Pith. sign in

REVIEW 4 major objections 5 minor 25 references

This paper claims that aerial vision-language navigation should be decomposed: an MLLM only grounds language to a pixel, depth back-projection fixes a 3D goal, and a geometric planner executes the flight.

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

2026-08-03 05:32 UTC pith:K2YGRYX5

load-bearing objection Fly0 is a sensible modular UAV VLN system with solid simulation gains; the NE overclaim and underspecified real-world baseline protocol need fixing before I trust the headline numbers. the 4 major comments →

arxiv 2602.15875 v2 pith:K2YGRYX5 submitted 2026-02-02 cs.RO cs.AI

Fly0: Persistent Metric Anchoring for Zero-Shot Aerial Vision-Language Navigation

Zhenxing Xu , Yihong Lu , Weidong Bao , Zhengqiu Zhu , Jingxuan Zhou , Zhichuang Wang , Ji Wang , Lihua Liu
show 1 more author
Wei He
This is my paper
classification cs.RO cs.AI
keywords aerial vision-language navigationzero-shot navigationmultimodal large language modelssemantic groundingdepth back-projectionEgo-Plannerobject permanenceUAV 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.

This paper claims that zero-shot aerial vision-language navigation fails when a multimodal language model is used as the low-level flight controller, and succeeds when the model is restricted to semantic grounding. In Fly0, the MLLM turns an instruction into a single 2D pixel in the current camera view; a depth-based back-projection lifts that pixel to a fixed 3D world coordinate; and a gradient-based geometric planner generates collision-free trajectories toward that coordinate. The paper reports that this decomposition outperforms state-of-the-art baselines by over 20 percentage points in success rate and roughly halves navigation error, while cutting average task time nearly in half relative to other zero-shot agents. The broader claim is that the bottleneck in visual-language navigation is semantic grounding, not control, and that a persistent 3D goal gives the drone object permanence even when the target leaves the field of view.

Core claim

The paper's central claim is that semantic grounding and motion execution are separable, and that separating them resolves the latency, oscillation, and context-loss problems of MLLM-as-controller agents. Fly0's perception module outputs a target pixel p2D = (x,y) from the instruction and RGB frame; the geometric module reads the depth d at that pixel and back-projects it through the camera intrinsics and the drone's pose chain to obtain a persistent world goal G_world = R_t(R_IC * dK^{-1}[x,y,1]^T + t_IC) + t_t. Ego-Planner then optimizes a B-spline trajectory with smoothness, collision, and dynamic-feasibility terms at 50 Hz, while the MLLM is re-queried only about every two seconds. Becau

What carries the argument

The load-bearing mechanism is the persistent 3D target anchor: a single back-projection identity G_world = R_t(R_IC G_cam + t_IC) + t_t, with G_cam = d K^{-1}[x,y,1]^T, turns one MLLM pixel prediction into a fixed metric destination. The anchor is refreshed at 0.5 Hz by re-grounding, and executed by Ego-Planner, a gradient-based local planner that minimizes a B-spline cost combining smoothness (jerk), a cubic collision barrier, and velocity/acceleration limits. This division of labor means the MLLM never issues a control command and never needs to remember the target across frames.

Load-bearing premise

The reported 20+ point success-rate advantage assumes the eight baseline models were fairly adapted to the custom quadrotor and evaluated under the same RTK-based success criterion; the paper does not describe that adaptation, so an unfair setup would make the gap an artifact.

What would settle it

Re-run the real-world comparison giving all eight baselines the same onboard RGB-D and LiDAR streams and the same RTK success threshold; if the average success-rate gap over the strongest zero-shot baseline falls below the claimed 20 points, the central claim fails. A cheaper check: verify that replacing sensor depth with MLLM-estimated depth reproduces the paper's reported drop from 70.43% to 56.47% on AerialVLN.

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

If this is right

  • If the decomposition is correct, MLLM-as-controller agents carry an inherent handicap: their latency and trajectory oscillation are structural, not tuning problems.
  • Object permanence follows directly: once the target is anchored in world coordinates, occlusion, field-of-view loss, and viewpoint change do not interrupt navigation.
  • The 0.5 Hz re-grounding strategy implies the system's compute is dominated by planning, making the approach roughly twice as fast as zero-shot baselines and compatible with edge deployment.
  • The ablation shows sensor depth is critical: replacing it with MLLM-estimated depth drops success rate from 70.43% to 56.47%, so the method's precision is inherited from the depth sensor and calibration chain.
  • The dependency test shows the grounding task is backbone-agnostic across several MLLMs, implying the semantic step does not demand frontier-scale reasoning.

