acceptodds
Withdrawn from ICLR 2027

Last-Meter Precision Navigation for UAVs: A Diffusion-Refined Aerial Visual Servoing Approach

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In this work, we study the last-meter precision navigation for UAVs, e.g., autonomously reaching a target within the final 10 meters using monocular vision. This task is challenging due to scale ambiguity, rotation discontinuities, and the need for fine-grained spatial reasoning. Existing methods often fail under large viewpoint changes or lack generalization to unseen environments. To this end, we propose DreamNav, a coarse-to-fine diffusion-refined aerial visual servoing framework. In the first coarse-estimation stage, a robust regression policy employs a trigonometric parameterization to predict rotation by jointly modeling sine and cosine components, effectively mitigating optimization instabilities caused by angular periodicity. Given this coarse estimate, the second diffusion-refined stage utilizes a pre-trained world model to simulate future visual observations for candidate actions, selecting the trajectory that minimizes visual discrepancy with the target through a process of visual imagination. To support rigorous evaluation, we contribute PairUAV, a large-scale benchmark comprising 4.8 million image pairs across 72 scenes, curated from the University-1652 dataset. Extensive experiments show DreamNav outperforms strong visual servoing and foundation model baselines in accuracy and generalization, with zero-shot transfer to unseen scenes.

Primary area:
applications to robotics, autonomy, planning
Keywords:
UAV navigation, visual servoing, diffusion models, image-goal navigation, relative pose estimation, aerial robotics

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

This paper studies last-meter UAV navigation using a source image and a target image. The authors introduce PairUAV, a benchmark containing 4.8 million image pairs constructed from the University-1652 dataset, and propose DreamNav, a two-stage coarse-to-fine framework.

In Stage I, the model combines correspondence-based geometric cues with visual features to regress relative heading and radial-range changes, using trigonometric regression to handle angular periodicity. In Stage II, a pose-conditioned diffusion model synthesizes observations for candidate actions around the coarse prediction and selects the candidate whose generated image best matches the target.

Experiments on 39 unseen scenes show that diffusion refinement improves navigation success across two visual backbones. With MASt3R, the success rate increases from 67.70% to 78.86%, outperforming the strongest reported baseline, π0.5 (70.66%).

Critical strengths
  1. Visual servoing for terminal UAV guidance is an interesting and practically relevant direction, especially in scenarios where coarse GNSS localization is insufficient for precise target alignment.

  2. The authors train on 33 scenes and evaluate on 39 held-out scenes, which is a reasonable experimental design for assessing scene-level generalization.

Critical weaknesses
  1. The task formulation is overly constrained and does not sufficiently support the claimed last-meter precision navigation capability. The proposed task is highly constrained: the camera pitch is fixed at 45°, both views must face the same reference center, and the model only predicts heading and radial-range changes rather than general 3D displacement. I find the practical motivation for these assumptions unclear, and even Figure 1 does not clearly illustrate a scenario satisfying them. Moreover, success is defined within a 10 m threshold, which seems insufficient to demonstrate the claimed last-meter precision.

  2. Although Stage II improves performance, the paper does not compare diffusion refinement against simpler alternatives, such as feature-based matching or geometric optimization. It is therefore unclear whether the improvement comes specifically from diffusion-based visual imagination or simply from local candidate search.

Questions for authors
  1. Could the authors provide a concrete real-world UAV navigation scenario that motivates the fixed 45° camera pitch, shared reference center, and 2-DoF action space? How does the motivating example in Figure 1 satisfy these assumptions?

  2. How does the proposed diffusion-based refinement compare with simpler non-generative approaches, such as feature-based matching or geometric optimization over the same candidate action space?

  3. PairUAV is built upon University-1652, which uses Google Earth imagery. Given Google's potential restrictions on using satellite imagery for machine learning and University-1652's restrictions on redistributing derivative datasets, could the authors clarify whether the training, evaluation, and release of PairUAV comply with the applicable licensing terms?

Reviewer_RjKGOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper identifies a potentially unexplored problem, which they call the last-meter navigation for UAVs: flying autonomous drones at the end of their destination may impose extra challenges (like the interference of GNSS signals and higher navigation performance requirements), so it should be studied as a separate problem. Under the lack of appropriate datasets, they propose a new one, called PairUAV, by sampling images pairs on an urban environment. Then they propose DreamNav, which is a world-model based method for their problem. DreamNav uses a coarse pose regressor that estimates 2D heading and range from a pair of current and goal image. A second step in the method uses a diffusion model to generate pose-conditioned images and selects the image (and associated pose) that best matches the goal image. These methods are then tested on PairUAV and are reported to perform better than others in terms of pose prediction performance.

I think that there are some interesting and potentially valuable contributions in the paper, but the submission appears to lack sufficient familiarity with aerial navigation and some related tasks, which causes me to have many fundamental concerns about its content. Even if the idea of last-meter aerial navigation is an interesting one, I am not sure if it's indeed what the paper is tackling. In my view, what it studies would be more accurately described as "elevated relative 2D pose estimation". Though this setting could also be well motivated. it doesn't seem to (as it was presented) reasonably describe the challenges and limitations that existing methods would face in this last-meter aerial navigation scenario and in general the submission fails to properly contextualize their work and connect it to previous methods. Consequently, the choice of baselines for comparison appears questionable and incomplete, while the proposed benchmark doesn't seem consistent with the paper's own description of the problem.

