Pith. sign in

REVIEW 4 major objections 4 minor 77 references

This paper argues that sustained occlusion, not visible-frame error, is the defining failure mode of UAV target tracking, and that a spatially grounded vision-language-action model trained for recovery can track through it.

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 ↗

T0 review · deepseek-v4-flash

2026-08-02 00:25 UTC pith:N73BUWF2

load-bearing objection Real open-loop gains and a novel occlusion-centric training recipe, but the 'closed-loop recovery' claim is only half-tested because evaluation feeds ground-truth state history - still deserves a serious referee. the 4 major comments →

arxiv 2607.15004 v1 pith:N73BUWF2 submitted 2026-07-16 cs.RO

CosFly-VLA: A Spatially Aware Vision-Language-Action Model for UAV Tracking

Ruilong Ren , Songsheng Cheng , Yunpeng Zhou , Hanxuan Chen , Xiangyue Wang , Tianle Zeng , Shuai Yuan , Binbo Li
show 4 more authors
Hanzhong Guo Ji Pei Da Zhang Kangli Wang
This is my paper
classification cs.RO
keywords UAV target trackingVision-Language-Action modelsOcclusion recoverySpatially grounded pretrainingCurriculum learningChain-of-thought reasoningClosed-loop reinforcement learningCARLA simulation
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.

CosFly-VLA is built on a simple organizing claim: UAV tracking fails most when the target disappears behind buildings, so the policy should be trained and evaluated for recovery rather than for visible-frame following. The paper reports that a 0.8B vision-language-action model, pretrained on aerial spatial reasoning and fine-tuned through occlusion-graded curricula, chain-of-thought reasoning, and closed-loop reinforcement learning, reduces open-loop displacement error by about a third relative to generalist VLA baselines and improves closed-loop success rate by up to 17 percentage points on seen maps. The central mechanism is a structured output interface that predicts the target box, its visibility, and an eight-step flight-command chunk from shared hidden states. If the results hold, occlusion-robust UAV tracking is achievable with a model of modest size, at least in the simulated environments studied. The strongest caveat, stated by the authors, is that evaluation feeds ground-truth state history rather than the model's own predicted feedback.

Core claim

The paper's central claim is that sustained occlusion, not visible-frame error, is the defining failure mode of UAV target tracking, and that a vision-language-action policy which jointly predicts target visibility, target location, and an eight-step waypoint chunk can maintain a spatial hypothesis about an unseen target and re-acquire it after occlusion. To realize this, CosFly-VLA couples a frozen vision-language backbone with three lightweight task heads selected through meta-query spans: a flow-matching action expert for continuous flight commands, an MLP bounding-box head, and a visibility head. Training proceeds through spatially grounded continued pretraining on a 500k-sample aerial s

What carries the argument

The load-bearing object is the structured prediction interface: boundary-tagged meta-query spans that route hidden states from a frozen Qwen3.5 backbone to three decoupled heads—a 12-layer flow-matching Diffusion Transformer action expert for the 8-step 4-DoF waypoint chunk, an MLP bounding-box head, and a visibility head. This lets the model answer visible targets from appearance and occluded targets from spatial reasoning without serializing numbers through text. The training recipe supplies the spatial and occlusion priors: CPT injects aerial depth/distance/3-D reasoning, the curriculum SFT keeps long-occlusion windows (including 4-of-5 and 5-of-5 invisible inputs), CoT teaches recovery-o

Load-bearing premise

The evaluation feeds the policy ground-truth state history (previous boxes, poses, and visibility flags) rather than its own predicted feedback, so the reported closed-loop recovery gains may not persist when the policy must rely on its own predictions.

What would settle it

Run the same closed-loop evaluation while feeding back the model's own predicted target boxes and visibility flags instead of ground truth; if the success-rate advantage over generalist baselines shrinks or reverses, the central recovery claim is not robust. A second check: audit whether episodes with the longest occlusion streaks are the ones where the CoT trace's predicted reappearance region matches the target's actual reappearance.

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

If this is right

  • Tracking benchmarks for UAVs should grade sustained occlusion explicitly: the paper defines Easy/Medium/Hard by line-of-sight visibility and shows that the hardest slices drive most of the gains.
  • Keeping long-occlusion windows in the training data, instead of filtering them out, is what lets the model learn re-acquisition; removing this data would erase the Hard-slice gains.
  • Closed-loop RL adds success and safety beyond offline imitation: it improves success rate and stand-off distance error, though it trades off a little track continuity on seen-test.
  • Spatially grounded continued pretraining is the step that most helps cross-map generalization, while CoT supervision most helps seen-test hard occlusion.
  • A model of 0.8B parameters can achieve these gains; the 2B/9B variants are not evaluated in this paper.

Where Pith is reading between the lines

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

  • [Editorial inference] If predicted-state feedback is used in deployment, the recovery gains may shrink because the policy will see its own drifted boxes and visibility flags; this is the natural next stress test.
  • [Editorial inference] The same recipe—spatial continued pretraining, occlusion curriculum, and recovery-oriented reasoning—could transfer to other partial-observability tracking settings, such as ground robots tracking targets behind walls or vehicles.
  • [Editorial inference] A testable extension would be to measure whether the CoT reasoning traces are causally responsible for the Hard-slice gains or merely correlated with extra training on occlusion-heavy windows.
  • [Editorial inference] Since all evaluation is in simulation, real-world transfer hinges on whether the spatial pretraining and visibility estimator survive domain shift in lighting, weather, and occluder geometry.