Where Pith is reading between the lines

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

  • Beyond the paper: the same 'ground to pixel, project to 3D, plan geometrically' split should transfer to ground robots and manipulators, where a persistent goal frame could replace per-frame policy inference.
  • Beyond the paper: the fixed 0.5 Hz re-grounding cadence is a natural tuning knob; re-grounding more often when depth uncertainty or target distance is high could reduce the undershoot/overshoot the paper reports.
  • Beyond the paper: since the paper's failure cases are repeated-object ambiguity and depth error, future gains are more likely to come from better depth fusion and disambiguation prompts than from larger MLLMs.
  • Beyond the paper: the backbone-insensitivity result suggests that for this pipeline the MLLM is a replaceable sensor, making it a testable claim that any VLM with adequate referring-grounding ability will yield equivalent navigation performance.

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 / 5 minor

Summary. The paper proposes Fly0, a modular zero-shot aerial vision-language navigation (VLN) system that decouples semantic grounding from geometric planning. A multimodal LLM (MLLM) maps a natural-language instruction to 2D pixel coordinates in the current RGB frame; the system then back-projects this point to 3D using depth and camera intrinsics; a LiDAR-based local planner (Ego-Planner) generates collision-free trajectories toward the persistent 3D goal. The authors claim that this architecture resolves the granularity mismatch, high inference latency, and fragile context maintenance of MLLM-as-controller baselines, and report large improvements in Success Rate (SR) and Navigation Error (NE) on the AerialVLN and OpenFly simulation datasets plus a custom real-world dataset of 800 instructions. The manuscript also includes ablations of the depth-input and trajectory-optimization components and a dependency test across five MLLM backbones.

Significance. If the empirical claims hold, the paper makes a useful contribution by showing that a decoupled semantic-grounding / geometric-planning pipeline can outperform both end-to-end and MLLM-as-controller baselines in zero-shot aerial VLN, while reducing latency and improving interpretability. The central architectural idea is clearly stated and the modularity is attractive. Strengths include the public code release, the explicit ablation of depth input and trajectory optimization, and the MLLM backbone-dependency study, which supports the claim that the method is not tied to a single model. However, the headline quantitative claims rest on an empirical comparison whose real-world component is not yet sufficiently documented, and the reported numbers do not match the abstract's '~50% NE reduction' claim in the way one would expect. These issues are fixable but currently prevent full confidence in the central claim.

major comments (4)
  1. [§5.1.5 and Table 1 (real-world columns)] The real-world comparison is the only evidence for the generalization claim, but the deployment protocol for the eight baselines is not described. It is unclear how AirSim-trained end-to-end policies (OpenUAV, UAV-Flow, AerialVLN, OpenFly) were adapted to the TTF-400LIVO-RTK platform, what sensor streams they received (RGB-only, RGB-D, LiDAR), how their discrete actions were converted to PX4 velocity commands, or whether the same RTK-based success criterion (d_th=5m) was applied to them. Without this protocol, the >20% SR margin (62.86% vs 35.07% for SPF) could be an artifact of baseline deployment mismatches rather than of the proposed method. Please provide a full protocol or restrict the real-world claim accordingly.
  2. [Abstract and §5.2.1] The abstract states that Fly0 reduces NE 'by approximately 50%', but Table 1 shows much smaller reductions against the strongest zero-shot baseline: on AerialVLN, 39.57m to 27.19m is a 31% reduction; on OpenFly, 46.57m to 29.47m is 37%; in the real world, 20.48m to 13.76m is 33%. The ~50% figure only appears when comparing against weaker end-to-end baselines (e.g., OpenFly: 67.78→27.19 is 60%, OpenUAV real-world: 27.39→13.76 is 50%). The claim should be reworded to name the comparison pairs explicitly and avoid the impression that the reduction is uniform against all state-of-the-art baselines.
  3. [§5.1.4] All metrics in Table 1 are point estimates. The text states that each instruction is averaged over 20 independent trials, but no variance, confidence interval, or significance test is reported. Given the small number of trials, it is not clear whether the headline margins (e.g., SR 70.43% vs 46.72% on AerialVLN) are statistically reliable. Reporting distributions or CIs for the main comparisons would materially strengthen the RQ1/RQ2 conclusions.
  4. [§4.3.2 and Table 2] The trajectory optimizer introduces several tunable parameters (λ_s=1.0, λ_c=10.0, λ_d=1.0, d_safe=0.5m, v_max=4.0m/s, a_max=3.0m/s²). No sensitivity analysis is provided, so it is unclear how robust the reported SR/NE improvements are to these choices. A small sensitivity sweep would help distinguish the benefit of decoupled planning from the effect of favorable planner tuning.
minor comments (5)
  1. [Table 1] Add units to the Time columns and state explicitly whether Time is computed only for successful episodes or all episodes. The caption says 'All baseline methods are configured with the MLLMs specified in their official implementations,' but four of the baselines are end-to-end and do not use MLLMs; the caption should be split or clarified.
  2. [Appendix A] The prompt hardcodes image dimensions as '720x480'. Confirm that this matches all simulation and real-world deployments, or describe how the coordinates are rescaled when the input resolution differs.
  3. [§5.1.3 (Real-World Dataset)] Specify how the 800 real-world instructions were generated/annotated and whether the same instruction set and success criterion were given to each baseline. This is important for auditing the real-world comparison.
  4. [§5.3 (Ablation)] The 'wo. DI' variant replaces sensor depth with MLLM-based depth estimation 'following the methodology adopted by SPF'. Please provide a concrete description of this implementation, since the result depends on it.
  5. [General] Minor typos and layout issues: 'UA V' appears with an inconsistent space (e.g., in the Introduction and Section 2); the phrase 'Section 4.1' in Appendix A should be cross-checked; Figure 5/6 captions are descriptive but do not indicate failure cases, despite the failure-case analysis in §5.2.4.

