Pith. sign in

REVIEW 3 major objections 52 references

ActiveFly-Bench is the first UAV benchmark that links high-level scene questions to fine-grained body-and-gimbal control for active aerial perception, and current agents still fail mainly at the planning and viewpoint steps.

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 · grok-4.5

2026-07-14 13:41 UTC pith:ZXXXZJT5

load-bearing objection Solid hierarchical UAV benchmark that actually connects EQA to 7-DoF control; useful data and real deployment, soft on stats and gold-standard sensitivity. the 3 major comments →

arxiv 2607.10180 v1 pith:ZXXXZJT5 submitted 2026-07-11 cs.RO cs.AI

ActiveFly-Bench: Aligning Embodied Question Answering with Vision-Language-Action for Aerial Embodied Perception

Weichen Zhang , Shiquan Yu , Yinan Zhu , Peizhi Tang , Shilong Ji , Zhiyuan Deng , Tianyi Lyu , Haoyang Wang
show 4 more authors
Xin Zeng Chen Gao Yong Li Xinlei Chen
This is my paper
classification cs.RO cs.AI
keywords UAVembodied question answeringvision-language-actionactive perceptionobservation behavior planningfine-grained controlaerial navigationgimbal control
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.

Language-guided aerial perception requires a drone not only to understand a question but to move and re-aim its camera so that the answer becomes visible. Existing benchmarks either stay in simulation with coarse navigation or omit the intermediate plan that turns a question into an observation strategy. ActiveFly-Bench closes that gap by decomposing the problem into three aligned tasks built from the same trajectories: Air-EQA (answer a question that cannot be solved from the start view), Observation Behavior Planning (state how the drone should move and aim), and FLUC (execute the corresponding fine-grained 7-DoF actions). Roughly 10k trajectories from real outdoor, real indoor, and simulated scenes, plus 1.3k question pairs, support training and evaluation. A closed-loop agent is also flown on a physical platform. Experiments with strong vision-language and vision-language-action models show that end-to-end success remains low even when final answers are sometimes correct, because planning and precise viewpoint control break first.

Core claim

The paper shows that a hierarchical, semantically aligned split into Air-EQA, Observation Behavior Planning, and fine-grained language-guided UAV control (including gimbal pitch) is both necessary and sufficient to evaluate whether a UAV agent can turn an open-vocabulary question into an informative viewpoint and then answer it. On this testbed, representative commercial VLMs paired with open VLA controllers achieve high question accuracy but substantially lower joint success, because agents routinely fail at behavior planning or miss the required final pose even when they pass near the target.

What carries the argument

The three-task hierarchy Air-EQA, OBP, and FLUC, all derived from the same human-collected and augmented trajectories so that question, observation plan, and 7-DoF action stay aligned. End-to-end embodied-perception success is defined as the product of correct OBP, oracle FLUC success, and correct Air-EQA.

Load-bearing premise

The claim depends on treating short human pilot trajectories, lightly noise-augmented, plus fixed position and orientation tolerances as a reliable gold standard for what counts as an optimal observation viewpoint.

What would settle it

An agent that reliably achieves high joint success (correct plan, oracle viewpoint success under the stated 3 m / 10° thresholds, and correct answer) on held-out real outdoor and indoor splits, while human pilots still judge the resulting viewpoints natural and informative; if no such agent appears, or if many labeled successes still leave the target poorly framed under modestly tighter orientation checks, the benchmark’s diagnosis of current bottlenecks would be undermined.

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

If this is right

  • UAV agents can be scored separately on planning, control, and answering, isolating which module fails.
  • Training data now exist for joint body-and-gimbal control conditioned on short observation plans rather than long navigation scripts.
  • Real-world closed-loop flight with ground-station inference becomes a standard evaluation requirement, not an optional demo.
  • The joint success metric makes “escape” cases (right answer despite wrong plan or missed target) measurable and penalizable.
  • Sim-to-real transfer for aerial vision-language-action models can be tested on matched indoor and outdoor trajectory categories.

Where Pith is reading between the lines

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

  • The same three-stage split could turn existing indoor EQA datasets into planning-plus-control benchmarks for ground robots with pan-tilt cameras.
  • If Observation Behavior Planning stays the dominant error source, a lightweight specialized planner may improve sample efficiency more than simply scaling the general vision-language model.
  • The multi-model blind filter used to discard questions answerable from the start frame could serve as a reusable quality gate for any active-perception dataset.
  • As control precision rises, fixed viewpoint tolerances may need adaptive tightening; otherwise success rates will saturate while true viewpoint quality remains limited.