I thus believe that the submission could be turned into a valuable contribution, but that would require extensive revisions and reframing, so currently I recommend its rejection.

Critical strengths

The proposed DreamNav method appears to be very reasonable and lightweight, which seems to have translated into small training times and small computational requirements. I think that's important in the context of world models, where other methods can be expensive to train and limit accessibility. Tough that seems to be a result of using well established techniques like LoRa's and pre-trained components instead of any novel idea, that does not invalidate this strength in my perception. The benchmark task is also formulated in a simple way and I believe that decomposing a larger problem into more manageable subproblems brings similar benefits and facilitates research in general. Even if in my perception that only resulted from a misunderstanding of the issue the paper was trying to tackle, I do think there's value in simplifications and perhaps a reasonable one could be properly made and justified without too much effort.

I also think that the idea of last-meter navigation as a problem (or at least some version of it) is indeed an interesting setting that has obvious practical motivation.

Critical weaknesses

In my perception, the main weaknesses is that problem they are trying to tackle doesn't seem to be formulated correctly: the general idea and motivation related to package delivery seems relevant and of practical usage, but how that is contextualized and translated into a task (as well as which methods are selected for comparison) appears inadequate.

To start, the authors only consider GNSS as a way of aerial navigation and fail to mention the extensive literature on methods such Visual-Inertial Odometry (VIO) or Simultaneous Localization and Mapping (SLAM). A significant motivation of these works is to be able operate in the absence of GNSS. To quickly name a few open-source options: Orb-SLAM, Kimera, OpenVINS, VINS-Mono, ROVIOLI. In general, the main benefit of GNSS is to prevent drift accumulation over large scales, so to claim reliance on them at short distance appears to be an inversion of reality. (though there are methods to enhance the precision of GPS, I think it's fair to consider them out of scope)

Take for example VINS-MONO [1], which is almost a decade old by this point. They report about 1% drift without loop closure, which means that that they could fly for 1km before their position estimate drifted by 10m (the threshold for success considered by the authors of this submission). In other words, these were already very capable of keeping their position estimate error low on the absence of any GNSS system or even a map. Following the paper's motivation, It would seem unlikely that GNSS (assuming it to be necessary) would become unreliable or unavailable from such a large distance due to the proximity to buildings or other urban reasons. One could imagine that many issues could diminish the practicality of these and other methods in the last meter setting (e.g., performance degradation, hardware cost) and there's definitely a case to be made for world models and related ideas, but these topics are never considered. Maybe the true issue is the fact that the goal position is not known accurately and only an image is available, but that is a different problem and appears poorly connected to GNSS availability without proper explanation (i.e., one problem is of locating the vehicle and the other is of locating the goal).

Nevertheless, is the paper really tackling the last meter aspect of navigation? The first half of the paper suggests that it is concerned with what happens close to the goal (around 10m) which seems to be an acceptable threshold where long-scale navigation turns into short-scale navigation in urban scenarios. This is supported both in text and figures. Yet, later in the experiments they consider the opposite of problem of being within 10m, since that's the success criteria. Granted, they also use the MAE as a metric and thus quantify precision beyond that threshold, but overall the problem seem to be that of the last 50 or so meters instead. It's not clear if the dataset would even be theoretically capable of providing meter-level precision since that would require enough pixel disparity and the views seem to be directed at distant regions. Not to mention the fact that navigation to such target image would have the monocular vehicle pointing a way from where it is moving, making it unable to perceive the building or other nearby obstacles, which seems hazardous. I will also comment that figure 2 highlights vertical position uncertainty as a key concern, but the formulation seems entirely horizontal and thus unable to capture that aspect.

Moreover, is the paper really dealing with aerial navigation if it's entirely constrained to a planar region without even pitching? Though I accept that there are situations where that could be a reasonable approximation (e.g., a plane in cruise), that doesn't seem representative of how quadcopters and similar vehicles operate. Perhaps that could help extend the applicability of 2D world models to a more complex aerial setting, but that connection seems to be mostly aesthetic as the paper appears to be essentially a 2D planar one, where the plane of movements happens to be far from the ground. The only value I see in that is in creating a more diverse and potentially harder dataset. This questionable aerial context is then used to justify not making a comparison to other world models when it seems they could be applied (e.g. perhaps RAE-NWM [2] could be applied with a single action), while only presenting unsatisfactory and unclear comparisons (even with a decade old RL method like AI2THOR).

Finally, is the paper really about navigation if there's only a single command (difference in 2D pose) and there are no dynamics and no history of actions or observations? It seems much closer to relative pose estimation or visual place recognition in the sense that these methods could be easily adapted to the problem (e.g., SALAD [3]).

In summary, how the problem is framed, materialized into a task and connected to previous work seems to be unsatisfactory.

