Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Aerial language navigation works by splitting the job between two drones: one reasons over the whole area, the other flies low and confirms the target.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Two drones at different altitudes, one guided by a vision-language model and one by a local navigator, reach targets more often than single-drone baselines on a new UAV navigation benchmark.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A genuinely new task and dataset for dual-UAV VLN, but the paper's central claim that dual-altitude collaboration drives the gains is not tested because every ablation keeps both UAVs. the 5 major comments →

arxiv 2508.15232 v1 pith:2SBRKMRM submitted 2025-08-21 cs.CV

AeroDuo: Aerial Duo for UAV-based Vision and Language Navigation

Ruipu Wu , Yige Zhang , Jinyu Chen , Linjiang Huang , Shifeng Zhang , Xu Zhou , Liang Wang , Si Liu This is my paper
classification cs.CV
keywords Vision-and-Language NavigationUAV navigationDual-altitude collaborationMultimodal large language modelTarget-oriented instructionAerial embodied AIMulti-agent planningNavigation benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Vision-and-language navigation for drones has been stuck at very low success rates when the pilot only gives a target-oriented sentence. This paper proposes that the bottleneck is a single viewpoint and introduces DuAl-VLN, a task in which two UAVs cooperate: one flies high to build a broad picture of the area, the other flies low to navigate safely and verify the target. To support the task, the authors build HaL-13k, a dataset of 13,838 paired high-low altitude trajectories with language instructions, and propose AeroDuo, a system pairing a multimodal language model on the high UAV with a lightweight three-stage planner on the low UAV. On unseen maps, AeroDuo raises success rate to 16.57% from 6.86% for the best single-UAV baseline it compares against. The paper's central claim is that dual-altitude collaboration, not simply a better single agent, is what lets a UAV system act on target descriptions alone.

Core claim

On its own terms, the paper establishes a new task and a new benchmark. DuAl-VLN removes the need for iterated human dialogue or step-by-step route descriptions: the two-UAV system receives a single target-oriented instruction and must locate the object. AeroDuo operationalizes a division of labor. The high-altitude UAV's Pilot-LLM takes a stitched orthographic map, the historical trajectory, and the instruction, and outputs a probability map over the terrain, which is converted into a centroid waypoint; the low-altitude UAV then executes the waypoint with an RL collision-avoidance controller and calls a grounded detector to stop once the target is found. The reported results show 16.57% suc

What carries the argument

The load-bearing mechanism is the target probability map predicted by Pilot-LLM over a global orthophoto map—a ground-plane image formed by reprojecting and stitching the high UAV's bird's-eye views using its trajectory and point clouds. The map converts a spatial-reasoning problem that large language models handle poorly (direct coordinate output) into a segmentation-like map prediction that can be pretrained with referring segmentation and depth estimation. On the low UAV, the Multi-Stage Pathfinder turns that map into a sub-goal centroid, computes collision-free waypoints with A* on an occupancy map, follows them with a point-cloud-based RL controller, and terminates with an open-vocabula

Load-bearing premise

The load-bearing premise is that the best single-UAV baseline is a fair counterfactual for the two-UAV system; since AeroDuo's advantages—a global orthophoto map and a multimodal language model—are bundled with the second drone, no experiment shows that the second drone itself, rather than those components, produces the 9.71-point gain.

What would settle it

Run AeroDuo with the high-altitude UAV's map and policy replaced by the low-altitude UAV's own onboard bird's-eye image while keeping the same Pilot-LLM: if success rate stays above roughly 15%, the second aircraft is unnecessary. Alternatively, keep the two-UAV setup but replace Pilot-LLM with a non-LLM segmentation-based map predictor: if success stays high, the language model is not the source of the gain. Either result would settle where the effect lives.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A target sentence like "red car near a T-intersection" can replace step-by-step route descriptions and real-time human assistance, reducing operator workload in aerial search.
  • The learned probability-map interface lets a multimodal language model contribute to navigation without asking it to regress exact coordinates, a known weak point of LLMs.
  • The dual-UAV design communicates only minimal coordinate information between agents, so the collaboration is bandwidth-efficient in principle.
  • The HaL-13k dataset provides synchronized high-low trajectories with multi-modal sensor streams, enabling further work on altitude-dependent perception and joint planning.
  • The method generalizes across both the unseen-map split and the unseen-object split of the benchmark, with the unseen-map split being the harder one on the reported numbers.

