Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

FlightGPT: Towards Generalizable and Interpretable UAV Vision-and-Language Navigation with Vision-Language Models

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read FlightGPT claims a 9.22-point success-rate gain over the strongest baseline in unseen city navigation.

desk verdict FlightGPT's 7B SFT+GRPO model posts a genuinely strong CityNav result, but the paper never documents whether the baselines share its AerialVLN look-ahead planner, so the 9.22-point SOTA margin is not yet fully interpretable. read the letter →

arxiv 2505.12835 v1 pith:ZOUQKLTI submitted 2025-05-19 cs.CL cs.CV

Hengxing Cai , Jinhan Dong , Jingjun Tan , Jingcheng Deng , Sihang Li , Zhifeng Gao , Haidong Wang , Zicheng Su
show 2 more authors
Agachai Sumalee Renxin Zhong
This is my paper · ORCID
classification cs.CLcs.CV
keywords UAVvision-and-languagenavigationvision-languagemodelsgrouprelativepolicyoptimizationchain-of-thoughtreasoningsupervisedfine-tuningCityNavgeneralizationtounseenenvironmentsaerial
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that a 7-billion-parameter vision-language model, fine-tuned in two stages, can outperform much larger zero-shot models and prior map-based predictors on city-scale UAV vision-and-language navigation. FlightGPT first runs supervised fine-tuning on chain-of-thought demonstrations produced by a stronger VLM, then applies group relative policy optimization (GRPO) with rewards for goal accuracy, landmark reasoning, and format compliance. On the unseen test split of CityNav, it reports a success rate of 21.20%, 9.22 percentage points above the strongest baseline, Qwen2.5-VL-32B at 11.98%. If the claim holds, generalization in aerial vision-and-language navigation can be bought with training strategy and structured reasoning rather than raw model scale.

What carries the argument

The load-bearing machinery is the two-stage training pipeline on top of a vision-language model, with target-coordinate prediction as the intermediate representation. Stage one is supervised fine-tuning on chain-of-thought demonstrations generated by a stronger VLM, which gives the policy a structured initialization. Stage two is Group Relative Policy Optimization (GRPO), a reinforcement-learning update that compares sampled outputs within a group to estimate advantages, driven by a composite reward: goal accuracy (an exponential decay around the 20-meter success threshold), landmark reasoning (intersection-over-union of predicted and ground-truth landmark boxes), and format compliance (whether <think> and <answer> tags parse). At inference, FlightGPT takes a semantic map annotated with the drone's position and known landmarks plus the natural-language target description, emits a chain-of-thought and a predicted target location, and hands that location to the AerialVLN look-ahead mechanism, which simulates future trajectories to produce executable actions. The look-ahead mechanism is what closes the loop between the VLM's coordinate output and the drone's discrete action space.

What would settle it

Rerun Qwen2.5-VL-32B and MGP on the CityNav test-unseen split with the exact semantic-map prompt and AerialVLN look-ahead planner used by FlightGPT; if their success rates stay near 11.98% and 6.38%, the SFT+GRPO recipe is responsible for the gain, but if they rise toward 21.20%, the planner, not the training, carries the improvement.

Watch

Extended reading notes

Core claim

The paper's central claim is that FlightGPT, a framework built on the 7B vision-language model Qwen2.5-VL-7B, achieves the best reported results on the CityNav UAV vision-and-language navigation benchmark by combining supervised fine-tuning with a GRPO reinforcement-learning stage and chain-of-thought reasoning. On the test-unseen split, FlightGPT reports a success rate of 21.20%, navigation error of 76.2 meters, oracle success rate of 35.38%, and SPL of 19.24, compared with 11.98% success rate for Qwen2.5-VL-32B, the strongest baseline, and 6.38% for the map-based goal predictor MGP. The framework first predicts a target coordinate from a semantic map and a natural-language instruction, then uses a look-ahead planning mechanism to convert that coordinate into executable drone actions. The contribution is therefore a training-and-representation recipe: structured reasoning plus reinforcement learning, rather than a larger model, is what the authors claim drives generalization to unseen urban environments.

Load-bearing premise

The load-bearing premise is that all baselines were evaluated with the same input and action-planning protocol as FlightGPT, including the look-ahead mechanism that converts a predicted coordinate into executable actions; the paper does not verify this for the zero-shot VLM baselines.