References

  • [1] Qin, Tong, Peiliang Li, and Shaojie Shen. "Vins-mono: A robust and versatile monocular visual-inertial state estimator." IEEE transactions on robotics 34.4 (2018): 1004-1020.
  • [2] Zhang, Mingkun, et al. "Rae-nwm: Navigation world model in dense visual representation space." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.
  • [3] Izquierdo, Sergio, and Javier Civera. "Optimal transport aggregation for visual place recognition." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.
Questions for authors

Questions and comments (relevant to the assessment):

  • Have the authors checked the copyright requirements for using Google Earth data and believe that there are no issues present? Their usage seems permissible to me and I don't mean to be nitpicky, but I'm not a legal expert and upon inspection their services seem to request a certain type of attribution (see https://about.google/brand-resource-center/products-and-services/geo-guidelines/#required-attribution).
  • The way the authors describe and handle the pose wrap-around issues indicates a lack of familiarity on the literature of neural pose regression and the relevant sections should probably be revised. If that's the case, I suggest the following starting point: Zhou, Yi, et al. "On the continuity of rotation representations in neural networks." 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2019.
  • Why did the authors use DINO-ResNet-50 and not the DINO-ViT or even the more modern and performant DINOv2 or DINOv3? That seems odd without explanation.
  • Table 2 seems to suggest that the backbone is of critical importance for the success of the method. Would it be possible to create a comparison with other methods using the same backbone?
  • Can the authors provide more information about how the other methods are trained? I ask that because it would seem unfair to compare the training time of DreamNav finetuned with LoRa against π0.5\pi_{0.5} without LoRa or a similar idea (if applicable). It would also be questionable to compare a RL method like AI2THOR against supervised ones (unless perhaps only the backbone was used and trained in a supervised fashion?)

Easily fixable issues, suggestions or comments (not important to assessment):

  • I don't think the authors cited the original paper for MASt3R and nor the original DINO paper (which had the ResNet they use)
  • Lines 451-452 seems to have some placeholder values
  • I'll mention that the text goes slightly beyond the page limit, which technically violates the submission instructions, but I don't take that into consideration for my review and leave it to the AC to decide the appropriate course of action.
  • Table 1 is good, but I think it would help to have some notion of how diverse the other datasets are, even if only in the appendix. For example, PairUAV has 4.8M instances, but from 1652 buildings. Had they used, for example, 10 buildings, then these 4.8M instances would be much less informative and would probably be very redundant.
  • I wonder if operating entirely within the latent space (without decoding back into pixel space) would result in improved performance, as is the case in many problems. I think it would be easy to try as it may only requires encoding the goal image, so I leave it as a suggestion.
Reviewer_zxZeOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents PairUAV, a heading–range prediction benchmark derived from University-1652, and DreamNav, which refines coarse pose predictions by generating views for nearby pose candidates and selecting the one with the lowest image MSE to the target.

Critical strengths
  1. The benchmark repurposes existing UAV imagery for image-goal relative-pose prediction at scale.
  2. The paper compares several types of baselines and reports improvements from refinement with two backbones.
Critical weaknesses
  1. Unclear range and heading annotations. Is rr the horizontal distance to the reference center or the 3D slant range? The paper states that Δr\Delta r is not the actual travel distance, but the underlying range coordinate remains unclear. Is θ\theta the camera’s azimuth around the reference center or its own yaw? Are the ground-truth values taken from rendering metadata, actual UAV flight records, or reconstructed from image ordering?
  2. Diffusion accuracy. The authors state that “Accurately estimating subtle differences in relative pose remains highly challenging.” Yet Stage II requires the generator to be highly sensitive to range and heading, producing clear visual responses even to small changes. If small changes in range or heading produce nearly identical images, how can the candidates be reliably distinguished? Moreover, pixel-wise MSE is sensitive to texture and appearance errors and does not necessarily reflect geometric accuracy. The smooth view changes in 3. Figure 3 show responsiveness to pose conditioning, but do not establish geometric accuracy for subtle pose changes. Necessity of stage 2. In Table 2, the gains from diffusion refinement are limited compared with those from switching to a geometry-aware backbone. This suggests that a purely geometric refinement approach may be more appropriate. Without such a comparison, the use of diffusion is not sufficiently justified.
  3. Typos and inconsistent results. Lines 451–452 contain “XX” placeholders. Lines 454–455 and Figure 4 report SRs of 42.00%/80.00%, inconsistent with 22.40%/78.86% in Table 2 (lines 332–334). Figure 3(a) shows conflicting Stage-II outputs and adds meters to degrees to calculate endpoint error, contradicting the 3D distance definition in lines 308–309.
Questions for authors
  1. What exactly do rr and θ\theta represent, and how are their ground-truth values obtained?
  2. What evidence shows that image MSE can reliably distinguish small changes in range and heading and select the more accurate pose?
  3. Does diffusion outperform simpler geometric refinement using the same coarse predictions?
  4. Can you resolve the “XX” placeholders, inconsistent success rates, and conflicting outputs and endpoint-error calculation in Figure 3(a)?

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

Related papers

Discussion (0)

Sign in to comment.