acceptodds
Under review as a conference paper at ICLR 2027

Beyond Expert Trajectory Matching: Geometry and Risk-Aware Imitation Learning for Safer UAV Vision-Language Navigation

Abstract

UAV vision-language navigation (UAV-VLN) requires UAVs to follow natural-language instructions while avoiding obstacles using visual observations. A common approach to UAV-VLN is imitation learning, where navigation models are trained to mimic actions from expert demonstrations. However, vanilla imitation learning cannot distinguish multiple expert trajectories in terms of task-relevant qualities, and one particularly important quality for UAV navigation is safety. In reality, expert trajectories include ones that remain comfortably in open spaces, and ones that pass dangerously close to obstacles. Regardless of the difficulties in obstacle avoidance, vanilla imitation learning treats these trajectories as equally desirable supervision, which is not helpful for lowering collision rates. In order to address this, we adopt a simple but effective approach named GeoRisk-IL. Specifically, we introduce a collision-risk penalty on predicted actions using the distance between future UAV locations and obstacles. An essential aspect of our design is to also enhance the 3D awareness of the UAV-VLN model. This provides the spatial understanding needed for meaningful risk estimation. Interestingly, the collision-risk penalty becomes more effective when combined with geometric representation alignment, despite the latter providing little to no benefit on its own, highlighting a strong synergy between geometric awareness and risk-aware learning. On both synthetic and real-world datasets, the proposed method demonstrates significant improvements over previous methods, verifying its effectiveness.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language Navigation, Imitation Learning, Geometric Representation Learning, Collision Avoidance, Vision-Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TuyqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GeoRisk-IL augments imitation learning for UAV vision-language navigation with geometric feature alignment to a frozen DA3 model and a clearance penalty on predicted trajectories. Both components are used only during training. Experiments on OpenUAV and HUGE-Bench report improved navigation success and lower collision rates, with ablations showing additional benefits from combining the two objectives.

Critical strengths

The same-backbone comparisons show substantial improvements. On OpenUAV Unseen Map, success rate increases from 23.49% to 44.89%, while collision rate decreases from 63.26% to 38.94%. These results provide meaningful evidence that the added training objectives are useful.

The auxiliary geometry and risk computations are removed at inference. This makes the approach practical without adding these components to the deployed navigation pipeline.

Critical weaknesses

Physical flight validation is missing. The paper does not report closed-loop experiments with a physical drone in real environments. Since the main contribution concerns safer UAV navigation, physical experiments are important for assessing whether the collision reduction transfers under sensing and localization errors, imperfect trajectory tracking, and inference delay. The practical safety benefits remain unvalidated under these conditions.

The methodological novelty appears limited. Geometric feature alignment and a distance-based trajectory penalty are straightforward training objectives. The main added value appears to be their combination and the observed interaction. However, the synergy ablation is restricted to Test Seen, without corresponding validation on unseen maps or HUGE-Bench. Stronger evidence is needed to establish the scope and significance of this combination.

Questions for authors

1: Additional validation: Can you provide a controlled physical-flight comparison against Qwen3-VL Base, reporting success rate, minimum obstacle clearance, and safety interventions? This would directly address the practical safety limitation.

2: Additional validation: Does the interaction between geometric alignment and the clearance penalty persist beyond Test Seen? A corresponding four-way ablation on an unseen split or HUGE-Bench would help answer this.

3: Clarification: Please specify the voxel size, depth history window, and any filtering applied before voxelization. Also explain the L1 assistant rules, including whether interventions alter actions or episode termination and how they affect the reported metrics.

4: Clarification: LongFly’s UO Full NE in Table 1 is 66.74, while Easy and Hard NE are 54.84 and 57.07. If Full averages these subsets, it should lie between their means. Please check the values and evaluation coverage. This concerns a cited baseline and does not directly invalidate the comparison against Qwen3-VL Base.

3 comments

Reviewer Tuyq — Part 1/3

Thank you for recognizing our strengths in 'the substantial improvements in same-backbone comparisons', 'the improved navigation success on unseen maps' and for your constructive feedback.

W1: Physical flight validation is missing. Since the main contribution concerns safer UAV navigation, physical experiments are important for assessing whether the collision reduction transfers under sensing and localization errors, imperfect trajectory tracking, and inference delay.