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

4 major / 4 minor

Summary. CosFly-VLA formulates UAV target tracking under sustained occlusion as a closed-loop recovery problem and proposes a VLA architecture with a meta-query interface, decoupled action/bbox/visibility heads, and a four-stage recipe (spatially grounded CPT, curriculum SFT, CoT, closed-loop RL). The paper reports open-loop ADE reductions of 34.1% (seen) and 35.3% (unseen) over OpenVLA for SFT+CoT, and closed-loop SR gains of +17pp (seen) and +2pp (unseen) for the RL variant, along with ablations and qualitative examples. The authors are transparent that all evaluations use ground-truth state history in both regimes and that predicted-state feedback is left to future work.

Significance. If the closed-loop claims were fully supported, this would be a meaningful advance: occlusion-robust recovery is an important and under-addressed capability for aerial VLAs, and the training recipe (large-scale CPT, curriculum SFT, CoT, RL) is a substantial engineering contribution. The open-loop protocol, occlusion-graded splits, and cumulative ablations are clearly defined and internally consistent. However, the closed-loop evaluation is not closed over the policy's own perception state, and the main baselines are not trained on the same task interface or data. These issues make the headline closed-loop recovery claim conditional rather than established; the open-loop structured-prediction results are the strongest evidence currently presented.

major comments (4)
  1. [Section 3, Eq. (1); Section 4.3] The 'closed-loop recovery' claim is undermined by the use of ground-truth state history in both regimes. Section 3 states 'Unless otherwise stated, s_prev is populated from ground-truth state history in both open-loop and closed-loop evaluation,' and Section 4.3 confirms 'the box/visibility history used for evaluation remains ground truth.' The policy is therefore never forced to recover from its own incorrect boxes or visibility flags; reported SR/ADE gains may largely reflect privileged state, not recovery competence. Please add experiments with predicted-state feedback (even if performance degrades) or consistently restrict the claim to 'action-closed-loop under oracle state history' and soften the abstract/title. Also clarify whether the ground-truth target box is supplied during invisible frames; if so, that is an especially strong oracle.
  2. [Tables 4–6; Section 5.2] The baseline comparison is not controlled. OpenVLA, π0, and π0.5 are not fine-tuned on the same SFT/CPT interface or task data (the paper describes them as 'general VLA baselines' but does not report any fine-tuning), yet they are used as primary references for relative improvements of 34.1%/35.3% ADE and +17pp SR. Without fine-tuning these baselines on the same training data (or clearly labeling them as zero-shot/unadapted), the headline gains conflate task adaptation with model quality. Please add fine-tuned baselines or explicitly frame the comparison as zero-shot and avoid 'strongest general VLA baseline' language.
  3. [Eq. (9); Table 6] The RL reward is composed of exactly the closed-loop evaluation quantities: stand-off distance with d*=28 m, IoU with the ground-truth box, success, and collision. Since the RL stage also receives ground-truth state history, the closed-loop improvements over the SFT checkpoint are expected and do not establish generalization to settings where these quantities must be estimated from the policy's own outputs. Please provide an evaluation with predicted boxes/visibility in the loop, or an ablation of reward terms, and report multi-seed variance (currently absent, as acknowledged in Section 7).
  4. [Abstract; Table 6] The headline '29.8% and 2.5%' pairs a large seen-test SR gain with a negligible unseen-test gain of +2pp (80%→82%), which is unlikely to be significant without confidence intervals. The cross-map generalization claim is therefore carried mainly by open-loop ADE, which is not a closed-loop metric. Please disaggregate the claims and explicitly state that the unseen closed-loop improvement is small and unquantified, rather than presenting the two numbers as comparable evidence.
minor comments (4)
  1. [Section 4.3 / Section 5.3] Please specify the map names/IDs in the 11-map RL split and confirm explicitly that Town01_Opt, Town03_Opt, and Town05_Opt used in unseen-test are disjoint from the RL training maps. The paper asserts disjointness in Section 5.1 but does not provide the map list, making the claim unverifiable.
  2. [Section 5.2] The phrase 'under the same prompt and I/O contract' is ambiguous about whether the VLA baselines were fine-tuned or evaluated zero-shot. Clarify this in the protocol description.
  3. [Section 3, Eq. (2)] The episode-level Hard definition (occ_rate≥20%) and the window-level Hard definition (3–4 invisible history frames) are both called 'Hard.' The distinction is stated, but a short cross-reference or naming convention (e.g., 'episode-Hard' vs 'window-Hard') would reduce reader confusion.
  4. [Table 2] The RL row says '11-map split: 440 train + 110 test' but does not state whether the 20 calibration paths are included in the evaluation or only used for normalization. Please clarify their role.

Circularity Check

0 steps flagged

No significant circularity: the empirical training-recipe evaluation is self-contained, and the stated oracle-state-history caveat is an evaluation-validity limitation rather than a circular derivation.

full rationale

The paper's central claims are empirical and held-out. Open-loop ADE reductions are measured on occlusion-rebalanced held-out windows against generalist VLA baselines, with unseen-test maps disjoint from the SFT pool. The RL stage optimizes Eq. (9) on 440 training paths and is evaluated on separate seen/unseen episodes; although the reward shares terms with the reported dErr and success metrics, this is standard objective-metric alignment, not a parameter fitted to the evaluation set. The explicit limitation that s_prev is ground-truth state history in both open-loop and closed-loop evaluation (Section 3; reiterated in Sections 6-7 and the conclusion) narrows the 'closed-loop recovery' claim to action-loop closure under privileged perception-state history, and means predicted-state feedback remains untested; this is a real validity caveat but not a circular definition or a by-construction reduction. Self-citations to CosFly-Track [55] and related internal work supply the dataset and framing but do not substitute for the empirical comparisons. No equation-level reduction of a reported result to its training input was found.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central claim rests on many hand-set hyperparameters and on simulation-only, oracle-history evaluation. No new physical entities are introduced, but the empirical result is carried by a large internal dataset and a training recipe whose components are not independently verifiable from the paper.