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

3 major / 0 minor

Summary. The paper introduces ActiveFly-Bench, a hierarchical benchmark for language-guided UAV embodied perception that links high-level Aerial Embodied Question Answering (Air-EQA) to intermediate Observation Behavior Planning (OBP) and low-level Fine-grained Language-guided UAV Control (FLUC, 7-DoF body+gimbal). It releases ~10k multi-source trajectories (sim + real indoor/outdoor) and ~1.3k aligned QA/OBP pairs, defines an end-to-end EP success metric Sep = Sobp · OSfluc · Seqa, and evaluates modular VLM+VLA agents (GPT-5.4/Gemini/Qwen + OpenVLA/π0.5) plus human upper bounds. Results show low EP success, high Air-EQA “escape,” and planning/viewpoint bottlenecks; a closed-loop ActiveFly agent is also deployed on a physical UAV with reported latency.

Significance. If the hierarchical construction and reported gaps hold, the work supplies a useful, previously missing testbed that forces joint evaluation of cyberspace reasoning, observation planning, and viewpoint-aware control for aerial agents—beyond pure VLN or indoor EQA. Strengths include multi-source trajectory collection with human pilots, multi-VLM blind filtering for Air-EQA, explicit EP composition, escape analysis (§6.4), real-world closed-loop deployment with latency breakdown (Table 3), and public data/code. These make the qualitative claim that current VLM+VLA stacks struggle on planning and precise viewpoint adjustment credible and actionable for the community.

major comments (3)
  1. §6.1 / Appendix A.5 and Table 2: Success thresholds δ_loc=3 m and δ_ori=10° (and the OSR definition that only requires any intermediate pose to meet them) are load-bearing for FLUC SR/OSR and thus for EP. The manuscript does not justify these values against typical target sizes, camera FOV, or pilot variance, nor report sensitivity. Without that analysis (or error bars over seeds/splits), the absolute SR/OSR numbers and the claimed “viewpoint adjustment” bottleneck are hard to interpret as robust.
  2. §4.1–4.3 and §6.2–6.4: Gold-standard observation behaviors and answers rest on short human pilot trajectories plus Gaussian waypoint perturbation and multi-VLM blind filtering. Residual information leakage or pilot idiosyncrasy is acknowledged as a risk but not quantified (e.g., inter-annotator agreement, fraction of retained “edge” EQAs after CoT review, or human–human EP agreement beyond the single “Human Agent” row). Because EP multiplies three binary indicators, even moderate label noise can inflate escape rates and understate true agent capability; a small reliability study is needed to underwrite the central “agents still struggle” claim.
  3. Table 2 and §5: Several VLM+VLA cells for Gemini/Qwen + OpenVLA/π0.5 leave FLUC metrics blank (“-”), while EP is still reported for some combinations. It is unclear whether those agents were not run end-to-end, failed to produce valid actions, or were evaluated only on partial pipelines. Clarifying the evaluation protocol and filling or explicitly excluding those cells is required for the comparative claim that π0.5-based agents outperform OpenVLA-based ones on EP.

Circularity Check

0 steps flagged

No significant circularity: empirical benchmark whose metrics and tasks are defined against external human trajectories and answers, not quantities derived from the same fitted parameters.

full rationale

ActiveFly-Bench is a systems/benchmark paper that decomposes language-guided UAV perception into Air-EQA, OBP and FLUC, constructs aligned datasets from human-piloted trajectories (real + sim), and evaluates off-the-shelf VLMs/VLAs under standard success metrics (SR/OSR/NE/nDTW, MCQ accuracy, APL, and the product EP indicator). The hierarchical construction (§3–4) and the EP definition Sep = Sobp · OSfluc · Seqa are definitional bookkeeping, not a derivation that reduces a claimed prediction to its own inputs. Ground-truth answers, observation-behavior descriptions and 7-DoF trajectories are human-annotated (with multi-VLM blind filtering for leakage); success thresholds (3 m / 10°) are fixed external criteria. Self-citations (EmbodiedCity, UAV-Flow, etc.) appear only as data sources or related systems and are not load-bearing uniqueness theorems or ansatzes that force the reported results. There is therefore no self-definitional loop, no fitted-parameter-as-prediction, and no circular self-citation chain. Score 0 is the honest finding.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 2 invented entities