A: Thank you for pointing out this limitation. We agree that physical-flight validation is important, but outdoor tests are beyond our capabilities at this moment due to regulatory reasons. We apologize for being unable to provide physical flight results within the rebuttal period. Nevertheless, we follow [3-5] and conduct simulated flight experiments that account for multiple conditions encountered in real-world flight. We evaluated Qwen3-VL Base and GeoRisk-IL on OpenUAV Test Unseen Map, mimicking real-world conditions and perturbations in Table R3-1.

Under the perturbations, GeoRisk-IL continues to reduce CR by 24.63 percentage points and increase SR by 18.99 percentage points over Base. These results show that its collision-reduction and task-completion benefits persist under joint sensing and execution perturbations.

We will add this and are committed to adding real-world flight experiments in the future.

Table R3-1. Flight simulation settings to mimic real-world noise and perturbations.

FactorSetting
Motion limitsHorizontal control-reference limits: acceleration 2.02.0, braking 3.6 m/s²; ascent 33, descent 2 m/s; yaw rate 30∘/s30^\circ/\mathrm{s}.
Continuous flight and inertiaVelocity carries over between action segments during continuous trajectory tracking.
Observation-to-execution delayA delay of 0.5 s follows each observation, during which the UAV continues flying.
State-estimation errorsRandom position errors: horizontal within 0.30 m and vertical within 0.10 m; yaw error up to 2∘2^\circ.
Image degradationAdditive Gaussian noise on RGB observations with standard deviation 2/2552/255.

Table R3-2. OpenUAV Test Unseen Map results under default and perturbed conditions.

MethodPerturbationsCR ↓SR ↑OSR ↑
Qwen3-VL Base×\times63.2623.4927.77
GeoRisk-IL×\times38.9444.8950.52
Qwen3-VL Base✓\checkmark63.6722.7627.77
GeoRisk-IL✓\checkmark39.0441.7545.62

[3] E. Kaufmann et al., “Champion-level drone racing using deep reinforcement learning,” Nature, vol. 620, no. 7976, pp. 982–987, 2023, doi: 10.1038/s41586-023-06419-4.

[4] S.-A. Yu et al., “FlightBench: Benchmarking learning-based methods for ego-vision-based quadrotors navigation,” IEEE Robotics and Automation Letters, vol. 10, no. 7, pp. 6888–6895, 2025, doi: 10.1109/LRA.2025.3573167.

[5] G. Sui et al., “ProtoVis-Nav: Prototypical visual imagination for visual-spatial aligned UAV navigation,” 2026, arXiv:2611.17330.

Reviewer Tuyq — Part 2/3

W2.1: The methodological novelty appears limited. Geometric feature alignment and a distance-based trajectory penalty are straightforward training objectives. The main added value appears to be their combination and the observed interaction.

A: We feel there are some misunderstandings. Our focus is not on ‘a new formulation’ or ‘a new architecture’. Instead, our contribution lies in:

  1. Identifying imitation learning (IL)’s potential issue of treating all GT trajectories uniformly. A collision-free expert trajectory may still pass close to obstacles, leaving little tolerance for prediction errors. Mimicking the GT trajectories alone cannot distinguish such a demonstration from one with a larger safety margin, and we need other paradigms to help to lower the crash rate.

  2. A simple-yet-effective collision-penalty-based extension to vanilla IL to counteract this issue. We are very delighted to see that all reviewers agree on the effectiveness of the proposed method: “the added training objectives are useful” (your assessment), “the method appears effective” (Reviewer nvy5), and “simple and modular” (Reviewer 6jGt). Since this design can greatly alleviate the collision issue, we find it worth sharing with the community. We purposefully choose its simplest form and present it without any bells and whistles to highlight the effectiveness of this design.

  3. Identifying the synergy between the collision penalty and the geometric supervision. In our experiment (see Table 3 of the main paper), as opposed to the gains reported in SpatialFly [1], we find that simply adding geometric signals hardly brings any improvement on TravelUAV dataset: SR changes from 64.74% to 64.46%, while CR remains 29.90%. However, when paired with the collision penalty, geometric supervision increases SR from 71.30% to 77.64% and reduces CR from 21.44% to 17.00%. This shows that the benefit of geometric supervision depends on the presence of an explicit risk-learning objective.

We thank you for noticing this synergy between collision penalty and geometry supervision. However, we want to clarify that it does not cover everything we find worth sharing with the community in this paper.