free parameters (7)
  • stand-off distance d* = 28 m
    Hand-set target distance in the RL reward (Eq. 9) and in the dErr evaluation metric; central to reward design.
  • RL reward weights wd, wiou = 1.0, 0.5
    Hand-set balance between distance error and IoU in Eq. 9; affects learned behavior.
  • action normalization scales = (5, 5, 2, 30)
    Per-dimension normalization of (Δx, Δy, Δz, Δψ) for the flow-matching action head; chosen by hand.
  • curriculum bucket weights = 1:1.5:2:3
    Stage-3 SFT weights over cruise/turn/loose-avoid/tight-avoid buckets; hand-set to emphasize failure-prone samples.
  • occlusion thresholds = vis_frac<0.6; distance gate (2,50) m; Easy/Medium/Hard at 5%/20% occ_rate
    Hand-set thresholds in Eq. 2 define difficulty grades used throughout evaluation and data mining.
  • loss weights (λa, λb, λv, λlm) = (1.0, 1.0, 0.05, 0.0); CoT λlm=0.1
    Hand-set multi-task loss weights in Eq. 8.
  • long-occlusion mining parameters = max_invisible=4, min_visible_history=0, max_alt_jump=3.0 m; 15% long-occ quota
    Hand-set data-mining choices that determine the hard-occlusion training distribution.
axioms (6)
  • domain assumption CARLA simulator dynamics and photorealism are a valid proxy for real UAV tracking.
    All closed-loop results are in CARLA; Section 7 explicitly says real-world robustness is not established.
  • domain assumption Ground-truth state history is a valid substitute for predicted-state feedback in closed-loop evaluation.
    Section 3 states s_prev is GT in both regimes; Section 7 leaves predicted-box feedback to future work.
  • domain assumption The geometric ray-cast visibility test (five LOS rays, 0.5 m margin) accurately labels target occlusion.
    Eq. 2 defines visibility; all difficulty grades and data mining depend on it.
  • domain assumption External aerial VQA/grounding datasets improve UAV occlusion recovery when used as continued pretraining.
    Table 1 CPT mixture; no experiment isolates external vs internal CPT data.
  • domain assumption Qwen3.5 as frozen backbone with LoRA retains sufficient spatial reasoning for UAV tracking.
    Architecture choice in Section 4.2.1; no head-to-head against fully fine-tuned backbone.
  • standard math Flow-matching / rectified-flow objective is a suitable action distribution model.
    Eq. 7 training loss; standard flow matching [31, 11, 4].

pith-pipeline@v1.3.0-alltime-deepseek · 25014 in / 11924 out tokens · 122586 ms · 2026-08-02T00:25:57.218373+00:00 · methodology

0 comments
read the original abstract

Dynamic target tracking is essential for Unmanned Aerial Vehicles (UAVs) operating in complex urban environments, where both the target and the camera viewpoint change continuously. Existing Vision-Language-Action (VLA) policies can track visible targets effectively, but their performance often degrades when buildings, vegetation, or roadside objects block the line of sight. During sustained occlusion, a policy may lose the target state, execute actions toward an incorrect region, and amplify this error through subsequent observations until re-acquisition becomes impossible. To this end, we present CosFly-VLA, a spatially aware VLA model that jointly grounds the target, estimates its visibility, and generates continuous flight actions through a structured prediction interface. To train this policy, we use a large-scale recipe over diverse data sources. Spatially Grounded Continued Pretraining (CPT) on a 500k mixed pool injects UAV-view depth, distance, and 3-D spatial reasoning. A three-stage Curriculum-based Supervised Fine-Tuning (SFT) process then specializes the tracker through multi-head warm-up followed by two-stage curriculum learning over natural and hard / long-occlusion data. Chain-of-Thought (CoT) training subsequently teaches recovery-oriented reasoning traces before structured answers. Finally, a closed-loop Reinforcement Learning (RL) stage optimizes tracking behavior with a multi-component reward covering stand-off tracking, grounding quality, collision avoidance, and task success. Relative to OpenVLA, CosFly-VLA-0.8B reduces open-loop Average Displacement Error (ADE) by 34.1% on seen-test and 35.3% on unseen-test. Closed-loop optimization improves Success Rate (SR) by 29.8% and 2.5%, respectively. These results demonstrate progress from visible-frame imitation toward spatially grounded action-closed-loop control, evaluated under a shared oracle state history.

Figures

Figures reproduced from arXiv: 2607.15004 by Binbo Li, Da Zhang, Hanxuan Chen, Hanzhong Guo, Ji Pei, Kangli Wang, Ruilong Ren, Shuai Yuan, Songsheng Cheng, Tianle Zeng, Xiangyue Wang, Yunpeng Zhou.