Circularity Check

0 steps flagged

No significant circularity: Fly0's claims are empirical system comparisons, not derivations that reduce to their inputs.

full rationale

Fly0 makes no formal derivation whose output equals an input by construction. The pipeline (MLLM 2D grounding -> depth back-projection -> Ego-Planner) is a composition of external components (pretrained MLLMs, Ego-Planner [Zhou et al., 2020]) with the paper's own projection equations (4)-(6) and trajectory costs (7)-(10); none of these define performance in terms of the reported SR/NE. The headline superiority claim rests on Table 1, an empirical benchmark comparison, not on a fitted parameter renamed as a prediction. Ablations (Table 2) independently remove depth input and trajectory optimization, and Table 3 swaps MLLM backbones, showing parity that is not assumed by the method. No load-bearing self-citation or imported uniqueness theorem appears: baselines are external works and the planner is cited from other authors. The paper itself flags real limitations in §5.2.4 (semantic ambiguity, depth-induced projection errors), and those are stated as failure modes, not as support for the central claim. The only substantive concern is external to circularity: the real-world baseline deployment protocol (sensor inputs, action conversion, sim-to-real adaptation) is underspecified in §5.1.5 and Appendix C, so the >20% SR margin could in principle be an evaluation-fairness artifact. That is a correctness/evidence risk, not a circularity reduction, and under the stated rules it does not raise the circularity score.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The paper contributes an architecture and an empirical evaluation, not a derivation. The central claim depends on the reliability of the MLLM's coordinate grounding, the validity of a single-pixel depth projection, and the fairness of the baseline comparisons; these are domain assumptions rather than mathematical axioms. No free parameters are fitted to optimize the headline results, but several hand-set thresholds (success radius, safety margin, re-grounding rate) directly shape the reported metrics.