We will clarify this.

References

[1] W. Jiang et al., “SpatialFly: Geometry-guided representation alignment for UAV vision-and-language navigation in urban environments,” 2026, arXiv:2603.21046.

W2.2: However, the synergy ablation is restricted to Test Seen, without corresponding validation on unseen maps or HUGE-Bench. Stronger evidence is needed to establish the scope and significance of this combination.

A: To test whether the collision penalty strengthens the benefit of geometric alignment beyond Test Seen, we added four-way ablations on OpenUAV Test Unseen Object and Test Unseen Map.

On Test Unseen Object, adding geometric alignment to Base increases SR by 0.96 percentage points, whereas adding it to Penalty only increases SR by 3.66 percentage points. On Test Unseen Map, the corresponding gains are 1.56 and 8.15 percentage points. Thus, on both unseen splits, geometric alignment yields larger SR gains when the collision penalty is present.

Table R3-3. Four-way ablation on OpenUAV Test Unseen Object (UO) and Test Unseen Map (UM).

SplitVariantGeometryCollisionCR ↓NE ↓SR ↑OSR ↑SPL ↑
UOQwen3-VL Base×\times×\times25.4438.2769.6378.7065.00
UOGeometry only✓\checkmark×\times26.2341.7470.5976.9565.84
UOPenalty only×\times✓\checkmark19.2433.3375.9982.8370.30
UOGeoRisk-IL✓\checkmark✓\checkmark15.9031.2379.6585.6973.76
UMQwen3-VL Base×\times×\times63.26108.9623.4927.7721.70
UMGeometry only✓\checkmark×\times60.44105.6525.0527.7722.91
UMPenalty only×\times✓\checkmark41.7590.4336.7440.9233.01
UMGeoRisk-IL✓\checkmark✓\checkmark38.9478.2444.8950.5239.59

Reviewer Tuyq — Part 3/3

Q1: Additional validation: Can you provide a controlled physical-flight comparison against Qwen3-VL Base, reporting success rate, minimum obstacle clearance, and safety interventions?

A: Please see W1.

Q2: Additional validation: Does the interaction between geometric alignment and the clearance penalty persist beyond Test Seen?

A: Please see W2.2.

Q3: Clarification: Please specify the voxel size, depth history window, and any filtering applied before voxelization. Also explain the L1 assistant rules, including whether interventions alter actions or episode termination and how they affect the reported metrics.

A: We construct training-time obstacles on OpenUAV and HUGE-Bench using h=100/255≈0.392h=100/255\approx0.392 m voxels and depth from the current and two preceding observations. Each non-overlapping 8×88\times8 window supplies its minimum depth and corresponding pixel location. Indices round⁡(p/h)\operatorname{round}(p/h) are deduplicated; occupied voxel centers form the obstacle set.

L1 provides one textual hint at each OpenUAV decision using the expert trajectory and recent UAV motion. Its main rules prioritize ascent and descent guidance: take off when the nearby reference point on the expert trajectory is more than 3 m above the UAV, and landing when the nearby reference point on the expert trajectory is more than 7 m below or the UAV is within 10 m horizontally of the expert trajectory’s endpoint. For the remaining decisions, an angular deviation greater than 20∘20^\circ between recent motion and the direction toward the nearby reference point on the expert trajectory produces left or right; otherwise, the hint is cruise. These hints neither replace predicted trajectories with expert actions nor directly terminate episodes, but they guide the model’s trajectory predictions through the input prompt.

Episodes end when the learned stop head triggers (Appendix A.4), a collision occurs, or the limit of 200 decisions per episode is reached. Base and GeoRisk-IL used identical L1 assistance. We also evaluated both models without any assistance; please refer to Table R1-1 (Reviewer 6jGt W1).

Q4: Clarification: LongFly’s UO Full NE in Table 1 is 66.74, while Easy and Hard NE are 54.84 and 57.07. If Full averages these subsets, it should lie between their means. Please check the values and evaluation coverage. This concerns a cited baseline and does not directly invalidate the comparison against Qwen3-VL Base.

A: Thank you for pointing this out. The LongFly results in our Table 1 were taken directly from Table III of the original paper [2]. We re-checked the source and confirmed that our UO NE entries exactly match the values reported in the original paper. However, we believe this inconsistency among the reported NE values does not affect our main conclusions. We will also contact the LongFly authors for clarification.