Where Pith is reading between the lines

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

  • The reported gain is not fully isolated to the second aircraft: the high-altitude UAV brings both the global orthophoto map and the MLLM reasoning, so a single-UAV AeroDuo using the same map and model could be the actual driver. Running that variant is the missing experiment.
  • Because all evaluation happens in simulation with a 20-metre success radius, transfer to real drones—where localization error, latency, and communication limits are larger—remains an open question the paper does not address.
  • The closeness of the unseen-object and unseen-map success rates suggests that new layouts, not new object categories, are the harder generalization problem; future work could concentrate on map-level generalization.
  • The probability-map design is a reusable trick for embodied tasks where large models are poor at direct coordinate output: turn regression into map prediction and pretrain with segmentation-style auxiliary tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper introduces a new task, Dual-Altitude UAV Collaborative VLN (DuAl-VLN), in which a high-altitude UAV performs broad environmental reasoning and a low-altitude UAV performs precise navigation. To support the task, the authors construct HaL-13k, a dataset of 13,838 paired high/low-altitude trajectories with target-oriented instructions, including held-out map and held-out object validation sets. The proposed framework, AeroDuo, couples a multimodal LLM (Pilot-LLM) on the high-altitude UAV, which predicts a target probability map from a stitched orthophoto, with a Multi-Stage Pathfinder (A*, RL-based collision avoidance, and GroundingDINO) on the low-altitude UAV. Experiments compare AeroDuo against single-UAV baselines (CMA, TravelUAV, TravelUAV-L1) and report higher success rates, e.g., 16.57% SR vs. 6.86% on unseen maps.

Significance. If the central claim were fully supported, this would be a meaningful contribution: it is, to my knowledge, one of the first UAV-VLN benchmarks with explicit dual-altitude collaboration, and the dataset scale (13,838 trajectories) plus held-out scene/object splits are useful assets for the community. The modular framework is also a strength: the high-altitude Pilot-LLM is pretrained on auxiliary tasks, and the low-altitude pathfinder is decomposed into clearly separated stages, with minimal coordinate-only communication between the UAVs. However, the paper's headline conclusion that dual-altitude collaboration causes the observed improvement is not established by the current experimental design, because no comparison removes the high-altitude UAV while keeping the rest of the stack fixed. The reported numbers also lack error bars, and several dataset-construction choices are underdescribed. These issues are fixable but require additional experiments and clarification.