free parameters (6)
  • Ego-Planner weights (λs, λc, λd) = 1.0, 10.0, 1.0
    Hand-chosen smoothing/safety/feasibility weights in §4.3.2; no sensitivity analysis is given, so performance depends on these values.
  • Collision clearance margin d_safe = 0.5 m
    Set in §4.3.2; determines when the collision penalty activates.
  • Kinematic limits v_max, a_max = 4.0 m/s, 3.0 m/s^2
    Feasibility bounds in Eq. 10; chosen for agility, not fitted to data.
  • Success threshold δ = 5 m
    Evaluation criterion in §5.1.4; directly affects SR. Average trajectory lengths are 200 m in simulation and 50 m real-world, so this threshold is generous.
  • MLLM re-grounding frequency = 0.5 Hz (every 2 s)
    Low-frequency re-querying strategy in §4.1; chosen to correct projection errors but not systematically tuned.
  • MLLM prompt image size assumption = 720x480
    The prompt in Appendix A bakes in this resolution; if actual input differs, coordinates scale improperly.
axioms (6)
  • domain assumption The MLLM reliably grounds the instruction to a single correct center pixel, returning (0,0) when the target is absent.
    Prompt in Appendix A assumes this behavior; §5.2.4 admits failure under object repetition and ambiguity.
  • domain assumption The depth value at the single grounded pixel is a valid range measurement for the target.
    Eq. 4 uses d = D_t(x,y); sensor noise, occlusions, and thin structures (wires) make this unreliable, as acknowledged in §5.2.4.
  • domain assumption Camera-IMU extrinsics and body-to-world pose are accurate enough for the 3D anchor.
    Eq. 5 requires R_IC, t_IC, R_t, t_t; any calibration or odometry error directly shifts G_world.
  • domain assumption LiDAR point clouds provide sufficient local obstacle coverage for Ego-Planner.
    The local map in §4.3.1 is a sliding-window ring buffer; sparse returns or obstacles outside the LiDAR FOV cannot be represented.
  • domain assumption All baselines were evaluated under the same platform and protocol as Fly0 in the real-world dataset.
    The paper gives no adaptation details for sim-trained end-to-end baselines deployed on the TTF-400LIVO-RTK platform (§5.1.5, Appendix C).
  • domain assumption Ego-Planner is a reliable external planner for collision-free quadrotor trajectories.
    The paper uses the published Ego-Planner algorithm as a black-box component from Zhou et al. 2020.

pith-pipeline@v1.3.0-alltime-deepseek · 12867 in / 13873 out tokens · 119300 ms · 2026-08-03T05:32:04.218540+00:00 · methodology

0 comments
read the original abstract

Current Visual-Language Navigation (VLN) methodologies face a trade-off between semantic understanding and control precision. While Multimodal Large Language Models (MLLMs) offer superior reasoning, deploying them as low-level controllers leads to high latency, trajectory oscillations, and poor generalization due to weak geometric grounding. To address these limitations, we propose Fly0, a framework that decouples semantic reasoning from geometric planning. The proposed method operates through a three-stage pipeline: (1) an MLLM-driven module for grounding natural language instructions into 2D pixel coordinates; (2) a geometric projection module that utilizes depth data to localize targets in 3D space; and (3) a geometric planner that generates collision-free trajectories. This mechanism enables robust navigation even when visual contact is lost. By eliminating the need for continuous inference, Fly0 reduces computational overhead and improves system stability. Extensive experiments in simulation and real-world environments demonstrate that Fly0 outperforms state-of-the-art baselines, improving the Success Rate by over 20\% and reducing Navigation Error (NE) by approximately 50\% in unstructured environments. Our code is available at https://github.com/xuzhenxing1/Fly0.

Figures

Figures reproduced from arXiv: 2602.15875 by Jingxuan Zhou, Ji Wang, Lihua Liu, Weidong Bao, Wei He, Yihong Lu, Zhengqiu Zhu, Zhenxing Xu, Zhichuang Wang.

Figure 1
Figure 1. Figure 1: Illustration of three navigation architectures. (a) End-to￾End models directly output actions. (b) Methods using MLLM as a controller typically rely on waypoint prediction and topological maps. (c) Our proposed Fly0 framework leverages the semantic reasoning of MLLMs for precise Coordinate Acquisition and Ge￾ometric Progression, enabling the generation of smooth metric trajectories in a zero-shot manner. F… view at source ↗
Figure 2
Figure 2. Figure 2: Schematic diagram of the proposed visual-language navigation system. The framework leverages an MLLM to bridge the gap between semantic instructions and metric navigation. By grounding the user’s command into a 2D target position, the system employs an un-projection module to derive the corresponding spatial coordinates. This precise 3D localization enables the Ego-Planner (Zhou et al., 2020) to compute op… view at source ↗
Figure 3
Figure 3. Figure 3: The execution pipeline from visual perception to trajectory optimization. First, the Input Stream combines RGB-D images and instructions to locate the target in the image frame. Next, the Back-Projection Engine maps this 2D point to the 3D world frame via the pinhole camera model. Finally, based on the projected 3D target and local LiDAR sensing, the Ego-Planner optimizes the flight path by solving a gradi… view at source ↗
Figure 4
Figure 4. Figure 4: presents the complete prompt template utilized to guide the MLLM in translating visual inputs and navigation instructions into precise 2D coordinates. Please locate the destination specified by the user within the image based on the user's complete instruction. The image dimensions are 720x480 pixels, with the origin (0,0) at the top-left corner. Please note the following when understanding the user's inst… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative visualization of a sequential navigation task in a complex urban environment. The UAV executes a multi-step composite instruction (bottom) sequentially. The three rows correspond to the three sub-tasks: approaching the tree, navigating to the streetlight, and reaching the final destination. The columns display the onboard First-Person View (FPV) with LiDAR perception, followed by Frontal and La… view at source ↗
Figure 6
Figure 6. Figure 6: Extensive qualitative evaluation across diverse simulation environments. The figure presents six additional independent trials ranging from urban streets to parks and island terrains. Each row depicts the successful execution of a long-horizon composite instruction (displayed at the bottom). These results demonstrate the robustness of Fly0 in handling heterogeneous scene topologies and its capability to gr… 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