As an empirical systems/benchmark paper the load-bearing content is definitional and experimental rather than axiomatic. Free parameters are mainly evaluation thresholds and system hyper-parameters; axioms are standard robotics and ML assumptions; invented entities are the three named tasks and the ActiveFly agent architecture.

free parameters (4)
  • position success threshold δ_loc = 3 m
    Set to 3 m for SR/OSR; directly controls reported success rates on FLUC and EP.
  • orientation success threshold δ_ori = 10°
    Set to 10° for yaw/pitch; jointly with δ_loc defines viewpoint success.
  • history frames n for final QA = 16
    Number of sampled images fed to the VLM for Air-EQA answer; fixed at 16.
  • Gaussian perturbation std for trajectory augmentation = 0.1 m / 0.05 rad
    0.1 m position / 0.05 rad yaw used to generate 5–40 demos per instruction; affects training distribution.
axioms (3)
  • domain assumption A short human-piloted trajectory that makes a previously unobservable target clearly visible constitutes a valid gold-standard observation behavior for the corresponding Air-EQA question.
    Underpins both FLUC labels and OBP answers (§4.1–4.3).
  • domain assumption Multi-VLM blind filtering (all of GPT/Gemini/Qwen answering correctly from the start frame) plus human review sufficiently removes information leakage and ambiguous questions.
    Used to validate Air-EQA and OBP pairs (Appendix A.3).
  • ad hoc to paper Modular VLM (planning/answering) + VLA (control) with stop-and-infer closed loop is a representative architecture for evaluating current UAV agents.
    Defines the ActiveFly agent and all reported baselines (§5).
invented entities (2)
  • Air-EQA / OBP / FLUC task hierarchy no independent evidence
    purpose: Decompose language-guided aerial active perception into aligned high-level QA, intermediate behavior planning, and low-level 7-DoF control.
    Core contribution of the benchmark; no independent prior definition of this exact three-way split for UAVs.
  • ActiveFly closed-loop agent no independent evidence
    purpose: Concrete VLM+VLA system that can be deployed on a physical UAV to execute the three tasks.
    Introduced to demonstrate real-world feasibility; architecture is a composition of existing model classes.

pith-pipeline@v1.1.0-grok45 · 26556 in / 2916 out tokens · 31260 ms · 2026-07-14T13:41:33.050644+00:00 · methodology

0 comments
read the original abstract

We introduce ActiveFly-Bench, the first benchmark to bridge cyberspace reasoning and physical-world interaction for UAV embodied perception. The benchmark decomposes active perception into three hierarchical tasks: Aerial Embodied Question Answering (Air-EQA), Observation Behavior Planning (OBP), and Fine-grained Language-guided UAV Control (FLUC), explicitly connecting high-level task understanding, behavior planning, and low-level control. The datasets are collected from both real-world and simulated outdoor environments for training and evaluation. We further develop ActiveFly, a closed-loop UAV agent that integrates visual-language reasoning with fine-grained control, and deploy it on a physical UAV platform. Experiments with representative VLMs and VLA models show that current UAV agents still struggle with behavior planning, viewpoint adjustment, and robust task completion in active perception. These results establish ActiveFly-Bench as a new testbed for embodied aerial intelligence.

Figures

Figures reproduced from arXiv: 2607.10180 by Chen Gao, Haoyang Wang, Peizhi Tang, Shilong Ji, Shiquan Yu, Tianyi Lyu, Weichen Zhang, Xinlei Chen, Xin Zeng, Yinan Zhu, Yong Li, Zhiyuan Deng.