major comments (5)
  1. [§5.2, Table 1; §5.3, Table 2] The central claim that dual-altitude collaboration improves navigation is not supported by the reported comparisons. AeroDuo differs from the TravelUAV/CMA baselines in several non-collaborative dimensions: the Pilot-LLM (Qwen2-VL) backbone, the stitched global orthophoto map, large-scale MLLM pretraining, and the modular Multi-Stage Pathfinder. Table 2 ablates Pretrain/GMC/KWD/CFN, but every row retains the high-altitude UAV; the 'no GMC' row still feeds the current BEV image from the high-altitude UAV, not the low-altitude UAV's forward view. Thus, the observed SR gain could come from the MLLM, the orthophoto map, or the pretraining, even if two-altitude collaboration contributed nothing. Please add a control that removes the high-altitude UAV (or replaces its BEV inputs with the low-altitude UAV's camera) while keeping the same perception/planning stack, and report SR/SST for that var
  2. [Abstract; §1] The phrase 'a significant improvement of 9.71% in navigation success rates' is misleading. The difference 16.57% - 6.86% is 9.71 percentage points, not a 9.71% relative improvement (which would be approximately 142%). Please correct the wording throughout to '9.71-point improvement' or state the relative improvement explicitly.
  3. [§5.1, Tables 1–2] No error bars, confidence intervals, or multiple-seed variance are reported. Each validation split contains 175 episodes, so a 16.57% versus 14.86% SR difference corresponds to 29 versus 26 successes, which may be within sampling noise. Please report results over at least three seeds or provide exact binomial confidence intervals for the main comparisons and ablations.
  4. [§3.2; §4.4] The HaL-13k dataset construction is underdescribed. The paper does not state how the 'target-oriented language instructions' were generated (e.g., template-based, human-written, or LLM-generated), whether they were filtered for resolvability, or what object/attribute distribution they cover. It also does not specify the split of the 14 scenes into training and unseen-map sets beyond '2 scenes', the number of trajectories per scene, or the criteria for selecting held-out object categories. These details are essential for assessing generalization claims and for reproducibility.
  5. [§4.3, Eq. (4)] The sub-goal is computed as a sum over M_tau(i,j)*[i,j] without normalization. Since M_tau is the output of a sigmoid (Eq. (3)), it is not a normalized probability distribution unless additional normalization is applied. As written, the 'centroid' is a weighted sum up to scale, which would bias the waypoint. Please either normalize M_tau before applying Eq. (4) or explicitly state the normalization step.
minor comments (5)
  1. [§5.1] Typo: 'Path Lengh' should be 'Path Length'.
  2. [§5.2] The Random baseline is described as selecting from 'four possible directions: forward, left, right, up, or down', which lists five actions.
  3. [§5.3, Table 2] The row labels in Table 2 are ambiguous: the positions of checkmarks under Pretrain/GMC/KWD/CFN do not unequivocally identify which components are active in each row. Please use explicit configuration names (e.g., 'Pretrain only', '+ GMC', '+ KWD', '+ CFN', 'Full') to make the ablation readable.
  4. [§4.4] The finetuning stage uses a Gaussian centered at the future position to generate ground-truth probability maps, but the choice of the Gaussian sigma and the future horizon k is not specified. These are free parameters that should be reported.
  5. [Figure 1] The annotations '1' and '2' in Figure 1 are not explained in the caption; please add a legend or clear referencing in the text.

Circularity Check

0 steps flagged

No significant circularity: the paper's gains are empirical comparisons, and the Pilot-LLM supervision is standard imitation learning rather than a self-fulfilling prediction.

full rationale

The closest candidate for circularity is the Pilot-LLM finetuning stage (Sec. 4.4), where ground-truth probability maps are generated from the future positions of the A*-planned low-altitude trajectories in HaL-13k. This is supervised imitation of the expert planner, not a prediction that is forced by construction: the evaluation metrics (SR, SST, SPL, OSR, NE) measure whether the low-altitude UAV reaches the true target within 20m on held-out unseen maps and unseen objects, and no metric compares the model's output to the training labels or to the A* path used to generate them. The model must generalize from BEV/orthophoto inputs and language instructions to new environments, so the reported success rates are not equivalent to re-displaying the training labels. The main comparison against CMA, TravelUAV, and TravelUAV(L1) is an empirical benchmark; although the absence of a single-UAV AeroDuo ablation means the 9.71% gain cannot be uniquely attributed to dual-altitude collaboration rather than to the MLLM, orthophoto map, pretraining, or modular pathfinder, that is an experimental-confound/validity limitation, not a circular reduction of the claim to its inputs. OpenUAV [54] is a same-group prior platform and baseline, but it is used as a simulator/dataset source and comparison point, not as load-bearing evidence for the dual-altitude collaboration claim. No self-citation chain, uniqueness theorem, or ansatz-smuggling is load-bearing. Therefore no circular step is identifiable by the standards required here.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The core result rests on a stack of external components (Qwen2-VL, GroundingDINO, A*, PPO) and on the dataset's expert trajectories. The only custom free parameters are the training-label Gaussian spread and detector threshold. The main load-bearing assumptions are the accuracy of the orthographic map stitching and sim-to-sim transfer of the navigator.