25 extracted references · 7 linked inside Pith

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

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

    Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., S \"u nderhauf, N., Reid, I., Gould, S., and Van Den Hengel, A. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 3674--3683, 2018

    2018

  3. [3]

    K., Chuang, J., V \'a zquez, M., and Savarese, S

    Chen, K., Chen, J. K., Chuang, J., V \'a zquez, M., and Savarese, S. Topological planning with transformers for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11276--11286, 2021

    2021

  4. [4]

    Speaker-follower models for vision-and-language navigation

    Fried, D., Hu, R., Cirik, V., Rohrbach, A., Andreas, J., Morency, L.-P., Berg-Kirkpatrick, T., Saenko, K., Klein, D., and Darrell, T. Speaker-follower models for vision-and-language navigation. Advances in neural information processing systems, 31, 2018

    2018

  5. [5]

    Aerial vision-and-language navigation via semantic-topo-metric representation guided llm reasoning

    Gao, Y., Wang, Z., Jing, L., Wang, D., Li, X., and Zhao, B. Aerial vision-and-language navigation via semantic-topo-metric representation guided llm reasoning. arXiv preprint arXiv:2410.08500, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Openfly: A comprehensive platform for aerial vision-language navigation

    Gao, Y., Li, C., You, Z., Liu, J., Li, Z., Chen, P., Chen, Q., Tang, Z., Wang, L., Yang, P., et al. Openfly: A comprehensive platform for aerial vision-language navigation. arXiv preprint arXiv:2502.18041, 2025

    arXiv 2025

  7. [7]

    Cross-modal map learning for vision and language navigation

    Georgakis, G., Schmeckpeper, K., Wanchoo, K., Dan, S., Miltsakaki, E., Roth, D., and Daniilidis, K. Cross-modal map learning for vision and language navigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 15460--15470, 2022

    2022

  8. [8]

    Towards learning a generic agent for vision-and-language navigation via pre-training

    Hao, W., Li, C., Li, X., Carin, L., and Gao, J. Towards learning a generic agent for vision-and-language navigation via pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 13137--13146, 2020

    2020

  9. [9]

    Vln bert: A recurrent vision-and-language bert for navigation

    Hong, Y., Wu, Q., Qi, Y., Rodriguez-Opazo, C., and Gould, S. Vln bert: A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pp.\ 1643--1653, 2021

    2021

  10. [10]

    Y., Lin, Y.-S., Lee, Y., Su, C.-H., Lee, J.-Y., Tsai, S.-R., Lin, C.-Y., Chen, K.-W., Ke, T.-W., and Liu, Y.-L

    Hu, C. Y., Lin, Y.-S., Lee, Y., Su, C.-H., Lee, J.-Y., Tsai, S.-R., Lin, C.-Y., Chen, K.-W., Ke, T.-W., and Liu, Y.-L. See, point, fly: A learning-free vlm framework for universal unmanned aerial navigation. In Conference on Robot Learning, pp.\ 4697--4708. PMLR, 2025

    2025

  11. [11]

    Visual language maps for robot navigation

    Huang, C., Mees, O., Zeng, A., and Burgard, W. Visual language maps for robot navigation. arXiv preprint arXiv:2210.05714, 2022

    Pith/arXiv arXiv 2022

  12. [12]

    Code as policies: Language model programs for embodied control

    Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., and Zeng, A. Code as policies: Language model programs for embodied control. arXiv preprint arXiv:2209.07753, 2022