Figure 1
Figure 1. Figure 1: Overview of ActiveFly-Bench. The benchmark targets language-guided embodied perception for UAVs. It decomposes the task into three hierarchical and semantically aligned subtasks: Aerial Embodied Question Answering, Observation Behavior Planning, and Fine-Grained Language-Guided UAV Control. Given a high-level task such as “What is under the tree?”, the UAV first performs reasoning and planning to infer the… view at source ↗
Figure 2
Figure 2. Figure 2: Examples of the three tasks in ActiveFly-Bench. Air-EQA spans five question categories: object recognition, attribute recognition, counting, spatial understanding, and world knowledge, with each example illustrated by the first and last frames. Observation Behavior Planning bridges Air-EQA and FLUC. FLUC includes 10 elementary and 4 complex active perception trajectories gimbal up/down, to support fine-gra… view at source ↗
Figure 3
Figure 3. Figure 3: Dataset Generation Pipeline and Statistics. (a) The data generation pipeline of three tasks. (b) The distribution of FLUC trajectory lengths. (c) Word-count distributions of language annotations. (d) Dataset composition statistics. (e) A word cloud generated from the language annotations. smooth trajectory, which is replayed by the UAV to automatically generate 5 to 40 demonstrations for each instruction. … view at source ↗
Figure 4
Figure 4. Figure 4: Category-level performance (left), error breakdown (middle) and case study of "escaped" case (right). The escaped case shows that although the agent deviates from the correct trajectory, it can still acquire the critical visual information (purple box) required to answer the embodied question. 6 Experiments We conduct a comprehensive evaluation of popular VLMs and VLA models on our benchmark. We adapt VLA … view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of real-world deployment of ActiveFly agent. The top and bottom rows show the first- and third-person views, respectively. Overall, Pi-0.5 is more likely to follow instructions and pass through the target region, which improves oracle success and benefits Air-EQA by exposing the agent to richer scene semantics. However, its longer trajectories reduce efficiency, leading to lower APL and large… view at source ↗
Figure 6
Figure 6. Figure 6: Example of Invalid EQA A.3.3 VLM-based Blind Screening of EQA The complete VLM-based Blind Screening pipeline utilized in this benchmark is outlined below: • Multi-Model Voting: The single-model judgment was replaced with a multi-model voting mechanism. Only when all models answer correctly on the start frame is the EQA flagged as “invalid”. This effectively minimizes the probability of models guessing the… view at source ↗
Figure 7
Figure 7. Figure 7: More Examples of EQA A.4 OBP Details Observation Behavior Planning (OBP) is presented in a QA format. The standard template for the question is "What action is needed to answer the question: <EQA question>¨ ¨ . The choices consist of the actual trajectory instruction as the correct answer, alongside three misleading instructions that are either physically impossible to execute or would result in a trajecto… view at source ↗
Figure 8
Figure 8. Figure 8: More Examples of OBP 21 [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] 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

52 extracted references · 16 linked inside Pith

  1. [1]

    Aligning cyber space with physical world: A comprehensive survey on embodied ai

    Yang Liu, Weixing Chen, Yongjie Bai, Xiaodan Liang, Guanbin Li, Wen Gao, and Liang Lin. Aligning cyber space with physical world: A comprehensive survey on embodied ai. IEEE/ASME Transactions on Mechatronics, 2025

    2025

  2. [2]

    Ddl: Empowering delivery drones with large-scale urban sensing capability.IEEE Journal of Selected Topics in Signal Processing, 18(3):502–515, 2024

    Xuecheng Chen, Haoyang Wang, Yuhan Cheng, Haohao Fu, Yuxuan Liu, Fan Dang, Yunhao Liu, Jinqiang Cui, and Xinlei Chen. Ddl: Empowering delivery drones with large-scale urban sensing capability.IEEE Journal of Selected Topics in Signal Processing, 18(3):502–515, 2024

    2024

  3. [3]

    Fuel: Fast uav exploration using incremental frontier structure and hierarchical planning.IEEE Robotics and Automation Letters, 6(2):779–786, 2021

    Boyu Zhou, Yichen Zhang, Xinyi Chen, and Shaojie Shen. Fuel: Fast uav exploration using incremental frontier structure and hierarchical planning.IEEE Robotics and Automation Letters, 6(2):779–786, 2021

    2021

  4. [4]

    Ego-planner: An esdf-free gradient-based local planner for quadrotors.IEEE Robotics and Automation Letters, 6(2):478– 485, 2020

    Xin Zhou, Zhepei Wang, Hongkai Ye, Chao Xu, and Fei Gao. Ego-planner: An esdf-free gradient-based local planner for quadrotors.IEEE Robotics and Automation Letters, 6(2):478– 485, 2020

    2020

  5. [5]