free parameters (3)
  • Gaussian sigma for ground-truth probability maps = not reported
    Used in Sec. 4.4 to spread the target probability map around the future low-altitude position; its value affects training labels and is chosen by hand.
  • GroundingDINO detection confidence threshold = not reported
    Used in Sec. 4.3's Target Localization to decide when the target is found; not specified in the paper.
  • maximum number of stitched historical images = 5
    Limits the orthophoto map size (Sec. 4.2); an arbitrary design choice that affects map coverage.
axioms (4)
  • domain assumption The orthographic projection pipeline in Eq. (1) yields an accurate global map from up to five stitched BEV images.
    Stated in Sec. 4.2; if stitching is inaccurate, the Pilot-LLM's probability map is unreliable.
  • domain assumption The RL navigator trained in Isaac Sim transfers to the OpenUAV simulator without a domain gap.
    Sec. 4.3 asserts the point-cloud-only navigator 'facilitates seamless adaptation across different simulated environments', but no sim-to-sim transfer evaluation is provided.
  • domain assumption GroundingDINO can detect all target object categories from the language description in low-altitude views.
    Sec. 4.3 uses it as the target localization detector; its zero-shot open-set ability is assumed.
  • domain assumption The expert trajectories in HaL-13k are optimal enough to serve as training labels for navigation.
    Sec. 3.2 generates low-altitude paths via A* on occupancy maps; the optimality of A* paths is assumed.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AeroDuo: Aerial Duo for UAV-based Vision and Language Navigation." pith.science (2026). https://pith.science/paper/2SBRKMRM

@misc{pith2026250815232,
  author       = {Pith},
  title        = {Pith review of: AeroDuo: Aerial Duo for UAV-based Vision and Language Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2SBRKMRM}},
  note         = {Machine review of arXiv:2508.15232}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Aerial Vision-and-Language Navigation (VLN) is an emerging task that enables Unmanned Aerial Vehicles (UAVs) to navigate outdoor environments using natural language instructions and visual cues. However, due to the extended trajectories and complex maneuverability of UAVs, achieving reliable UAV-VLN performance is challenging and often requires human intervention or overly detailed instructions. To harness the advantages of UAVs' high mobility, which could provide multi-grained perspectives, while maintaining a manageable motion space for learning, we introduce a novel task called Dual-Altitude UAV Collaborative VLN (DuAl-VLN). In this task, two UAVs operate at distinct altitudes: a high-altitude UAV responsible for broad environmental reasoning, and a low-altitude UAV tasked with precise navigation. To support the training and evaluation of the DuAl-VLN, we construct the HaL-13k, a dataset comprising 13,838 collaborative high-low UAV demonstration trajectories, each paired with target-oriented language instructions. This dataset includes both unseen maps and an unseen object validation set to systematically evaluate the model's generalization capabilities across novel environments and unfamiliar targets. To consolidate their complementary strengths, we propose a dual-UAV collaborative VLN framework, AeroDuo, where the high-altitude UAV integrates a multimodal large language model (Pilot-LLM) for target reasoning, while the low-altitude UAV employs a lightweight multi-stage policy for navigation and target grounding. The two UAVs work collaboratively and only exchange minimal coordinate information to ensure efficiency.

Figures

Figures reproduced from arXiv: 2508.15232 by Jinyu Chen, Liang Wang, LinJiang Huang, Ruipu Wu, Shifeng Zhang, Si Liu, Xu Zhou, Yige Zhang.