Figure 1
Figure 1. Figure 1: Task overview. The blue trajectory denotes the pedestrian’s motion, and the green trajectory denotes the UAV’s executed flight path. At each step, the agent observes five egocentric frames, previous target boxes, and the task prompt. It predicts three synchronized outputs: target visibility, the current-frame bounding box, and an 8-step 4-DoF waypoint action chunk. When the target is visible, the policy ca… view at source ↗
Figure 2
Figure 2. Figure 2: CosFly-VLA architecture. Multi-frame observations, language instructions, state history, and boundary￾tagged meta-query tokens are encoded by a frozen Qwen3.5 backbone with LoRA adapters. The meta-query hidden states are routed to three structured prediction heads: a 12-layer flow-matching DiT action expert for the 8-step 4-DoF waypoint-delta action chunk, a MLP bounding-box head, and a visibility head. Th… view at source ↗
Figure 3
Figure 3. Figure 3: CosFly-VLA training recipe. (a) Spatially Grounded CPT adapts the Qwen3.5 backbone with aerial spatial-perception supervision. (b) Curriculum-based SFT is a three-stage supervised process: multi-head warm-up, natural-distribution curriculum learning, and hard / long-occlusion curriculum learning. (c) CoT Training teaches the model to emit recovery-oriented reasoning traces before structured answers. (d) Cl… view at source ↗
Figure 4
Figure 4. Figure 4: Closed-loop reinforcement learning pipeline. Starting from the SFT checkpoint, CosFly-VLA is served as an action-head policy in the CARLA environment. The rollout collector gathers multiple on-policy trajectories and one expert anchor for each start condition, computes rewards from stand-off distance, target IoU, success, and collision terms, and transfers the trajectories to the trainer. The trainer compu… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative CoT reasoning under long occlusion. The seen-test and unseen-test examples each show four historical frames and the current frame, followed by a structured reasoning trace that identifies the last visible target state, the primary occluder, relative motion, the estimated current target region, and a recovery-oriented flight plan. and the same five-frame ground-truth history buffer [PITH_FULL_I… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative closed-loop comparison on representative CARLA episodes. Rows show Examples A–D; columns compare π0.5, CosFly-VLA-0.8B (SFT), and CosFly-VLA-0.8B (RL). The thick bright-yellow curve denotes the ground-truth UAV path, and the thick bright-magenta dashed curve denotes pedestrian motion. Executed UAV trajectories are shown as thick solid curves, using vivid orange for π0.5, bright cyan-green for C… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

77 extracted references · 14 linked inside Pith

  1. [1]

    Claude Opus 4.6, February 2026

    Anthropic. Claude Opus 4.6, February 2026

    2026

  2. [2]

    Learning vision-based pursuit-evasion robot policies

    Andrea Bajcsy, Antonio Loquercio, Ashish Kumar, and Jitendra Malik. Learning vision-based pursuit-evasion robot policies. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 9197–9204. IEEE, 2024

    2024

  3. [3]

    Capera: Captioning events in aerial videos.Remote Sensing, 15(8):2139, 2023

    Laila Bashmal, Yakoub Bazi, Mohamad Mahmoud Al Rahhal, Mansour Zuair, and Farid Melgani. Capera: Captioning events in aerial videos.Remote Sensing, 15(8):2139, 2023

    2023

  4. [4]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky.π 0: A visi...

    2024

  5. [5]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov...

    2023

  6. [6]

    Rt-1: Robotics transformer for real-world control at scale, 2023

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang- Huei Lee, Sergey Levine, Yao Lu, Utsav Malla,...

    2023

  7. [7]

    Gr-3 technical report, 2025

    Chilam Cheang, Sijin Chen, Zhongren Cui, Yingdong Hu, Liqun Huang, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Xiao Ma, Hao Niu, Wenxuan Ou, Wanli Peng, Zeyu Ren, Haixin Shi, Jiawen Tian, Hongtao Wu, Xin Xiao, Yuyang Xiao, Jiafeng Xu, and Yichu Yang. Gr-3 technical report, 2025

    2025

  8. [8]

    Track a*: Fast visibility-aware trajectory planning for active target tracking.arXiv preprint arXiv:2605.05338, 2026

    Hanxuan Chen, Kangli Wang, and Ji Pei. Track a*: Fast visibility-aware trajectory planning for active target tracking.arXiv preprint arXiv:2605.05338, 2026

    Pith/arXiv arXiv 2026

  9. [9]

    Cosfly: Plan in the matrix, fly in the world.arXiv preprint arXiv:2605.19120, 2026

    Hanxuan Chen, Xiangyue Wang, Songsheng Cheng, Ruilong Ren, Jie Zheng, Shuai Yuan, Tianle Zeng, Hanzhong Guo, Binbo Li, Kangli Wang, et al. Cosfly: Plan in the matrix, fly in the world.arXiv preprint arXiv:2605.19120, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    Vision-and-language navigation for uavs: Progress, challenges, and a research roadmap.arXiv preprint arXiv:2604.13654, 2026

    Hanxuan Chen, Jie Zheng, Siqi Yang, Tianle Zeng, Siwei Feng, Songsheng Cheng, Ruilong Ren, Hanzhong Guo, Shuai Yuan, Xiangyue Wang, et al. Vision-and-language navigation for uavs: Progress, challenges, and a research roadmap.arXiv preprint arXiv:2604.13654, 2026

    Pith/arXiv arXiv 2026

  11. [11]

    Diffusion policy: Visuomotor policy learning via action diffusion

    Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. InProceedings of Robotics: Science and Systems (RSS), 2023

    2023

  12. [12]

    D-V AT: End-to-end visual active tracking for micro aerial vehicles.IEEE Robotics and Automation Letters, 9(6):5046–5053, 2024

    Alberto Dionigi, Simone Felicioni, Mirko Leomanni, and Gabriele Costante. D-V AT: End-to-end visual active tracking for micro aerial vehicles.IEEE Robotics and Automation Letters, 9(6):5046–5053, 2024

    2024

  13. [13]

    Open3d-vqa: A benchmark for embodied spatial concept reasoning with multimodal large language model in open space

    Weichao Du, Tianyu Zhang, Lixiang Wang, Chen Gao, Yong Liu, Xinlei Chen, and Yueting Pang. Open3d-vqa: A benchmark for embodied spatial concept reasoning with multimodal large language model in open space. In Proceedings of the 33rd ACM International Conference on Multimedia (ACM MM), 2025

    2025

  14. [14]

    Gemini 3.1 Pro Model Card, February 2026

    Google DeepMind. Gemini 3.1 Pro Model Card, February 2026

    2026

  15. [15]

    AwareVLN: Reasoning with self-awareness for vision-language navigation.arXiv preprint arXiv:2605.22816, 2026