Editorial extensions

If this is right

  • A 7B vision-language model trained with SFT plus GRPO can outperform a 32B zero-shot VLM on unseen city-scale navigation, so model scale is not the only lever for generalization.
  • The two-stage recipe nearly doubles navigation efficiency on unseen test splits: SPL rises from 10.76 for Qwen2.5-VL-32B to 19.24 for FlightGPT.
  • Adding the RL stage on top of SFT is what buys out-of-distribution robustness: SFT-only reaches 11.20% success on test-unseen, while SFT+RL reaches 21.20%.
  • Structured chain-of-thought with landmark bounding boxes and target coordinates makes the navigation decision inspectable, and the paper's automated scoring of 5,000 outputs indicates SFT+RL produces more complete and coherent reasoning than RL alone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the reported gain is mostly attributable to the look-ahead planner, the same SFT+GRPO coordinate-prediction recipe could transfer to other map-based navigation settings by swapping the action-conversion step, since the model itself only needs to output a map coordinate.
  • Editorial inference: because the SFT stage replaces the teacher's predicted target with the ground-truth location, SFT does not actually teach the model to connect its landmark boxes to its final answer; the RL IoU reward may be the only term enforcing that connection, which could be tested by ablating that reward alone.
  • Editorial inference: the format reward's success at enforcing structured output suggests a general, parse-free way to inject interpretability into closed-loop vision-language agents, namely rewarding tag compliance rather than hand-designing action decoders.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes FlightGPT, a UAV vision-and-language navigation system built on Qwen2.5-VL-7B. The system takes a semantic map and a natural language instruction, performs chain-of-thought reasoning inside <think> tags, and outputs a target coordinate inside <answer> tags, which is then converted into executable actions via the AerialVLN look-ahead mechanism. The training pipeline has two stages: supervised fine-tuning on CoT-annotated demonstrations generated by Qwen2.5-VL-32B, followed by GRPO reinforcement learning with a composite reward combining goal accuracy, landmark IoU, and output format compliance. On CityNav, the paper reports state-of-the-art results, including a test-unseen success rate of 21.20% compared to 11.98% for Qwen2.5-VL-32B, and an ablation showing SFT+RL outperforms SFT-only and RL-only. The paper also evaluates reasoning quality with GPT-4o scores and discusses limitations including the simulation-to-reality gap and deployment feasibility.

Significance. If the reported results hold, the paper makes a useful empirical contribution: it shows that a 7B VLM trained with SFT warm-up and GRPO can outperform larger zero-shot VLMs and prior map-based goal predictors on a city-scale navigation benchmark, and it provides a structured, interpretable CoT output format. The internal ablation is informative and the release of code and data is a concrete strength. The significance of the headline SOTA claim, however, depends on whether all baselines are evaluated under the same input, prompting, and action-planning protocol; the paper does not currently establish that parity, so the central claim is not fully verified.