    Aerialvln: Vision-and-language navigation for uavs

    Shubo Liu, Hongsheng Zhang, Yuankai Qi, Peng Wang, Yanning Zhang, and Qi Wu. Aerialvln: Vision-and-language navigation for uavs. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15384–15394, 2023

    2023

  6. [6]

    Towards realistic UA V vision-language navigation: Platform, benchmark, and methodology.CoRR, abs/2410.07087, 2024

    Xiangyu Wang, Donglin Yang, Ziqin Wang, Hohin Kwan, Jinyu Chen, Wenjun Wu, Hong- sheng Li, Yue Liao, and Si Liu. Towards realistic UA V vision-language navigation: Platform, benchmark, and methodology.CoRR, abs/2410.07087, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Openfly: A comprehensive platform for aerial vision-language navigation.arXiv preprint arXiv:2502.18041, 2025

    Yunpeng Gao, Chenhui Li, Zhongrui You, Junli Liu, Zhen Li, Pengan Chen, Qizhi Chen, Zhonghan Tang, Liansheng Wang, Penghui Yang, et al. Openfly: A comprehensive platform for aerial vision-language navigation.arXiv preprint arXiv:2502.18041, 2025

    arXiv 2025

  8. [8]

    Citynav: A large-scale dataset for real-world aerial navigation

    Jungdae Lee, Taiki Miyanishi, Shuhei Kurita, Koya Sakamoto, Daichi Azuma, Yutaka Matsuo, and Nakamasa Inoue. Citynav: A large-scale dataset for real-world aerial navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5912–5922, 2025

    2025

  9. [9]

    Openvln: Open- world aerial vision-language navigation.arXiv preprint arXiv:2511.06182, 2025

    Peican Lin, Gan Sun, Chenxi Liu, Fazeng Li, Weihong Ren, and Yang Cong. Openvln: Open- world aerial vision-language navigation.arXiv preprint arXiv:2511.06182, 2025

    arXiv 2025

  10. [10]

    Reid, Stephen Gould, and Anton van den Hengel

    Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 10

    2018

  11. [11]

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

    Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. InComputer Vision – ECCV 2020, 2020

    2020

  12. [12]

    Touchdown: Natural language navigation and spatial reasoning in visual street environments

    Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. Touchdown: Natural language navigation and spatial reasoning in visual street environments. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12538–12547, 2019

    2019

  13. [13]

    Room-across- room: Multilingual vision-and-language navigation with dense spatiotemporal grounding

    Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across- room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. InProceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020

    2020

  14. [14]

    Embodied question answering

    Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1–10, 2018

    2018

  15. [15]

    Eqa-mx: Embodied question answering using multimodal expression

    Md Mofijul Islam, Alexi Gladstone, Riashat Islam, and Tariq Iqbal. Eqa-mx: Embodied question answering using multimodal expression. InInternational Conference on Learning Representations, volume 2024, pages 48079–48102, 2024

    2024

  16. [16]

    Embodied Question Answering in photorealistic environments with point cloud perception

    Erik Wijmans, Samyak Datta, Oleksandr Maksymets, Abhishek Das, Georgia Gkioxari, Stefan Lee, Irfan Essa, Devi Parikh, and Dhruv Batra. Embodied Question Answering in photorealistic environments with point cloud perception. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6659–6668, 2019

    2019

  17. [17]

    CityEQA: A hierarchical llm agent on embodied question answering benchmark in city space.arXiv preprint arXiv:2502.12532, 2025

    Yong Zhao, Kai Xu, Zhengqiu Zhu, Yue Hu, Zhiheng Zheng, Yingfeng Chen, Yatai Ji, Chen Gao, Yong Li, and Jincai Huang. CityEQA: A hierarchical llm agent on embodied question answering benchmark in city space.arXiv preprint arXiv:2502.12532, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Explore until confident: Efficient exploration for embodied question answering

    Allen Z Ren, Jaden Clark, Anushri Dixit, Masha Itkina, Anirudha Majumdar, and Dorsa Sadigh. Explore until confident: Efficient exploration for embodied question answering. InRobotics: Science and Systems (RSS), 2024

    2024

  19. [19]

    Openeqa: Embodied question answering in the era of foundation models

    Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. Openeqa: Embodied question answering in the era of foundation models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16488–16498, 2024

