin, Wentong Li, Yixuan Jiang, Ziyuan Huang, Bin Wang, Piji Li, Jianke Zhu, Jie Qin, and Sheng-Jun Huang. Agentvln: Towards agentic vision-and-language navigation.arXiv preprint arXiv:2603.17670, 2026

work page arXiv 2026
[39]

Navid-4d: Unleashing spatial intelligence in egocentric rgb-d videos for vision-and-language navigation

Haoran Liu, Weikang Wan, Xiqian Yu, Minghan Li, Jiazhao Zhang, Bo Zhao, Zhibo Chen, Zhongyuan Wang, Zhizheng Zhang, and He Wang. Navid-4d: Unleashing spatial intelligence in egocentric rgb-d videos for vision-and-language navigation. InIEEE International Conference on Robotics and Automation (ICRA), pages 10607–10615, 2025

2025
[40]

Navmorph: A self-evolving world model for vision-and-language navigation in continuous environments

Junyu Gao Xuan Yao and Changsheng Xu. Navmorph: A self-evolving world model for vision-and-language navigation in continuous environments. InIEEE International Conference in Computer Vision (ICCV), pages 5536–5546, 2025

2025
[41]

Etpnav: Evolving topological planning for vision-language navigation in continuous environments

Dong An, Hanqing Wang, Wenguan Wang, Zun Wang, Yan Huang, Keji He, and Liang Wang. Etpnav: Evolving topological planning for vision-language navigation in continuous environments. InIEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024

2024
[42]

Clash: Collaborative large-small hierarchical frame- work for continuous vision-and-language navigation,

Liuyi Wang, Zongtao He, Jinlong Li, Ruihao Xia, Mengxian Hu, Chenpeng Yao, Chengju Liu, Yang Tang, and Qijun Chen. Clash: Collaborative large-small hierarchical framework for continuous vision-and-language navigation.arXiv preprint arXiv:2512.10360, 2025

work page arXiv 2025
[43]

D3d-vlp: Dynamic 3d vision-language-planning model for embodied grounding and navigation.arXiv preprint arXiv:2512.12622, 2025

Zihan Wang, Seungjun Lee, Guangzhao Dai, and Gim Hee Lee. D3d-vlp: Dynamic 3d vision-language-planning model for embodied grounding and navigation.arXiv preprint arXiv:2512.12622, 2025

work page arXiv 2025
[44]

Available: https://arxiv.org/abs/2512.20940

Shuhao Ye, Sitong Mao, Yuxiang Cui, Xuan Yu, Shichao Zhai, Wen Chen, Shunbo Zhou, Rong Xiong, and Yue Wang. Etp-r1: Evolving topological planning with reinforcement fine-tuning for vision-language navigation in continuous environments.arXiv preprint arXiv:2512.20940, 2025

work page arXiv 2025
[45]

P 3Nav: End-to-end perception, prediction and planning for vision-and-language navigation,

Tianfu Li, Wenbo Chen, Haoxuan Xu, Xinhu Zheng, and Haoang Li. P 3nav: End-to-end perception, prediction and planning for vision-and-language navigation.arXiv preprint arXiv:2603.17459, 2026

work page arXiv 2026
[46]

Jiaqi Chen, Bingqian Lin, Xinmin Liu, Lin Ma, Xiaodan Liang, and Kwan-Yee K. Wong. Affordances-oriented planning using foundation models for continuous vision-language naviga- tion.arXiv preprint arXiv:2407.05890, 2024

work page arXiv 2024
[47]

Abot-n0: Technical report on the vla foundation model for versatile embodied navigation

Zedong Chu, Shichao Xie, Xiaolong Wu, Yanfen Shen, Minghua Luo, Zhengbo Wang, Fei Liu, Xiaoxu Leng, Junjun Hu, Mingyang Yin, Jia Lu, Yingnan Guo, Kai Yang, Jiawei Han, Xu Chen, et al. Abot-n0: Technical report on the vla foundation model for versatile embodied navigation. arXiv preprint arXiv:2602.11598, 2026. 12

work page arXiv 2026
[48]