Figure 1
Figure 1. Figure 1: In the DuAl-VLN task, the two UAVs operating at distinct altitudes achieve collaborative target search through [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of dataset statistics and trajectory collec [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The AeroDuo framework integrates multiple data sources to achieve precise UAV navigation. We input a global [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of Multi-Stage Pathfinder (MSP) on low [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of UAV target search performance. All [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. No Training, Better Flights: Test-Time Scaled VLMs for UAV Navigation

    cs.CV 2026-07 conditional novelty 4.0

    Test-time scaling—parallel candidate generation, iterative self-correction, and multi-criteria selection—improves a frozen UAV navigation VLM's success rate by about 2 percentage points on the TravelUAV benchmark.

Reference graph

Works this paper leans on

77 extracted references · 52 canonical work pages · cited by 1 Pith paper

  1. [1]

    Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünder- hauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. 2018. Vision-and- language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR

    work page 2018

  2. [2]

    Christopher Berner, Greg Brockman, Brooke Chan, Vicki Cheung, Przemysław Dębiak, Christy Dennison, David Farhi, Quirin Fischer, Shariq Hashme, Chris Hesse, et al. 2019. Dota 2 with large scale deep reinforcement learning. arXiv (2019)

    work page 2019

  3. [3]

    Prabin Bhandari, Antonios Anastasopoulos, and Dieter Pfoser. 2023. Are large language models geospatially knowledgeable?. In Proceedings of the 31st ACM International Conference on Advances in Geographic Information Systems . 1–4

    work page 2023

  4. [4]

    Ilker Bozcan and Erdal Kayacan. 2020. AU-AIR: A Multi-modal Unmanned Aerial Vehicle Dataset for Low Altitude Traffic Surveillance. arXiv preprint (2020)

    work page 2020

  5. [5]

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. 2023. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818 (2023)

    Pith/arXiv arXiv 2023

  6. [6]

    Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. 2017. Matterport3D: Learning from RGB-D Data in Indoor Environments. 3DV (2017)

    work page 2017

  7. [7]

    Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. 2019. Touchdown: Natural language navigation and spatial reasoning in visual street environments. In CVPR

    work page 2019

  8. [8]

    Jinyu Chen, Chen Gao, Erli Meng, Qiong Zhang, and Si Liu. 2022. Reinforced structured state-evolution for vision-language navigation. In CVPR

    work page 2022

  9. [9]

    Jiaqi Chen, Bingqian Lin, Ran Xu, Zhenhua Chai, Xiaodan Liang, and Kwan-Yee K Wong. 2024. Mapgpt: Map-guided prompting for unified vision-and-language navigation. arXiv preprint arXiv:2401.07314 (2024)

    Pith/arXiv arXiv 2024

  10. [10]

    Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. 2021. History aware multimodal transformer for vision-and-language navigation. NeurIPS (2021)

    work page 2021

  11. [11]

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality. See https://vicuna. lmsys. org (accessed 14 April 2023) 2, 3 (2023), 6

    2023

  12. [12]

    Stepan Dergachev and Konstantin Yakovlev. 2021. Distributed Multi-Agent Navigation Based on Reciprocal Collision Avoidance and Locally Confined Multi- Agent Path Finding. In CASE

    work page 2021

  13. [13]

    Yue Fan, Winson Chen, Tongzhou Jiang, Chun Zhou, Yi Zhang, and Xin Eric Wang

    work page

  14. [14]

    Yue Fan, Shilei Chu, Wei Zhang, Ran Song, and Yibin Li. 2020. Learn by observa- tion: Imitation learning for drone patrolling from videos of a human navigator. In IROS

    work page 2020

  15. [15]

    Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis- Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. 2018. Speaker-Follower Models for Vision-and-Language Navigation. NeurIPS (2018)

    work page 2018

  16. [16]

    Chen Gao, Jinyu Chen, Si Liu, Luting Wang, Qiong Zhang, and Qi Wu. 2021. Room-and-object aware knowledge reasoning for remote embodied referring expression. In CVPR

    work page 2021

  17. [17]