    Pith/arXiv arXiv 2022

  13. [13]

    Aerialvln: Vision-and-language navigation for uavs

    Liu, S., Zhang, H., Qi, Y., Wang, P., Zhang, Y., and Wu, Q. Aerialvln: Vision-and-language navigation for uavs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 15384--15394, 2023

    2023

  14. [14]

    Navagent: Multi-scale urban street view fusion for uav embodied vision-and-language navigation

    Liu, Y., Yao, F., Yue, Y., Xu, G., Sun, X., and Fu, K. Navagent: Multi-scale urban street view fusion for uav embodied vision-and-language navigation. arXiv preprint arXiv:2411.08579, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    A., Tadevosyan, G., Akhmetkazy, A., Cabrera, M

    Sautenkov, O., Yaqoot, Y., Lykov, A., Mustafa, M. A., Tadevosyan, G., Akhmetkazy, A., Cabrera, M. A., Martynov, M., Karaf, S., and Tsetserukou, D. Uav-vla: Vision-language-action system for large scale aerial mission generation. In 2025 20th ACM/IEEE International Conference on Human-Robot Interaction (HRI), pp.\ 1588--1592. IEEE, 2025

    2025

  16. [16]

    Lm-nav: Robotic navigation with large pre-trained models of language, vision, and action

    Shah, D., Osi \'n ski, B., Levine, S., et al. Lm-nav: Robotic navigation with large pre-trained models of language, vision, and action. In Conference on robot learning, pp.\ 492--504. PMLR, 2023

    2023

  17. [17]

    Y., and Zhang, L

    Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y.-F., Wang, W. Y., and Zhang, L. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 6629--6638, 2019

    2019

  18. [18]

    Towards realistic uav vision-language navigation: Platform, benchmark, and methodology

    Wang, X., Yang, D., Wang, Z., Kwan, H., Chen, J., Wu, W., Li, H., Liao, Y., and Liu, S. Towards realistic uav vision-language navigation: Platform, benchmark, and methodology. arXiv preprint arXiv:2410.07087, 2024

    Pith/arXiv arXiv 2024

  19. [19]

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

    Wang, X., Yang, D., Liao, Y., Zheng, W., Dai, B., Li, H., Liu, S., et al. Uav-flow colosseo: A real-world benchmark for flying-on-a-word uav imitation learning. arXiv preprint arXiv:2505.15725, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Geonav: Empowering mllms with explicit geospatial reasoning abilities for language-goal aerial navigation

    Xu, H., Hu, Y., Gao, C., Zhu, Z., Zhao, Y., Li, Y., and Yin, Q. Geonav: Empowering mllms with explicit geospatial reasoning abilities for language-goal aerial navigation. arXiv preprint arXiv:2504.09587, 2025

    arXiv 2025

  21. [21]

    Citynavagent: Aerial vision-and-language navigation with hierarchical semantic planning and global memory

    Zhang, W., Gao, C., Yu, S., Peng, R., Zhao, B., Zhang, Q., Cui, J., Chen, X., and Li, Y. Citynavagent: Aerial vision-and-language navigation with hierarchical semantic planning and global memory. arXiv preprint arXiv:2505.05622, 2025

    arXiv 2025

  22. [22]

    Cityeqa: A hierarchical llm agent on embodied question answering benchmark in city space

    Zhao, Y., Xu, K., Zhu, Z., Hu, Y., Zheng, Z., Chen, Y., Ji, Y., Gao, C., Li, Y., and Huang, J. Cityeqa: A hierarchical llm agent on embodied question answering benchmark in city space. arXiv preprint arXiv:2502.12532, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Navgpt: Explicit reasoning in vision-and-language navigation with large language models

    Zhou, G., Hong, Y., and Wu, Q. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 7641--7649, 2024

    2024

  24. [24]

    Ego-planner: An esdf-free gradient-based local planner for quadrotors

    Zhou, X., Wang, Z., Ye, H., Xu, C., and Gao, F. Ego-planner: An esdf-free gradient-based local planner for quadrotors. IEEE Robotics and Automation Letters, 6 0 (2): 0 478--485, 2020

    2020

  25. [25]

    Vision-language navigation with self-supervised auxiliary reasoning tasks

    Zhu, F., Zhu, Y., Chang, X., and Liang, X. Vision-language navigation with self-supervised auxiliary reasoning tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10012--10022, 2020

    2020