Efficient-vln: A training-efficient vision-language navigation model,

Duo Zheng, Shijia Huang, Yanyang Li, and Liwei Wang. Efficient-vln: A training-efficient vision-language navigation model.arXiv preprint arXiv:2512.10310, 2025

work page arXiv 2025
[49]

Let’s reward step-by-step: Step-aware con- trastive alignment for vision-language navigation in continuous environments.arXiv preprint arXiv:2603.09740, 2026

Haoyuan Li, Rui Liu, Hehe Fan, and Yi Yang. Let’s reward step-by-step: Step-aware con- trastive alignment for vision-language navigation in continuous environments.arXiv preprint arXiv:2603.09740, 2026

work page arXiv 2026
[50]

thinking with images

Weiye Zhu, Zekai Zhang, Xiangchen Wang, Hewei Pan, Teng Wang, Tiantian Geng, Rongtao Xu, and Feng Zheng. Navida: Vision-language navigation with inverse dynamics augmentation. arXiv preprint arXiv:2601.18188, 2026

work page arXiv 2026
[51]

Decovln: Decoupling observation, reasoning, and correction for vision-and-language navigation

Zihao Xin, Wentong Li, Yixuan Jiang, Bin Wang, Runmin Cong, Jie Qin, and Shengjun Huang. Decovln: Decoupling observation, reasoning, and correction for vision-and-language navigation. arXiv preprint arXiv:2603.13133, 2026

work page arXiv 2026
[52]

InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, Zhaokai Wang, Zhe Chen, Hongjie Zhang, Ganlin Yang, Haomin Wang, et al. Internvl3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency.arXiv preprint arXiv:2508.18265, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[53]

DINOv2: Learning Robust Visual Features without Supervision

Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[54]

Foundationstereo: Zero-shot stereo matching

Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, and Stan Birchfield. Foundationstereo: Zero-shot stereo matching. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5249–5260, 2025

2025
[55]

Open-nav: Exploring zero-shot vision-and-language navigation in continuous environment with open-source llms

Yanyuan Qiao, Wenqi Lyu, Hui Wang, Zixu Wang, Zerui Li, Yuan Zhang, Mingkui Tan, and Qi Wu. Open-nav: Exploring zero-shot vision-and-language navigation in continuous environment with open-source llms. InIEEE International Conference on Robotics and Automation (ICRA), pages 6710–6717, 2025

2025
[56]

Dreamnav: A trajectory-based imaginative framework for zero-shot vision-and-language navigation.arXiv preprint arXiv:2509.11197, 2025

Yunheng Wang, Yuetong Fang, Taowen Wang, Yixiao Feng, Yawen Tan, Shuning Zhang, Peiran Liu, Yiding Ji, and Renjing Xu. Dreamnav: A trajectory-based imaginative framework for zero-shot vision-and-language navigation.arXiv preprint arXiv:2509.11197, 2025

work page arXiv 2025
[57]

OpenAI GPT-5 System Card

Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, Akshay Nathan, Alan Luo, Alec Helyar, Aleksander Madry, Aleksandr Efremov, et al. Openai gpt-5 system card.arXiv preprint arXiv:2601.03267, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[58]

Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, Luke Marris, Sam Petulla, Colin Gaffney, Asaf Aharoni, Nathan Lintz, et al. Gemini 2.5: Pushing the frontier with advanced reasoning.arXiv preprint arXiv:2507.06261, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[59]

Preset” column denotes the maximum sampling radius for generating the noisy prior, while “Actual

Anthropic. System card: Claude opus 4 and claude sonnet 4. https://www.anthropic.com/ claude-4-system-card. 13 A Details of the Pilot Studies A.1 Experimental Setup for Visual Uncertainty To examine whether the performance degradation of current VLN agents is associated with visual uncertainty, we conduct controlled pilot studies on two representative ope...