References

[2] W. Jiang et al., “LongFly: Long-horizon UAV vision-and-language navigation with spatiotemporal context integration,” 2025, arXiv:2512.22010.

Reviewer_nvy5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes GeoRisk-IL, a training framework for UAV vision-language navigation that goes beyond pure expert trajectory imitation. Built on Qwen3-VL-2B with LoRA, it adds two training-only objectives: a collision-risk penalty that penalizes predicted trajectories with insufficient clearance from obstacles reconstructed from depth observations, and a geometric representation alignment loss that aligns the VLM’s visual-token features with those of a frozen Depth Anything 3 model.

Critical strengths
  1. The paper is well-motivated in both its problem formulation and method design. It clearly identifies that expert demonstrations vary in their safety margins and that vanilla imitation learning cannot account for this, and the proposed collision-risk penalty directly targets this issue.
  2. Though simple, the method appears effective. Both auxiliary objectives are used only during training and add no inference-time overhead, yet they substantially improve success rates and reduce collision rates over the base model.
Critical weaknesses
  1. The experimental settings of the compared methods are unclear, which raises concerns about fair comparison. The baseline results are taken directly from the original papers, and it is not specified whether these methods use different MLLM backbones, input views, or assistant settings. Notably, the Qwen3-VL Base already far outperforms all prior methods (e.g., 64.74% vs. 35.83% SR on Test Seen), which suggests that much of the advantage over prior work comes from the backbone and training setup rather than the proposed objectives.
  2. HUGE-Bench is very small, so the results there are not statistically significant and cannot clearly support the claims. Judging from the reported percentages, the test set contains only about 83 episodes, so the SR improvement over the base model (40.96% to 44.58%) corresponds to only about 3 episodes. No variance across runs is reported.
  3. The novelty is limited. Distance-based collision costs are common in trajectory optimization, and aligning navigation representations with geometry foundation model features has been explored in prior work such as SpatialFly. The main contribution lies in combining these two components within an imitation learning framework.
Questions for authors

Please see the weaknesses above.

2 comments

Reviewer nvy5 — Part 1/2

Thank you for recognizing our strengths in 'the simplicity and effectiveness of our method', 'the clear motivation behind our method design' and for your constructive feedback.

W1.1: The experimental settings of the compared methods are unclear; it is not specified whether these methods use different backbones, input views, or assistant settings.

A: Thank you for pointing this out. For detailed experimental settings, Table R2-1 summarizes the backbones, camera counts and directions, assistant settings, and the Test Seen performance of all compared methods. The results for prior methods are quoted directly from their original papers.

GeoRisk-IL uses Qwen3-VL-2B, the smallest model in the Qwen3-VL series, while prior VLM-based methods use 3B/7B backbones (Table R2-1). Notably, a very recent work ProtoVis-Nav (NeurIPS '26) [1] also uses Qwen3-VL-2B as its navigation backbone.

We will clarify this.

Table R2-1. Comparison settings and OpenUAV Test Seen (Full) performance (%)

MethodBackbone#camerasCamera config.AssistantSR ↑OSR ↑SPL ↑
Random ActionNone0NoneN/A0.140.210.07
CMABi-LSTMNot verifiedNot verifiedL18.3718.727.90
TravelUAVVicuna-7B5front, rear, left, right, downL116.1044.2614.30
NavFoMQwen2-7B4front, rear, left, rightL129.1749.2425.03
LongFlyQwen2.5-3B5front, rear, left, right, downL136.3965.8731.07
AirForesightVicuna-7B5front, rear, left, right, downL135.8369.2530.22
GeoRisk-IL (Ours)Qwen3-VL-2B2front, downL177.6480.4769.87

References

[1] G. Sui et al., “ProtoVis-Nav: Prototypical visual imagination for visual-spatial aligned UAV navigation,” 2026, arXiv:2611.17330. NeurIPS '26 conference page.

W1.2: The Qwen3-VL Base already far outperforms all prior methods, which suggests that much of the advantage over prior work comes from the backbone and training setup rather than the proposed objectives.

A: Thank you for this question. During rebuttal, we add a new experiment to analyze the performance improvements in our baseline design.

  1. We follow AeroVLA's implementation and correct a BGR/RGB mismatch bug in OpenUAV's evaluation code. We regarded this correction as a bug fix and therefore did not discuss it in the initial submission; we will add the relevant explanation in the revised manuscript.