major comments (3)
  1. [§4.1.3, Appendix B, and §3.2.1 step 3] The external baseline protocol is under-specified. FlightGPT's action pipeline includes the AerialVLN look-ahead mechanism to convert a predicted target coordinate into executable actions, and its prompt includes a semantic map with red landmark masks and current pose. The baseline descriptions in Appendix B do not state whether Seq2Seq, CMA, MGP, GPT-4o, Qwen2.5-VL, and LLaMA-3.2-11B-Vision receive the same semantic map, the same prompt template, and the same look-ahead controller. Because SR and SPL are end-to-end navigation metrics, the 9.22 percentage-point test-unseen gap in Table 2 could partly reflect differences in the action-execution component rather than the proposed SFT+GRPO training. The authors should report the exact evaluation protocol for each baseline and, preferably, run all VLM and MGP baselines through the same planner with the same prompt; without this, the headline SOTA claim is not verified.
  2. [§4.2.1, Table 2] All main results are reported from a single training run without error bars, multiple seeds, or significance tests. Given that the central claim depends on the gap between FlightGPT (SR 21.20) and the strongest baseline (SR 11.98), and that some val-seen versus test-unseen numbers vary substantially within a single model (e.g., Qwen2.5-VL-7B), the reader cannot distinguish a stable improvement from run-to-run noise. Please report at least three seeds with means and standard deviations or bootstrap confidence intervals, and specify the exact CityNav split construction and the test-unseen evaluation procedure, including how target coordinates and episode termination are handled.
  3. [§6 Limitations] The limitations section does not acknowledge the evaluation-protocol risk identified above. It lists the simulation-to-reality gap and deployment feasibility, but it does not mention that the external baselines may not have used the same look-ahead planner and prompting scheme as FlightGPT. Since that is the most load-bearing unverified premise for the state-of-the-art claim, the limitation should either be stated explicitly or addressed experimentally before the paper claims SOTA performance.
minor comments (6)
  1. [§4.1.4] The SFT data collection is described as 1,872 samples collected and filtered from Qwen2.5-VL-32B outputs, but the filtering criteria in §3.2.2 are stated qualitatively ('abnormal output formats', 'more than 20 meters away'). Please specify the exact numbers discarded at each step and the format-validity check used.
  2. [§3.2.2] The data generation step replaces the predicted target location with the ground truth for retained samples. This can create a mismatch between the generated CoT text and the final answer. Please state whether any analysis was done on the frequency of such replacements and whether the CoT text was also corrected to remain consistent with the ground-truth location.
  3. [§4.2.3, Table 3] The reasoning quality scores are averages of three GPT-4o evaluations over 5,000 samples, but no variance, inter-rater agreement, or correlation with downstream navigation performance is reported. Please add confidence intervals or a per-sample score distribution, and clarify how the random sample was drawn.
  4. [§3.1] There is a typo in the action list: 'forwad' should be 'forward'.
  5. [Appendix B] The CMA baseline is cited as 'Hu et al.' without a year or complete bibliographic entry; the reference list also appears to contain an incomplete entry for Hu et al. Please provide the full citation.
  6. [§4.1.1] The paper states that CityNav covers Birmingham and Cambridge but does not clarify which cities appear in val-seen, val-unseen, and test-unseen splits. A sentence describing the split construction would help readers interpret the generalization claims in Table 2.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: FlightGPT's SFT and RL are supervised by ground-truth CityNav targets and landmarks, and the SOTA claim is evaluated on held-out splits.

full rationale

FlightGPT's training and evaluation are not circular. The SFT stage uses Qwen2.5-VL-32B to generate CoT demonstrations, but these are filtered and the predicted target coordinates are replaced with CityNav ground truth; the SFT objective is next-token prediction on these corrected outputs. The RL stage applies GRPO with rewards computed against ground-truth target coordinates and landmark boxes from the training split. All reported navigation metrics (NE, SR, OSR, SPL) are evaluated on CityNav validation-seen, validation-unseen, and test-unseen splits that are not used for fitting, so the central SOTA claim is an external benchmark result rather than a consequence of the training objective. The only notable weakness, namely whether the baseline VLMs received the same look-ahead action planner as FlightGPT, is an evaluation-protocol fairness concern, not circularity in the paper's derivation chain. The GPT-4o-based reasoning-quality evaluation is auxiliary and does not support the load-bearing SOTA claim. No equation or fitted parameter in the paper reduces to the quantity being predicted, and no load-bearing self-citation chain appears.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the CityNav benchmark's maps and split definitions, on the AerialVLN look-ahead planner, on hand-chosen reward constants, and on LLM-as-judge evaluation. None of these are derived within the paper; they are imported from prior work or chosen by hand. The hand-chosen constants (d_cutoff, tau, format weights, equal reward weights) are not ablated, so their contribution to the reported 9.22 point gain is unknown. No new physical entities are introduced. Overall, the paper's contribution is an empirical system configuration rather than a parameter-free derivation.