    Wenxuan Guo, Xiuwei Xu, Yichen Liu, Xiangyu Li, Hang Yin, Huangxing Chen, Wenzhao Zheng, Jianjiang Feng, Jie Zhou, and Jiwen Lu. AwareVLN: Reasoning with self-awareness for vision-language navigation.arXiv preprint arXiv:2605.22816, 2026. 17 CosFly-VLA: A Spatially Aware Vision-Language-Action Model for UA V Tracking

    Pith/arXiv arXiv 2026

  16. [16]

    A dual process vla: Efficient robotic manipulation leveraging vlm, 2024

    ByungOk Han, Jaehong Kim, and Jinhyeok Jang. A dual process vla: Efficient robotic manipulation leveraging vlm, 2024

    2024

  17. [17]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Represen- tations (ICLR), 2022

    2022

  18. [18]

    Navdreamer: Video models as zero-shot 3d navigators, 2026

    Xijie Huang, Weiqi Gai, Tianyue Wu, Congyu Wang, Zhiyang Liu, Xin Zhou, Yuze Wu, and Fei Gao. Navdreamer: Video models as zero-shot 3d navigators, 2026

    2026

  19. [19]

    Physical Intelligence, Ali Amin, Raichelle Aniceto, Ashwin Balakrishna, Kevin Black, Ken Conley, Grace Connors, James Darpinian, Karan Dhabalia, Jared DiCarlo, Danny Driess, Michael Equi, Adnan Esmail, Yunhao Fang, Chelsea Finn, Catherine Glossop, Thomas Godden, Ivan Goryachev, Lachy Groom, Hunter Hancock, Karol Hausman, Gashon Hussein, Brian Ichter, Szym...

    2025

  20. [20]

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y . Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsc...

    2025

  21. [21]

    Herath M. P. C. Jayaweera and Samer Hanoun. A dynamic artificial potential field (D-APF) UA V path planning technique for following ground moving targets.IEEE Access, 8:192760–192776, 2020

    2020

  22. [22]

    A survey on vision-language-action models for autonomous driving

    Sicong Jiang, Zilin Huang, Kangan Qian, Ziang Luo, Tianze Zhu, Yang Zhong, Yihong Tang, Menglin Kong, Yunlong Wang, Siwen Jiao, et al. A survey on vision-language-action models for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 4583– 4595, 2025

    2025

  23. [23]

    A survey on detection, classification, and tracking of UA Vs using radar and communications systems.IEEE Communications Surveys & Tutorials, 28:3272–3310, 2025

    Wahab Khawaja, Martins Ezuma, Vasilii Semkin, Fatih Erden, Ozgur Ozdemir, and Ismail Guvenc. A survey on detection, classification, and tracking of UA Vs using radar and communications systems.IEEE Communications Surveys & Tutorials, 28:3272–3310, 2025

    2025

  24. [24]

    Openvla: An open-source vision-language-action model, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model, 2024

    2024

  25. [25]

    IB-AMG: Aircraft mission generation with inference-based vision-language-action model

    JinHai Li, Peng Chen, MoHan Li, and LuYi Ren. IB-AMG: Aircraft mission generation with inference-based vision-language-action model. In2025 40th Youth Academic Annual Conference of Chinese Association of Automation (YAC), pages 2323–2328, 2025

    2025

  26. [26]

    Hrvqa: A visual question answering benchmark for high-resolution aerial images.ISPRS Journal of Photogrammetry and Remote Sensing, 214:65–81, 2024

    Kun Li, George V osselman, and Michael Ying Yang. Hrvqa: A visual question answering benchmark for high-resolution aerial images.ISPRS Journal of Photogrammetry and Remote Sensing, 214:65–81, 2024

    2024

  27. [27]

    Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation, 2024

    Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu, Yizhong Zhang, Xiaofan Wang, Bei Liu, Jianlong Fu, Jianmin Bao, Dong Chen, Yuanchun Shi, Jiaolong Yang, and Baining Guo. Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation, 2024

    2024

  28. [28]

    Adaptive and background-aware vision transformer for real-time UA V tracking

    Shuiwang Li, Yangxiang Yang, Dan Zeng, and Xucheng Wang. Adaptive and background-aware vision transformer for real-time UA V tracking. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13989–14000, 2023

    2023

  29. [29]

    AutoTrack: Towards high-performance visual tracking for UA V with automatic spatio-temporal regularization

    Yiming Li, Changhong Fu, Fangqiang Ding, Ziyuan Huang, and Geng Lu. AutoTrack: Towards high-performance visual tracking for UA V with automatic spatio-temporal regularization. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  30. [30]

    Gr-rl: Going dexterous and precise for long-horizon robotic manipulation, 2025

    Yunfei Li, Xiao Ma, Jiafeng Xu, Yu Cui, Zhongren Cui, Zhigang Han, Liqun Huang, Tao Kong, Yuxiao Liu, Hao Niu, Wanli Peng, Jingchao Qiao, Zeyu Ren, Haixin Shi, Zhi Su, Jiawen Tian, Yuyang Xiao, Shenyu Zhang, Liwei Zheng, Hang Li, and Yonghui Wu. Gr-rl: Going dexterous and precise for long-horizon robotic manipulation, 2025. 18 CosFly-VLA: A Spatially Awar...