  2. We then replace GroundingDINO-based stopping, as used in TravelUAV [2] and AirForesight [3], with a learned stop head (Appendix A.4).

In Table R2-2, we examine the improvement from these two designs in our baseline. The variant using GroundingDINO-based stopping without the BGR/RGB fix achieves 33.85% SR. Correcting the BGR/RGB mismatch increases SR from 33.85% to 39.07%. Replacing GroundingDINO-based stopping with the learned stop head further increases SR to 64.74%. Moreover, GeoRisk-IL improves all metrics over Qwen3-VL Base in every setting. Even with GroundingDINO stop and BGR/RGB mismatch, its SR, OSR, and SPL still exceed all prior methods in Table R2-1.

Table R2-2. Navigation performance on OpenUAV Test Seen under different stopping rules and image-channel settings.

MethodStopping strategyBGR/RGB fixSR ↑OSR ↑SPL ↑
Qwen3-VL BaseGroundingDINOno33.8564.3930.37
GeoRisk-ILGroundingDINOno44.2271.6538.64
Qwen3-VL BaseGroundingDINOyes39.0762.3435.37
GeoRisk-ILGroundingDINOyes52.4770.8745.95
Qwen3-VL BaseStop headyes64.7468.6259.18
GeoRisk-ILStop headyes77.6480.4769.87

References

[2] X. Wang et al., “Towards realistic UAV vision-language navigation: Platform, benchmark, and methodology,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2025.

[3] Y. Liu et al., “AirForesight: Current-to-future spatial map imagination with cross-space planning consistency for UAV-VLN,” in Proc. 34th ACM Int. Conf. Multimedia (MM), 2026, doi: 10.1145/3767308.3836065.

Reviewer nvy5 — Part 2/2

W2: HUGE-Bench is very small, so the results there are not statistically significant and cannot clearly support the claims. Judging from the reported percentages, the test set contains only about 83 episodes, so the SR improvement over the base model (40.96% to 44.58%) corresponds to only about 3 episodes. No variance across runs is reported.

A: We agree that, given the small size of HUGE-Bench, the original single-run results do not establish the statistical significance of the reported improvements. We therefore independently repeated training and evaluation five times for both Qwen3-VL Base and GeoRisk-IL on this benchmark. We report the mean, 95% confidence interval, and two-sided paired tt-test pp-value for each metric.

While mean nDTW improves slightly over Qwen3-VL Base, GeoRisk-IL achieves significantly higher SR and CSPL and significantly lower CR (p<0.05p<0.05).

Table R2-3. HUGE-Bench performance, reported as mean [95% CI] across runs. SR and CR are percentages. The pp-values are from two-sided paired tt-tests (n=5n=5).

MethodnDTW ↑SR ↑CR ↓CSPL ↑
Qwen3-VL Base0.1815 [0.1705, 0.1925]41.20 [39.95, 42.46]26.75 [21.31, 32.18]0.3925 [0.3791, 0.4059]
GeoRisk-IL0.1970 [0.1790, 0.2149]45.54 [42.30, 48.79]20.24 [16.83, 23.65]0.4407 [0.4134, 0.4679]
pp-value0.16340.01450.00430.0249
Significant (p<0.05p<0.05)noyesyesyes

W3: The novelty is limited. Distance-based collision costs are common in trajectory optimization, and aligning navigation representations with geometry foundation model features has been explored in prior work such as SpatialFly. The main contribution lies in combining these two components within an imitation learning framework.

A: We feel this is a misunderstanding. Our focus is not on introducing ‘a new formulation’ or ‘a new architecture’. Instead, our contribution lies in:

  1. Identifying imitation learning (IL)’s potential issue of treating all GT trajectories uniformly. A collision-free expert trajectory may still pass close to obstacles, leaving little tolerance for prediction errors. Mimicking the GT trajectories alone cannot distinguish such a demonstration from one with a larger safety margin, and we need other paradigms to help to lower the crash rate.

  2. A simple-yet-effective collision-penalty-based extension to vanilla IL to counteract this issue. We are very delighted to see that all reviewers agree on the effectiveness of the proposed method: “the method appears effective” (your assessment), “simple and modular” (Reviewer 6jGt), and “the added training objectives are useful” (Reviewer Tuyq). Since this design can greatly alleviate the collision issue, we find it worth sharing with the community. We purposefully choose its simplest form and present it without any bells and whistles to highlight the effectiveness of this design.