free parameters (5)
  • d_cutoff = 80 meters
    Upper distance bound in Rgoal beyond which no reward is given (Section 3.2.3). Hand-chosen; not derived from the benchmark metric and no sensitivity analysis is reported.
  • tau = 100
    Decay temperature in the exponential goal reward. Hand-chosen; affects reward shaping for distances between 20 and 80 meters.
  • format_reward_weights = +0.5 for tags, +0.25 for landmark_bbox, +0.25 for target_location
    Hand-designed format compliance rewards in Section 3.2.3; the weights are arbitrary and no ablation is shown.
  • reward_component_weights = 1.0, 1.0, 1.0
    Rtotal is the unweighted sum Rgoal + RIoU + Rformat; equal weighting is a modeling choice, not derived or ablated.
  • sft_filter_distance = 20 meters
    Samples whose Qwen2.5-VL-32B predicted location is more than 20 meters from ground truth are discarded (Section 3.2.2). Threshold equals the success metric, but its effect on training is not ablated.
assumptions (6)
  • domain assumption CityNav semantic maps (red-masked landmarks, current position, heading, FOV) are a sufficient observation space for resolving target descriptions.
    Invoked in the input design (Section 3.2.2) and prompt template (Appendix A). If landmarks are missing or targets are not near a red-masked landmark, the CoT and landmark IoU reward lose their grounding.
  • domain assumption Targets are usually located near a red-masked landmark, as stated in the operational guidance prompt.
    The prompt in Appendix A instructs the model that the target is usually near a red-masked landmark; the landmark IoU reward depends on this regularity holding in CityNav.
  • domain assumption The AerialVLN look-ahead mechanism converts predicted target coordinates into effective actions, and this conversion is not the source of the reported gains.
    Used as the action planner after target prediction (Section 3.2.1, step 3). No control experiment isolates the planner contribution from the model contribution.
  • domain assumption GPT-4o scores of completeness, coherence, and fluency are a valid proxy for interpretability of CoT reasoning.
    Section 4.2.3 uses GPT-4o to score 5,000 outputs on three subjective criteria; no human agreement or validation is reported.
  • domain assumption The CityNav training split used for SFT and RL does not overlap with the val-unseen and test-unseen evaluation splits.
    The paper says samples were selected from the training set (Section 4.1.4) but does not explicitly confirm split hygiene or report any leakage checks for the unseen test set.
  • domain assumption Qwen2.5-VL-32B, after filtering and ground-truth replacement, provides high-quality SFT demonstrations.
    Section 3.2.2 relies on this without human evaluation of demonstration quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlightGPT: Towards Generalizable and Interpretable UAV Vision-and-Language Navigation with Vision-Language Models." pith.science (2026). https://pith.science/paper/ZOUQKLTI

@misc{pith2026250512835,
  author       = {Pith},
  title        = {Pith review of: FlightGPT: Towards Generalizable and Interpretable UAV Vision-and-Language Navigation with Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZOUQKLTI}},
  note         = {Machine review of arXiv:2505.12835}
}
read the original abstract

Unmanned Aerial Vehicle (UAV) Vision-and-Language Navigation (VLN) is vital for applications such as disaster response, logistics delivery, and urban inspection. However, existing methods often struggle with insufficient multimodal fusion, weak generalization, and poor interpretability. To address these challenges, we propose FlightGPT, a novel UAV VLN framework built upon Vision-Language Models (VLMs) with powerful multimodal perception capabilities. We design a two-stage training pipeline: first, Supervised Fine-Tuning (SFT) using high-quality demonstrations to improve initialization and structured reasoning; then, Group Relative Policy Optimization (GRPO) algorithm, guided by a composite reward that considers goal accuracy, reasoning quality, and format compliance, to enhance generalization and adaptability. Furthermore, FlightGPT introduces a Chain-of-Thought (CoT)-based reasoning mechanism to improve decision interpretability. Extensive experiments on the city-scale dataset CityNav demonstrate that FlightGPT achieves state-of-the-art performance across all scenarios, with a 9.22\% higher success rate than the strongest baseline in unseen environments. Our implementation is publicly available.

Figures

Figures reproduced from arXiv: 2505.12835 by the authors.

Figure 1
Figure 1. Workflow of FlightGPT for UAV VLN. FlightGPT takes multimodal input comprising a semantic map [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The two-stage training pipeline of FlightGPT. The pipeline consists of a supervised fine-tuning (SFT) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Reward (train) and success rate (test) over [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

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

    cs.CV 2026-07 conditional novelty 4.0 of 10

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

  2. UAVs Meet Agentic AI: A Multidomain Survey of Autonomous Aerial Intelligence and Agentic UAVs

    cs.RO 2025-06 conditional novelty 3.0 of 10

    A narrative survey defines 'agentic UAVs' as drones with perception, cognition, control, and communication layers and catalogs applications and challenges across eight domains.

Reference graph

Works this paper leans on

19 extracted references · 16 canonical work pages · cited by 2 Pith papers

  1. [1]

    Completeness - Does the text fully present all the necessary reasoning steps or information needed to answer the question or solve the problem? - Are there any missing steps, skipped logic, or gaps in reasoning?

    work page

  2. [2]

    Coherence - Is the reasoning internally consistent and logically connected? Are there any contradictions or abrupt jumps?

    work page

  3. [3]

    landmark_bbox

    Fluency - Is the language natural and smooth? Are grammar, sentence structure, and word choice appropriate? Scoring Guideline: - 1: Very poor - 2: Poor - 3: Fair - 4: Good - 5: Excellent Based on the above definitions and scoring scale, please evaluate the following text. Output one integer score (1–5) per criterion. Strictly follow the format below so it...

