acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Aligned Rewards For Aerial Vision-And-Language Navigation

Abstract

Aerial vision-and-language navigation (AVLN) requires an agent to localize a language-described goal on a city-scale aerial map by grounding visual landmarks and reasoning over their spatial relations. Recent vision-language model (VLM)- based methods formulate goal localization as coordinate generation and optimize the policy with reinforcement learning. Their composite rewards typically cal- ibrate goal accuracy with globally fixed distance scales, limiting discrimination beyond a cutoff and ignoring variation in landmark extent. To address this is- sue, we propose GS-Nav, a geometry-aligned learning framework for AVLN. Its core is a Scale-Adaptive Gaussian Reward, which converts the distance to the target into a dense preference signal and adjusts its bandwidth according to the spatial extent of the referred landmark. In addition, we identify a simple but ef- fective resolution-decoupling strategy: training with high-resolution maps while performing inference at a lower resolution. On the CityNav benchmark, GS-Nav obtains the best reported SR, OSR, SPL, and mean NE point estimates among the listed learned methods on validation unseen. Meanwhile, resolution-decoupled inference uses approximately 3.1×fewer visual tokens and reduces median query latency by about 4.2×. Controlled reward comparisons support geometry-aligned supervision within CityNav; resolution controls reproduce the inference trend at two scales of the Qwen2.5-VL family. Code and models will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Aerial Vision Language Navigation, Reinforcement Learning, Gaussian Reward, Landmark Grounding, Goal Localization, CityNav

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_WaVqOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GS-Nav addresses language-conditioned destination localization on aerial maps. A Qwen2.5-VL policy predicts a landmark box and destination coordinate, which is executed by a fixed controller. GRPO training combines a destination-centered Gaussian reward with landmark-dependent bandwidths, Gaussian landmark coverage, and format supervision. The paper also studies 4k training with 2k inference. On CityNav, it reports 20.56% validation-unseen SR and 23.82% test-unseen SR. Under matched inference conditions, the complete reward improves validation-unseen SR from 19.64% to 20.56%.

Critical strengths
  1. The paper addresses a practically relevant problem. Reliable language-guided aerial navigation requires connecting linguistic descriptions with spatial structure. Studying geometric supervision for destination localization is a meaningful direction toward improving this capability.

  2. The approach offers a simple and potentially practical way to improve an existing navigation pipeline. It combines changes to training supervision and inference resolution while retaining the underlying policy–controller interface, making the ideas relatively straightforward to incorporate into comparable systems.

Critical weaknesses
  • Limited methodological novelty. GUI-G² already combines size-adaptive Gaussian point rewards, Bhattacharyya coverage, and GRPO. The main changes are relocating the reward center and introducing candidate-wise maxima.

  • Unjustified destination tolerance. CityNav evaluates success within a fixed 20 m radius, but the reward scales tolerance using a reference landmark’s dimensions—even when the destination lies outside that landmark. This can assign higher rewards to unsuccessful predictions than to successful ones.

  • Candidate aggregation undermines landmark-specific calibration. In Eq. (3), a candidate with greater width and height dominates a smaller candidate everywhere. Independent point and coverage maxima also allow destination tolerance to come from a different landmark than the one predicted by the policy.

  • Weak incremental benefit from 4k training. At identical 2k inference, Table 3 shows only 20.28%→20.56% SR and 18.82→18.95 SPL, although mean NE improves by 3.99 m. The large accuracy and latency gains primarily reflect replacing 4k inference with 2k inference.

  • Evaluation restricted to an outdated model family. All experiments use Qwen2.5-VL.

  • Incomplete performance comparison. SA-GCS reports higher test-unseen SR/SPL—24.55/22.86 versus 23.82/21.78—but is omitted. Protocol differences require clarification before direct comparison, leaving the empirical advance inadequately contextualized.

  • No reported physical-flight validation. The supplied main paper reports CityNav trajectories and H20 query latency. Neither establishes physical UAV operation or end-to-end deployment performance.

Questions for authors
  • What bandwidth and reward components are used for the fixed-Gaussian baseline in Table 2? Which reported comparison establishes the benefit of landmark-dependent calibration separately from landmark supervision?

  • Do all Table 3 rows use the same evaluation episodes? Please clarify SR aggregation and treatment of invalid outputs, particularly how 10.53% is obtained if using a single run over 2,826 episodes.

  • What annotation, input, training-budget, or evaluation-protocol differences prevent comparison with SA-GCS, which reports higher test-unseen SR and SPL?

  • Does the submitted supplementary material contain physical UAV evaluation? If so, where are the platform, protocol, and results reported?

Reviewer_ahoYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies geometric reward design for VLM-based aerial vision-and-language navigation. It proposes GS-Nav, which combines a landmark-scale-adaptive Gaussian reward for destination coordinates, a Gaussian similarity reward for intermediate landmark grounding, and an output-format reward. The destination reward is centered at the annotated target, while its anisotropic bandwidths are determined by candidate landmark dimensions. The paper also investigates resolution-decoupled inference, using 4k maps for training and 2k maps for inference. Experiments on CityNav show competitive unseen-environment performance and substantial inference savings.