    2025

  31. [31]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  32. [32]

    Trackvla++: Unleashing reasoning and memory capabilities in vla models for embodied visual tracking, 2025

    Jiahang Liu, Yunpeng Qi, Jiazhao Zhang, Minghan Li, Shaoan Wang, Kui Wu, Hanjing Ye, Hong Zhang, Zhibo Chen, Fangwei Zhong, Zhizheng Zhang, and He Wang. Trackvla++: Unleashing reasoning and memory capabilities in vla models for embodied visual tracking, 2025

    2025

  33. [33]

    Indooruav: Benchmarking vision-language uav navigation in continuous indoor environments, 2025

    Xu Liu, Yu Liu, Hanshuo Qiu, Yang Qirong, and Zhouhui Lian. Indooruav: Benchmarking vision-language uav navigation in continuous indoor environments, 2025

    2025

  34. [34]

    End-to-end active object tracking and its real-world deployment via reinforcement learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(6):1317–1332, 2020

    Wenhan Luo, Peng Sun, Fangwei Zhong, Wei Liu, Tong Zhang, and Yizhou Wang. End-to-end active object tracking and its real-world deployment via reinforcement learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(6):1317–1332, 2020

    2020

  35. [35]

    Cognitivedrone: A vla model and evaluation benchmark for real-time cognitive task solving and reasoning in uavs, 2025

    Artem Lykov, Valerii Serpiva, Muhammad Haris Khan, Oleg Sautenkov, Artyom Myshlyaev, Grik Tadevosyan, Yasheerah Yaqoot, and Dzmitry Tsetserukou. Cognitivedrone: A vla model and evaluation benchmark for real-time cognitive task solving and reasoning in uavs, 2025

    2025

  36. [36]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195–4205, 2023

    2023

  37. [37]

    Qwen3-VL Technical Report.arXiv preprint arXiv:2511.21631, 2025

    Qwen Team. Qwen3-VL Technical Report.arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Qwen3.5: Towards native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026

    2026

  39. [39]

    UA V-VLA: Vision-language-action system for large scale aerial mission generation

    Oleg Sautenkov, Yasheerah Yaqoot, Artem Lykov, Muhammad Ahsan Mustafa, Grik Tadevosyan, Aibek Akhmetkazy, Miguel Altamirano Cabrera, Mikhail Martynov, Sausar Karaf, and Dzmitry Tsetserukou. UA V-VLA: Vision-language-action system for large scale aerial mission generation. In2025 20th ACM/IEEE International Conference on Human-Robot Interaction (HRI), page...

    2025

  40. [40]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  41. [41]

    Racevla: Vla-based racing drone navigation with human-like behaviour, 2025

    Valerii Serpiva, Artem Lykov, Artyom Myshlyaev, Muhammad Haris Khan, Ali Alridha Abdulkarim, Oleg Sautenkov, and Dzmitry Tsetserukou. Racevla: Vla-based racing drone navigation with human-like behaviour, 2025

    2025

  42. [42]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  43. [43]

    Hume: Introducing system-2 thinking in visual-language-action model.arXiv preprint arXiv:2505.21432, 2025

    Haoming Song, Delin Qu, Yuanqi Yao, Qizhi Chen, Qi Lv, Yiwen Tang, Modi Shi, Guanghui Ren, Maoqing Yao, Bin Zhao, et al. Hume: Introducing system-2 thinking in visual-language-action model.arXiv preprint arXiv:2505.21432, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    UA V-ground visual tracking: A unified dataset and collaborative learning approach.IEEE Transactions on Circuits and Systems for Video Technology, 34(5):3619–3632, 2024

    Dengdi Sun, Leilei Cheng, Song Chen, Chenglong Li, Yun Xiao, and Bin Luo. UA V-ground visual tracking: A unified dataset and collaborative learning approach.IEEE Transactions on Circuits and Systems for Video Technology, 34(5):3619–3632, 2024

    2024

  45. [45]

    Open-world drone active tracking with goal-centered rewards

    Haowei Sun, Jinwu Hu, Zhirui Zhang, Haoyuan Tian, Xinze Xie, Yufeng Wang, Xiaohua Xie, Yun Lin, Zhuliang Yu, and Mingkui Tan. Open-world drone active tracking with goal-centered rewards. InAdvances in Neural Information Processing Systems (NeurIPS), 2025

    2025

  46. [46]

    Air-vla: Vision-language-action systems for aerial manipulation, 2026

    Jianli Sun, Bin Tian, Qiyao Zhang, Chengxiang Li, Zihan Song, Zhiyong Cui, Yisheng Lv, and Yonglin Tian. Air-vla: Vision-language-action systems for aerial manipulation, 2026

    2026

  47. [47]

    Moving target tracking by unmanned aerial vehicle: A survey and taxonomy.IEEE Transactions on Industrial Informatics, 20(5):7056–7068, 2024

    Nianyi Sun, Jin Zhao, Qing Shi, Chang Liu, and Peng Liu. Moving target tracking by unmanned aerial vehicle: A survey and taxonomy.IEEE Transactions on Industrial Informatics, 20(5):7056–7068, 2024

    2024

  48. [48]

    Autofly: Vision-language-action model for uav autonomous navigation in the wild, 2026

    Xiaolou Sun, Wufei Si, Wenhui Ni, Yuntian Li, Dongming Wu, Fei Xie, Runwei Guan, He-Yang Xu, Henghui Ding, Yuan Wu, Yutao Yue, Yongming Huang, and Hui Xiong. Autofly: Vision-language-action model for uav autonomous navigation in the wild, 2026

    2026

  49. [49]

    Refdrone: A challenging benchmark for referring expression comprehension in drone scenes

    Zhichao Sun, Yepeng Liu, Zhiling Su, Huachao Zhu, Yuliang Gu, Yuda Zou, Zelong Liu, Gui-Song Xia, Bo Du, and Yongchao Xu. Refdrone: A challenging benchmark for referring expression comprehension in drone scenes. arXiv preprint arXiv:2502.00392, 2025

    arXiv 2025

  50. [50]

    Alpamayo-r1: Bridging reasoning and action prediction for generalizable autonomous driving in the long tail.arXiv preprint arXiv:2511.00088, 2025

    NVIDIA Alpamayo-R1 Team and Marco Pavone. Alpamayo-r1: Bridging reasoning and action prediction for generalizable autonomous driving in the long tail.arXiv preprint arXiv:2511.00088, 2025. 19 CosFly-VLA: A Spatially Aware Vision-Language-Action Model for UA V Tracking