    Chen Gao, Si Liu, Jinyu Chen, Luting Wang, Qi Wu, Bo Li, and Qi Tian. 2023. Room-object entity prompting and reasoning for embodied referring expression. IEEE TPAMI (2023)

    work page 2023

  18. [18]

    Yunpeng Gao, Chenhui Li, Zhongrui You, Junli Liu, Zhen Li, Pengan Chen, Qizhi Chen, Zhonghan Tang, Liansheng Wang, Penghui Yang, et al. 2025. OpenFly: A Versatile Toolchain and Large-scale Benchmark for Aerial Vision-Language Navigation. arXiv preprint arXiv:2502.18041 (2025)

    arXiv 2025

  19. [19]

    Yunpeng Gao, Zhigang Wang, Linglin Jing, Dong Wang, Xuelong Li, and Bin Zhao. 2024. Aerial Vision-and-Language Navigation via Semantic-Topo-Metric Representation Guided LLM Reasoning. arXiv preprint arXiv:2410.08500 (2024)

    Pith/arXiv arXiv 2024

  20. [20]

    Alessandro Giusti, Jérôme Guzzi, Dan C Cireşan, Fang-Lin He, Juan P Rodríguez, Flavio Fontana, Matthias Faessler, Christian Forster, Jürgen Schmidhuber, Gianni Di Caro, et al. 2015. A machine learning approach to visual perception of forest trails for mobile robots. IEEE RAL (2015)

    work page 2015

  21. [21]

    Zonghao Guo, Ruyi Xu, Yuan Yao, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, and Gao Huang. 2025. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. In ECCV

    work page 2025

  22. [22]

    Peter E Hart, Nils J Nilsson, and Bertram Raphael. 1968. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics 4, 2 (1968), 100–107

    1968

  23. [23]

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, et al

  24. [24]

    Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould

  25. [25]

    Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. 2019. Stay on the path: Instruction fidelity in vision-and-language navigation. arXiv preprint (2019)

    work page 2019

  26. [26]

    Katie Kang, Suneel Belkhale, Gregory Kahn, Pieter Abbeel, and Sergey Levine

    work page

  27. [27]

    Xianghao Kong, Jinyu Chen, Wenguan Wang, Hang Su, Xiaolin Hu, Yi Yang, and Si Liu. 2024. Controllable navigation instruction generation with chain of thought prompting. In European Conference on Computer Vision. Springer, 37–54

    work page 2024

  28. [28]

    Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee

  29. [29]

    Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. 2020. Room-Across-Room: Multilingual Vision-and-Language Navigation with Dense Spatiotemporal Grounding. In EMNLP

    work page 2020

  30. [30]

    Jungdae Lee, Taiki Miyanishi, Shuhei Kurita, Koya Sakamoto, Daichi Azuma, Yutaka Matsuo, and Nakamasa Inoue. 2024. CityNav: Language-Goal Aerial Navigation Dataset with Geographic Information. arXiv preprint (2024)

    work page 2024

  31. [31]

    Hongyu Li, Jinyu Chen, Ziyu Wei, Shaofei Huang, Tianrui Hui, Jialin Gao, Xiaom- ing Wei, and Si Liu. 2025. LLaVA-ST: A Multimodal Large Language Model for Fine-Grained Spatial-Temporal Understanding. arXiv preprint arXiv:2501.08282 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    Jialu Li and Mohit Bansal. 2024. Panogen: Text-conditioned panoramic envi- ronment generation for vision-and-language navigation. Advances in Neural Information Processing Systems 36 (2024)

    work page 2024

  33. [33]

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al . 2023. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)

    Pith/arXiv arXiv 2023

  34. [34]

    Shubo Liu, Hongsheng Zhang, Yuankai Qi, Peng Wang, Yanning Zhang, and Qi Wu. 2023. Aerialvln: Vision-and-language navigation for uavs. In ICCV

    work page 2023

  35. [35]