    work page

  4. [4]

    landmark_bbox

    Bounding Box for the Landmark Region:– Based on the aerial view, the northern edge of Chesterwood Gardens is highlighted in red. The area around the edge of the buildings and the sidewalk here aligns with the target description. {"landmark_bbox": [2693, 16, 4026, 1361]} 2 The target, a black car parked between a silver car and a black car, is located in t...

    work page

  5. [5]

    landmark_bbox

    Bounding box of the area: – Since we need to look behind the intersection at the parking lot, and considering the red masks as a guide, I’ll assume an area directly behind the intersection is our target. This means the parking lot would lie inside the yellow bounding box (highlighted on the image). The bounding box region of interest behind the red-masked...

    work page

  6. [8]

    Chesterwood Gardens

    Semantic Interpretation of the Instruction: – The car is described as a black car and is on the sidewalk near a building. – The location reference “Chesterwood Gardens” should point the drone towards the correct residential area

    work page

  7. [9]

    Given it is residential with houses and likely the target is near the edge of one of these buildings on the sidewalk, I would focus on the northern edge of Chesterwood Gardens

    Identifying the Correct Landmark Region: – The aerial view shows Chesterwood Gardens as a rectangular area marked by the red mask. Given it is residential with houses and likely the target is near the edge of one of these buildings on the sidewalk, I would focus on the northern edge of Chesterwood Gardens

    work page

  8. [10]

    The sidewalk along the northern edge of Chesterwood Gardens is a likely location

    Relative Position of the Target: – The target is described as on the sidewalk next to a building. The sidewalk along the northern edge of Chesterwood Gardens is a likely location

    work page

Show all 19 references
  1. [12]

    One Stop,

    Identify Possible Red-Masked Landmarks: – The image highlights red structures with some marked as "One Stop," which appears larger and more prominent. – The area behind "One Stop" is clearly outlined by the red mask

  2. [13]

    One Stop

    Locate the Target Relatively: – The target is described as being behind "One Stop." – Given the current field of view and the position indicated by the arrow, the target should be within the vicinity of the "One Stop" landmark

  3. [14]

    One Stop

    Bounding Box for Landmark: – The "One Stop" landmark is clearly marked by the red structure. – The bounding box for the landmark can be identified roughly around the "One Stop" structure. {"landmark_bbox": [1594, 0, 3864, 1502]} 3 The target description indicates that the targ...

  4. [15]

    These identify the streets and regions of interest

    Identify the relevant landmark red masks: – There are red masks over Chesterton Road and Ferry Path in the image. These identify the streets and regions of interest

  5. [16]

    It points towards the top-left of the image

    Determine the direction of movement: – The UA V’s current orientation is indicated by the arrow at the bottom-left. It points towards the top-left of the image

  6. [17]

    Locate the intersection: – The intersection of Chesterton Road and Ferry Path is where the red masks overlap at the bottom of the image

  7. [18]

    Given the red masks, it’s reasonable to conclude that any area behind these masks, specifically in the region near the intersection, is the location to look for

    Identify the target area: – The target parking lot is described as being behind the buildings at the corner. Given the red masks, it’s reasonable to conclude that any area behind these masks, specifically in the region near the intersection, is the location to look for

  8. [2019]

    10 OpenAI, :, Aaron Hurst, Adam Lerer, Adam P

    Self-monitoring navigation agent via auxil- iary progress estimation.arXiv preprint. 10 OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander M ˛ adry, Alex Baker-Whitcom...

    arXiv

  9. [2023]

    target_location

    Learning generalizable manipulation poli- cies with object-centric 3d representations.Preprint, arXiv:2310.14386. 15 A Prompt Template for FlightGPT Prompt System Message: You are an intelligent autonomous aerial vehicle (UA V) capable of real-world navigation and visual targe...

    arXiv

  10. [2024]

    InProceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (V olume 3: System Demonstra- tions), Bangkok, Thailand

    Llamafactory: Unified efficient fine-tuning of 100+ language models. InProceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (V olume 3: System Demonstra- tions), Bangkok, Thailand. Association for Computa- tional Linguistics. Yifeng Zhu, Zh...

  11. [2025]

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao

    Uav-vla: Vision-language-action system for large scale aerial mission generation.arXiv preprint arXiv:2501.05014. Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. 2...

    2025 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.