    Pith/arXiv arXiv 2025

  51. [51]

    Octo: An open-source generalist robot policy, 2024

    Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, You Liang Tan, Lawrence Yunliang Chen, Pannag Sanketi, Quan Vuong, Ted Xiao, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo: An open-source generalist robot policy, 2024

    2024

  52. [52]

    UA Vs meet LLMs: Overviews and perspectives toward agentic low-altitude mobility

    Yonglin Tian, Fei Lin, Yiduo Li, Tengchao Zhang, Qiyao Zhang, Xuan Fu, Jun Huang, Xingyuan Dai, Yutong Wang, Chunwei Tian, et al. UA Vs meet LLMs: Overviews and perspectives toward agentic low-altitude mobility. Information Fusion, 122:103158, 2025

    2025

  53. [53]

    Trackvla: Embodied visual tracking in the wild

    Shaoan Wang, Jiazhao Zhang, Minghan Li, Jiahang Liu, Anqi Li, Kui Wu, Fangwei Zhong, Junzhi Yu, Zhizheng Zhang, and He Wang. Trackvla: Embodied visual tracking in the wild. In Joseph Lim, Shuran Song, and Hae-Won Park, editors,Proceedings of The 9th Conference on Robot Learning, volume 305 ofProceedings of Machine Learning Research, pages 4139–4164. PMLR,...

    2025

  54. [54]

    Uav-flow colosseo: A real-world benchmark for flying-on-a-word uav imitation learning, 2025

    Xiangyu Wang, Donglin Yang, Yue Liao, Wenhao Zheng, wenjun wu, Bin Dai, Hongsheng Li, and Si Liu. Uav-flow colosseo: A real-world benchmark for flying-on-a-word uav imitation learning, 2025

    2025

  55. [55]

    Cosfly-track: A large-scale multi-modal dataset for uav visual tracking via multi-constraint trajectory optimization.arXiv preprint arXiv:2605.17776, 2026

    Xiangyue Wang, Hanxuan Chen, Songsheng Cheng, Ruilong Ren, Jie Zheng, Shuai Yuan, Tianle Zeng, Hanzhong Guo, Kangli Wang, and Ji Pei. Cosfly-track: A large-scale multi-modal dataset for uav visual tracking via multi-constraint trajectory optimization.arXiv preprint arXiv:2605.17776, 2026

    Pith/arXiv arXiv 2026

  56. [56]

    Hierarchical instruction-aware embodied visual tracking, 2025

    Kui Wu, Hao Chen, Churan Wang, Fakhri Karray, Zhoujun Li, Yizhou Wang, and Fangwei Zhong. Hierarchical instruction-aware embodied visual tracking, 2025

    2025

  57. [57]

    VLM can be a good assistant: Enhancing embodied visual tracking with self-improving vision-language models

    Kui Wu, Shuhang Xu, Hao Chen, Churan Wang, Zhoujun Li, Yizhou Wang, and Fangwei Zhong. VLM can be a good assistant: Enhancing embodied visual tracking with self-improving vision-language models. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13154–13161, 2025

    2025

  58. [58]

    Cognitive embodied learning for anomaly active target tracking.Communications Engineering, 4(1):224, Nov 2025

    Qihui Wu, Jiahao Li, Fuhui Zhou, Jiahuan Ji, Haoyang Wang, Hongtao Liang, and Kai-Kuang Ma. Cognitive embodied learning for anomaly active target tracking.Communications Engineering, 4(1):224, Nov 2025

    2025

  59. [59]

    Learning an adaptive and view-invariant vision transformer for real-time UA V tracking.IEEE Transactions on Circuits and Systems for Video Technology, 36(2):2403–2418, 2026

    You Wu, Yongxin Li, Mengyuan Liu, Xucheng Wang, Xiangyang Yang, Hengzhou Ye, Dan Zeng, Qijun Zhao, and Shuiwang Li. Learning an adaptive and view-invariant vision transformer for real-time UA V tracking.IEEE Transactions on Circuits and Systems for Video Technology, 36(2):2403–2418, 2026

    2026

  60. [60]

    Vla-an: An efficient and onboard vision-language-action framework for aerial navigation in complex environments, 2025

    Yuze Wu, Mo Zhu, Xingxing Li, Yuheng Du, Yuxin Fan, Wenjun Li, Zhichao Han, Xin Zhou, and Fei Gao. Vla-an: An efficient and onboard vision-language-action framework for aerial navigation in complex environments, 2025

    2025

  61. [61]

    Anti-distractor active object tracking in 3D environments.IEEE Transactions on Circuits and Systems for Video Technology, 32(6):3697–3707, 2022

    Mao Xi, Yun Zhou, Zheng Chen, Wengang Zhou, and Houqiang Li. Anti-distractor active object tracking in 3D environments.IEEE Transactions on Circuits and Systems for Video Technology, 32(6):3697–3707, 2022

    2022

  62. [62]

    AerialVLA: A vision-language-action model for uav navigation via minimalist end-to-end control.arXiv preprint arXiv:2603.14363, 2026

    Peng Xu, Zhengnan Deng, Jiayan Deng, Zonghua Gu, and Shaohua Wan. AerialVLA: A vision-language-action model for uav navigation via minimalist end-to-end control.arXiv preprint arXiv:2603.14363, 2026

    arXiv 2026

  63. [63]

    Similarity-guided layer-adaptive vision transformer for UA V tracking

    Chaocan Xue, Bineng Zhong, Qihua Liang, Yaozong Zheng, Ning Li, Yuanliang Xue, and Shuxiang Song. Similarity-guided layer-adaptive vision transformer for UA V tracking. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6730–6740, 2025

    2025

  64. [64]

    Handling occlusion in UA V visual tracking with query-guided redetection.IEEE Transactions on Instrumentation and Measurement, 73:1–17, 2024