    Youzhi Liu, Fanglong Yao, Yuanchang Yue, Guangluan Xu, Xian Sun, and Kun Fu. 2024. NavAgent: Multi-scale Urban Street View Fusion For UAV Embodied Vision-and-Language Navigation. arXiv preprint (2024)

    work page 2024

  36. [36]

    Antonio Loquercio, Ana I Maqueda, Carlos R Del-Blanco, and Davide Scaramuzza

    work page

  37. [37]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    Pith/arXiv arXiv 2017

  38. [38]

    Ryan Lowe, Yi I Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. 2017. Multi-agent actor-critic for mixed cooperative-competitive environments. NeurIPS (2017)

    work page 2017

  39. [39]

    András L Majdik, Charles Till, and Davide Scaramuzza. 2017. The Zurich urban micro aerial vehicle dataset. The IJRR (2017)

    work page 2017

  40. [40]

    Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al

  41. [41]

    Yash Vardhan Pant, Houssam Abbas, Rhudii A Quaye, and Rahul Mangharam

    work page

  42. [42]

    o (marl\

    Diego Perez-Liebana, Katja Hofmann, Sharada Prasanna Mohanty, Noburu Kuno, Andre Kramer, Sam Devlin, Raluca D Gaina, and Daniel Ionita. 2019. The multi- agent reinforcement learning in malm\" o (marl\" o) competition. arXiv preprint arXiv:1901.08129 (2019)

    Pith/arXiv arXiv 2019

  43. [43]

    Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. 2020. Reverie: Remote embodied visual referring expression in real indoor environments. In CVPR

    work page 2020

  44. [44]

    Yanyuan Qiao, Qianyi Liu, Jiajun Liu, Jing Liu, and Qi Wu. 2024. LLM as Copilot for Coarse-Grained Vision-and-Language Navigation. In European Conference on Computer Vision. Springer, 459–476

    work page 2024

  45. [45]

    Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. 2024. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14313–14323

    2024

  46. [46]

    Mikayel Samvelyan, Tabish Rashid, Christian Schroeder De Witt, Gregory Far- quhar, Nantas Nardelli, Tim GJ Rudner, Chia-Man Hung, Philip HS Torr, Jakob Foerster, and Shimon Whiteson. 2019. The starcraft multi-agent challenge. arXiv (2019)

    work page 2019

  47. [47]

    arXiv preprint arXiv:2108.10470 (2021)

    Isaac gym: High performance gpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470 (2021)

    Pith/arXiv arXiv 2021

  48. [48]

    Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. 2018. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. InField and Service Robotics: Results of the 11th International Conference . Springer, 621–635. MM ’25, October 27–31, 2025, Dublin, Ireland Ruipu Wu et al

    work page 2018

  49. [49]

    In ICCPS

    Fly-by-logic: Control of multi-drone fleets with temporal logic objectives. In ICCPS. IEEE, 186–197

    work page

  50. [50]

    Nikolai Smolyanskiy, Alexey Kamenev, Jeffrey Smith, and Stan Birchfield. 2017. Toward low-flying autonomous MAV trail navigation using deep neural networks for environmental awareness. In IROS

    work page 2017

  51. [51]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    Pith/arXiv arXiv 2023

  52. [52]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. 2024. Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191 (2024)

    Pith/arXiv arXiv 2024

  53. [53]

    Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. 2022. Less is more: Generating grounded navigation instructions from landmarks. In CVPR

    work page 2022

  54. [54]

    Xiangyu Wang, Donglin Yang, Ziqin Wang, Hohin Kwan, Jinyu Chen, Wenjun Wu, Hongsheng Li, Yue Liao, and Si Liu. 2024. Towards Realistic UAV Vision- Language Navigation: Platform, Benchmark, and Methodology

    work page 2024

  55. [55]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  56. [56]

    Yutaro Yamada, Yihan Bao, Andrew K Lampinen, Jungo Kasai, and Ilker Yildirim

    work page

  57. [57]