    2024

  20. [20]

    Beyond the destination: A novel benchmark for exploration-aware embodied question answering.arXiv preprint arXiv:2503.11117, 2025

    Kaixuan Jiang, Yang Liu, Weixing Chen, Jingzhou Luo, Ziliang Chen, Ling Pan, Guanbin Li, and Liang Lin. Beyond the destination: A novel benchmark for exploration-aware embodied question answering.arXiv preprint arXiv:2503.11117, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    UA V-Flow colosseo: A real-world benchmark for flying-on-a-word uav imitation learning.arXiv preprint arXiv:2505.15725, 2025

    Xiangyu Wang, Donglin Yang, Yue Liao, Wenhao Zheng, Bin Dai, Wenjun Wu, Hongsheng Li, and Si Liu. UA V-Flow colosseo: A real-world benchmark for flying-on-a-word uav imitation learning.arXiv preprint arXiv:2505.15725, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Aerialvla: A vision-language-action model for aerial navigation with online dialogue

    Jinyu Chen, Hongyu Li, Zongheng Tang, Xiaoduo Li, Wenjun Wu, and Si Liu. Aerialvla: A vision-language-action model for aerial navigation with online dialogue. InProceedings of the AAAI Conference on Artificial Intelligence, pages 18161–18169, 2026

    2026

  23. [23]

    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

  24. [25]

    Mavic series

    DJI. Mavic series. https://store.dji.com/cn/shop/mavic-series, 2026. DJI Online Store. Accessed: 2026-05-07. 11

    2026

  25. [26]

    Multi-Target Embodied Question Answering

    Licheng Yu, Xinlei Chen, Georgia Gkioxari, Mohit Bansal, Tamara L Berg, and Dhruv Batra. Multi-Target Embodied Question Answering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6309–6318, 2019

    2019

  26. [27]

    Industryeqa: Pushing the frontiers of embodied question answering in industrial scenarios.Advances in Neural Information Processing Systems, 38, 2026

    Yifan Li, Yuhang Chen, Anh Dao, Lichi Li, Zhongyi Cai, Zhen Tan, Tianlong Chen, and Yu Kong. Industryeqa: Pushing the frontiers of embodied question answering in industrial scenarios.Advances in Neural Information Processing Systems, 38, 2026

    2026

  27. [28]

    Multi-step reasoning for embodied question answering via tool augmentation.arXiv preprint arXiv:2510.20310, 2025

    Mingliang Zhai, Hansheng Liang, Xiaomeng Fan, Zhi Gao, Chuanhao Li, Che Sun, Xu Bin, Yuwei Wu, and Yunde Jia. Multi-step reasoning for embodied question answering via tool augmentation.arXiv preprint arXiv:2510.20310, 2025

    arXiv 2025

  28. [29]

    Knowledge-based em- bodied question answering.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11948–11960, 2023

    Sinan Tan, Mengmeng Ge, Di Guo, Huaping Liu, and Fuchun Sun. Knowledge-based em- bodied question answering.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11948–11960, 2023

    2023

  29. [30]

    VideoNavQA: Bridging the gap between visual and embodied question answering

    C˘at˘alina Cangea, Eugene Belilovsky, Pietro Liò, and Aaron Courville. VideoNavQA: Bridging the gap between visual and embodied question answering. InBritish Machine Vision Conference (BMVC), 2019

    2019

  30. [31]

    Memory-centric embodied question answering.arXiv preprint arXiv:2505.13948, 2025

    Mingliang Zhai, Zhi Gao, Yuwei Wu, and Yunde Jia. Memory-centric embodied question answering.arXiv preprint arXiv:2505.13948, 2025

    arXiv 2025

  31. [32]

    Episodic memory question answering

    Samyak Datta, Sameer Dharur, Vincent Cartillier, Ruta Desai, Mukul Khanna, Dhruv Batra, and Devi Parikh. Episodic memory question answering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19119–19128, 2022

    2022

  32. [33]

    Multi-agent embodied question answering in interactive environments

    Sinan Tan, Weilai Xiang, Huaping Liu, Di Guo, and Fuchun Sun. Multi-agent embodied question answering in interactive environments. InEuropean Conference on Computer Vision, pages 663–678. Springer, 2020

    2020

  33. [34]