    Yuanliang Xue, Tao Shen, Guodong Jin, Lining Tan, Nian Wang, Lianfeng Wang, and Jing Gao. Handling occlusion in UA V visual tracking with query-guided redetection.IEEE Transactions on Instrumentation and Measurement, 73:1–17, 2024

    2024

  65. [65]

    Ezreal: Enhancing zero-shot outdoor robot navigation toward distant targets under varying visibility.arXiv preprint arXiv:2509.13720, 2025

    Tianle Zeng, Jianwei Peng, Hanjing Ye, Guangcheng Chen, Senzi Luo, and Hong Zhang. Ezreal: Enhancing zero-shot outdoor robot navigation toward distant targets under varying visibility.arXiv preprint arXiv:2509.13720, 2025

    arXiv 2025

  66. [66]

    Can aerial vla models cooperate? evaluating closed-loop air-ground coordination with carla-air.arXiv preprint arXiv:2605.31066, 2026

    Tianle Zeng, Yanci Wen, Xueang Yu, and Hong Zhang. Can aerial vla models cooperate? evaluating closed-loop air-ground coordination with carla-air.arXiv preprint arXiv:2605.31066, 2026

    Pith/arXiv arXiv 2026

  67. [67]

    Carla-air: Fly drones inside a carla world–a unified infrastructure for air-ground embodied intelligence.arXiv preprint arXiv:2603.28032, 2026

    Tianle Zeng, Yanci Wen, and Hong Zhang. Carla-air: Fly drones inside a carla world–a unified infrastructure for air-ground embodied intelligence.arXiv preprint arXiv:2603.28032, 2026

    Pith/arXiv arXiv 2026

  68. [68]

    Aircopbench: A benchmark for multi-drone collaborative embodied perception and reasoning.arXiv preprint arXiv:2511.11025, 2025

    Jirong Zha, Yuxuan Fan, Tianyu Zhang, Geng Chen, Yingfeng Chen, Chen Gao, and Xinlei Chen. Aircopbench: A benchmark for multi-drone collaborative embodied perception and reasoning.arXiv preprint arXiv:2511.11025, 2025

    arXiv 2025

  69. [69]

    Igniting vlms toward the embodied space, 2025

    Andy Zhai, Brae Liu, Bruno Fang, Chalse Cai, Ellie Ma, Ethan Yin, Hao Wang, Hugo Zhou, James Wang, Lights Shi, Lucy Liang, Make Wang, Qian Wang, Roy Gan, Ryan Yu, Shalfun Li, Starrick Liu, Sylas Chen, Vincent Chen, and Zach Xu. Igniting vlms toward the embodied space, 2025. 20 CosFly-VLA: A Spatially Aware Vision-Language-Action Model for UA V Tracking

    2025

  70. [70]

    Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks, 2025

    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, 2025

    2025

  71. [71]

    UA V-Track VLA: Embodied aerial tracking via vision-language-action models.arXiv preprint arXiv:2604.02241, 2026

    Qiyao Zhang, Shuhua Zheng, Jianli Sun, Chengxiang Li, Xianke Wu, Zihan Song, Zhiyong Cui, Yisheng Lv, and Yonglin Tian. UA V-Track VLA: Embodied aerial tracking via vision-language-action models.arXiv preprint arXiv:2604.02241, 2026

    Pith/arXiv arXiv 2026

  72. [72]

    WorldVLN: Autoregressive world action model for aerial vision-language navigation.arXiv preprint arXiv:2605.15964, 2026

    Baining Zhao, Jiacheng Xu, Weicheng Feng, Xin Zhang, Zhaolu Wang, Haoyang Wang, Shilong Ji, Ziyou Wang, Jianjie Fang, Zhiheng Zheng, Weichen Zhang, Yu Shang, Wei Wu, Chen Gao, Xinlei Chen, and Yong Li. WorldVLN: Autoregressive world action model for aerial vision-language navigation.arXiv preprint arXiv:2605.15964, 2026

    Pith/arXiv arXiv 2026

  73. [73]

    Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn

    Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware, 2023

    2023

  74. [74]

    RSPT: Reconstruct surroundings and predict trajectory for generalizable active object tracking

    Fangwei Zhong, Xiao Bi, Yudi Zhang, Wei Zhang, and Yizhou Wang. RSPT: Reconstruct surroundings and predict trajectory for generalizable active object tracking. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 37, pages 3705–3714, 2023

    2023

  75. [75]

    UnrealZoo: Enriching photo-realistic virtual worlds for embodied ai

    Fangwei Zhong, Kui Wu, Churan Wang, Hao Chen, Hai Ci, Zhoujun Li, and Yizhou Wang. UnrealZoo: Enriching photo-realistic virtual worlds for embodied ai. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5769–5779, 2025

    2025

  76. [76]

    Airspatialbot: A spatially-aware aerial agent for fine-grained vehicle attribute recognition and retrieval.IEEE Transactions on Geoscience and Remote Sensing, 63:1–15, 2025

    Yue Zhou, Ran Ding, Xue Yang, Jiang Xue, and Xingzhao Liu. Airspatialbot: A spatially-aware aerial agent for fine-grained vehicle attribute recognition and retrieval.IEEE Transactions on Geoscience and Remote Sensing, 63:1–15, 2025

    2025

  77. [77]

    Multimodal mathematical reasoning embedded in aerial vehicle imagery: Benchmarking, analysis, and exploration

    Yue Zhou, Litong Feng, Mengcheng Lan, Xue Yang, Qingyun Li, Yiping Ke, Jiang Xue, and Wayne Zhang. Multimodal mathematical reasoning embedded in aerial vehicle imagery: Benchmarking, analysis, and exploration. ISPRS Journal of Photogrammetry and Remote Sensing, 2025. 21

    2025