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. 2024. Depth anything v2. NeurIPS 37 (2024), 21875–21911

    work page 2024

  58. [58]

    Abhik Singla, Sindhu Padakandla, and Shalabh Bhatnagar. 2019. Memory-based deep reinforcement learning for obstacle avoidance in UAV with limited environ- ment knowledge. IEEE TIST (2019)

    work page 2019

  59. [59]

    Haitian Zeng, Xiaohan Wang, Wenguan Wang, and Yi Yang. 2023. Kefa: A Knowl- edge Enhanced and Fine-grained Aligned Speaker for Navigation Instruction Generation. arXiv preprint arXiv:2307.13368 (2023)

    Pith/arXiv arXiv 2023

  60. [60]

    Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zhang, and He Wang. 2024. Navid: Video-based vlm plans the next step for vision-and-language navigation. arXiv preprint arXiv:2402.15852 (2024)

    Pith/arXiv arXiv 2024

  61. [61]

    Yue Zhang and Parisa Kordjamshidi. 2023. VLN-Trans, Translator for the Vision and Language Navigation Agent. In ACL

    work page 2023

  62. [62]

    Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. 2022. Target-Driven Structured Transformer Planner for Vision-Language Navigation. In ACM MM

    work page 2022

  63. [63]

    Zhonghan Zhao, Kewei Chen, Dongxu Guo, Wenhao Chai, Tian Ye, Yanting Zhang, and Gaoang Wang. 2024. Hierarchical auto-organizing system for open- ended multi-agent navigation. arXiv (2024)

    work page 2024

  64. [64]

    Zhefan Xu, Xinming Han, Haoyu Shen, Hanyu Jin, and Kenji Shimada. 2025. Navrl: Learning safe flight in dynamic environments. IEEE RAL (2025)

    work page 2025

  65. [65]

    Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. 2024. MLVU: A Comprehensive Bench- mark for Multi-Task Long Video Understanding. arXiv preprint arXiv:2406.04264 (2024)

    Pith/arXiv arXiv 2024

  66. [66]

    arXiv preprint arXiv:2310.14540 (2023)

    Evaluating spatial understanding of large language models. arXiv preprint arXiv:2310.14540 (2023)

    Pith/arXiv arXiv 2023

  67. [68]

    Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg

  68. [75]

    Duo Zheng, Shijia Huang, Lin Zhao, Yiwu Zhong, and Liwei Wang. 2024. To- wards learning a generalist model for embodied navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 13624–13634

    work page 2024

  69. [77]

    Fengda Zhu, Vincent CS Lee, and Rui Liu. 2024. Communicative and Coopera- tive Learning for Multi-agent Indoor Navigation. In Pacific-Asia Conference on Knowledge Discovery and Data Mining . Springer, 273–285

    work page 2024

  70. [2016]

    Modeling context in referring expressions. In ECCV. Springer, 69–85

    work page

  71. [2017]

    arXiv preprint arXiv:1707.06347 (2017)

    Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

    Pith/arXiv arXiv 2017

  72. [2018]

    IEEE RAL 3, 2 (2018), 1088–1095

    Dronet: Learning to fly by driving. IEEE RAL 3, 2 (2018), 1088–1095

    work page 2018

  73. [2019]

    arXiv preprint (2019)

    Generalization through simulation: Integrating simulated and real data into deep reinforcement learning for vision-based autonomous flight. arXiv preprint (2019)

    work page 2019

  74. [2020]

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

    work page

  75. [2021]

    A recurrent vision-and-language bert for navigation. In CVPR

    work page

  76. [2023]

    In Findings of the Association for Computational Linguistics: ACL 2023

    Aerial Vision-and-Dialog Navigation. In Findings of the Association for Computational Linguistics: ACL 2023 . Association for Computational Linguistics, Toronto, Canada, 3043–3061

    work page 2023

  77. [2024]

    MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework. In ICLR

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.