    Noisyeqa: Benchmarking embodied question answering against noisy queries.arXiv preprint arXiv:2412.10726, 2024

    Tao Wu, Chuhao Zhou, Yen Heng Wong, Lin Gu, and Jianfei Yang. Noisyeqa: Benchmarking embodied question answering against noisy queries.arXiv preprint arXiv:2412.10726, 2024

    Pith/arXiv arXiv 2024

  34. [35]

    CityNav: Language-goal aerial navigation dataset with geographic information.CoRR, abs/2406.14240, 2024

    Jungdae Lee, Taiki Miyanishi, Shuhei Kurita, Koya Sakamoto, Daichi Azuma, Yutaka Matsuo, and Nakamasa Inoue. CityNav: Language-goal aerial navigation dataset with geographic information.CoRR, abs/2406.14240, 2024

    Pith/arXiv arXiv 2024

  35. [36]

    Uav-on: A benchmark for open-world object goal navigation with aerial agents

    Jianqiang Xiao, Yuexuan Sun, Yixin Shao, Boxi Gan, Rongqiang Liu, Yanjin Wu, Weili Guan, and Xiang Deng. Uav-on: A benchmark for open-world object goal navigation with aerial agents. InProceedings of the 33rd ACM International Conference on Multimedia, pages 13023–13029, 2025

    2025

  36. [37]

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

    Xu Liu, Yu Liu, Hanshuo Qiu, Yang Qirong, and Zhouhui Lian. Indooruav: Benchmarking vision-language uav navigation in continuous indoor environments. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 23864–23872, 2026

    2026

  37. [38]

    Air-vla: Vision-language-action systems for aerial manipulation.arXiv preprint arXiv:2601.21602, 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.arXiv preprint arXiv:2601.21602, 2026

    arXiv 2026

  38. [39]

    VLA-AN: An efficient and onboard vision- language-action framework for aerial navigation in complex environments.arXiv preprint arXiv:2512.15258, 2025

    Yuze Wu, Mo Zhu, others, and Fei Gao. VLA-AN: An efficient and onboard vision- language-action framework for aerial navigation in complex environments.arXiv preprint arXiv:2512.15258, 2025

    arXiv 2025

  39. [40]

    Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  40. [41]

    π0.5: A Vision- Language-Action Model with Open-World Generalization.arXiv preprint arXiv:2504.16054, 2025

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al. π0.5: A Vision- Language-Action Model with Open-World Generalization.arXiv preprint arXiv:2504.16054, 2025. 12

    Pith/arXiv arXiv 2025

  41. [42]

    π0: A Vision-Language-Action Flow Model for General Robot Control.arXiv preprint arXiv:2410.24164, 2024

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. π0: A Vision-Language-Action Flow Model for General Robot Control.arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  42. [43]

    Embodiedcity: A benchmark platform for embodied agent in real-world city environment.arXiv preprint arXiv:2410.09604, 2024

    Chen Gao, Baining Zhao, Weichen Zhang, Jinzhu Mao, Jun Zhang, Zhiheng Zheng, Fanhang Man, Jianjie Fang, Zile Zhou, Jinqiang Cui, et al. Embodiedcity: A benchmark platform for embodied agent in real-world city environment.arXiv preprint arXiv:2410.09604, 2024

    Pith/arXiv arXiv 2024

  43. [44]

    Libero: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776–44791, 2023

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776–44791, 2023

    2023

  44. [45]

    Openai gpt-5 system card.arXiv preprint arXiv:2601.03267, 2025

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. Openai gpt-5 system card.arXiv preprint arXiv:2601.03267, 2025

    Pith/arXiv arXiv 2025

  45. [46]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    Pith/arXiv arXiv 2025

  46. [47]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  47. [48]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    2023

  48. [49]

    Fine-tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645, 2025

    Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645, 2025

    Pith/arXiv arXiv 2025

  49. [50]

    Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter.IEEE Robotics and Automation Letters, 6(2):3317–3324, 2021

    Wei Xu and Fu Zhang. Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter.IEEE Robotics and Automation Letters, 6(2):3317–3324, 2021

    2021

  50. [51]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  51. [52]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  52. [53]

    move up",

    Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10012–10022, 2020. 13 A Technical appendices and supplementary material A.1 Dataset Details A.1.1 Details of Trajectory/Image Curation in...

    2020