Critical strengths

The paper addresses a concrete weakness in coordinate-based AVLN supervision. A fixed distance cutoff can remove relative goal feedback when all sampled predictions fall outside its effective range. The proposed reward offers a simple way to incorporate landmark geometry into both destination supervision and intermediate grounding while retaining the existing coordinate-to-controller interface.

Critical weaknesses
  1. The paper does not show that its main contribution, the scale-adaptive reward, actually matters. The gains over simpler rewards are small and come from single runs, so they cannot be told apart from noise. The adaptive bandwidth is never tested on its own. Its design also sits uneasily with the task: it makes the reward more lenient around large landmarks, while success is still judged with a fixed radius. The paper does not show that this reward tracks navigation success better than a fixed Gaussian or the FlightGPT verifier. As a result, the claim that geometry-aligned supervision improves AVLN is not established.
  2. The headline results mostly reflect a change in inference protocol, not the proposed reward. The main comparison puts GS-Nav with 4k→2k inference next to prior methods evaluated at matched resolution. Under a shared protocol, most of the gap disappears. The resolution finding itself comes down to "4k inference degrades this model". It is shown for one model family, has no explanation, and looks more like a Qwen2.5-VL input-processing issue than a general training principle. This leaves the paper without a contribution that is both well supported and general.
Questions for authors
  1. The step from the Fixed Gaussian to GS-Nav changes both the bandwidth and the landmark term. How do adaptive point reward without the coverage term, and fixed point reward with Gaussian coverage, perform?
  2. All candidate Gaussians in Eq. (3) share the center p∗p^*, so the maximum always selects the widest bandwidth on each axis. How is this different from simply using the largest annotated box? On held-out rollouts, how well does each reward in Table 2 correlate with 20 m success?
  3. σ\sigma is defined in pixels of the 4k map. For typical CityNav landmarks, what does it correspond to in meters relative to the 20 m success radius, and how sensitive are the results to α\alpha?
  4. The motivation is that the truncated verifier gives no relative signal when a whole GRPO group lands beyond 80 m. How often does this happen during FlightGPT-composite training, and how does that fraction change under your reward?
  5. At which resolution is FlightGPT-RL in Table 1 trained and evaluated? Under the shared 4k→2k protocol in Table 2, the FlightGPT composite already reaches 19.64 SR. How does GS-Nav compare with FlightGPT under FlightGPT's own matched-resolution setting?
  6. The 4k setting yields 16,384 visual tokens, exactly Qwen2.5-VL's default max_pixels cap. Was the same processor setting used in training and at 4k inference? If so, why does 4k→4k (10.53 SR) perform much worse than 4k→2k (20.56)?
Reviewer_4XYpOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GS-Nav to address a problem of the existing method FlightGPT. In FlightGPT, the goal reward is truncated at a fixed distance threshold, so it cannot distinguish between predictions that are far from the target, and it also does not adapt to the size of landmarks. GS-Nav consists of two components. The first is the Scale-Adaptive Gaussian Reward, which evaluates the target position with an anisotropic Gaussian distribution based on the landmark size, and evaluates the predicted landmark region by its overlap with the ground-truth region. The second is Resolution-Decoupled Inference, which uses 4k maps for training and 2k maps for inference. The authors report that GS-Nav reduces the number of tokens and the latency at inference compared with 4k inference, and that it outperforms FlightGPT in the unseen environments of CityNav.

Critical strengths

The experiments on CityNav show that the model trained with 4k maps achieves higher accuracy when it is used with 2k maps at inference, and that this also reduces the number of tokens and the latency at inference.

Critical weaknesses
  • The overall framework and the evaluation protocol of this work follow the previous method, FlightGPT, and the initial weights also seem to be a reproduction of the SFT model of FlightGPT. The reward design is also almost a direct transfer of GUI-G^2. The differences are limited to moving the center of the Gaussian to the target position and taking the max over candidates. Therefore, the novelty is limited.
  • The motivation of the proposed method and the evidence for its effectiveness are not sufficiently verified. In the previous method, predictions beyond the threshold receive the same score, and when all predictions in a group are out of range, the relative signal of the goal term is lost. However, the paper does not measure how often this problem actually occurs, so the need to change the reward is unclear. In addition, the ablation study changes several components at the same time. Therefore, the contribution of scale adaptation alone, which is the core of the claim, is not separated.
Questions for authors
  • Please clearly explain the technical contribution that is specific to this work, compared with FlightGPT and GUI-G^2.
  • Please provide quantitative evidence for the motivation of the proposed method.
  • For FlightGPT and GS-Nav, please show the contributions of the reward and the inference resolution separately.
1 comment

Thank you for your response.

Regarding Q1, although the differences from existing methods have been explained, the technical novelty beyond engineering efforts is still unclear to me.

Regarding Q2, the additional experiments appropriately address my concern.

On the other hand, regarding Q3, I do not think the experimental results sufficiently support the effectiveness of the proposed method.

Overall, my main concerns, particularly regarding the technical novelty of the method, remain unresolved. Thus, I will keep my rating unchanged.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.