  3. Identifying the synergy between the collision penalty and the geometric supervision. In our experiment (see Table 3 of the main paper), as opposed to the gains reported in SpatialFly [4], we find that simply adding geometric signals hardly brings any improvement on TravelUAV dataset: SR changes from 64.74% to 64.46%, while CR remains 29.90%. However, when paired with the collision penalty, geometric supervision increases SR from 71.30% to 77.64% and reduces CR from 21.44% to 17.00%. This shows that the benefit of geometric supervision depends on the presence of an explicit risk-learning objective.

We agree that collision costs and geometric signals have been used in previous works. However, we believe our design differs from them in:

  1. We are among the first to adopt the collision penalty to solve the collision issue in IL-based UAV navigation. Thus, the resulting system does not require any inference-time cost-based trajectory optimization, thereby avoiding the associated computational overhead.

  2. Please see above. Geometric representations alone bring no significant improvements, but their combination with clearance penalties is very helpful.

We will clarify this.

References

[4] W. Jiang et al., “SpatialFly: Geometry-guided representation alignment for UAV vision-and-language navigation in urban environments,” 2026, arXiv:2603.21046.

Reviewer_6jGtOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper mainly addresses the risky actions in UAV vision-language navigation. The proposed method GeoRisk-IL combines expert trajectory supervision with an obstacle-clearance penalty and visual-feature alignment to a frozen Depth Anything 3 teacher. Both auxiliary objectives are used only during training. Results on OpenUAV and HUGE-Bench show higher success rates and fewer collisions than the same-backbone baseline. The ablation shows little benefit from geometry alignment alone, but a larger benefit when combined with the risk penalty.

Critical strengths
  • The method is simple and modular. Its auxiliary objectives can be integrated into existing imitation-learning pipelines with limited architectural changes, offering potential applicability to a broader range of navigation models.
  • The gains over the same-backbone baseline are substantial. On unseen OpenUAV maps under L1 assistance, success rises from 23.49% to 44.89%, while collisions fall from 63.26% to 38.94%. Improvements in success and collision rates on HUGE-Bench provide additional evidence for the effectiveness of the proposed methods.
Critical weaknesses
  • All OpenUAV experiments use the L1 assistant, which provides ongoing navigation hints from the test episode's oracle trajectory. Such information is generally unavailable in deployment. The results establish a benefit from the risk penalty under strong assistance, but do not show whether it still reduces collisions while preserving task completion without oracle hints, which weaken the safety claim.
  • The paper is motivated by differences in the safety margins of expert trajectories, but does not establish how common low-clearance demonstrations are or how much they contribute to baseline collisions. Collisions may also arise from perception failures, trajectory prediction errors, or error accumulation. The results support the effectiveness of the clearance penalty, but leave the link between demonstration risk and imitation-learning failures insufficiently established.
Questions for authors

Could the authors analyze the main causes of baseline collisions and assess how much low-clearance expert trajectories contribute? Do limited safety margins amplify the collision risk caused by perception or prediction errors? Besides, if possible, could author provide explicit evidence that the benefit of combining geometric alignment with risk supervision comes from better geometry representations?

3 comments

Reviewer 6jGt — Part 1/3

Thank you for recognizing our strengths in the simplicity and modularity of GeoRisk-IL, its potential applicability to existing imitation-learning pipelines, and strong performance. We also thank you for your constructive feedback. Below are point-to-point answers to each of the comments.

W1: All OpenUAV experiments use the L1 assistant, but do not show whether it still reduces collisions while preserving task completion without oracle hints.

A: Thank you for raising this concern. During rebuttal, we add new experiments to verify that GeoRisk-IL reduces collisions and preserves task completion without oracle hints. We evaluated the baseline model and GeoRisk-IL on OpenUAV test splits that do not have any assistance or directional hints. This essentially requires the UAV to explore the scenario for a specified target, which is more challenging than the current settings.

Across Test Seen, Test Unseen Object, and Test Unseen Map, GeoRisk-IL reduces CR by 30.95, 22.42, and 39.77 percentage points, respectively. NE, SR, OSR, and SPL also improve on all three splits.

Table R1-1. Navigation performance on OpenUAV without any assistance or directional hints. NE is in meters; CR, SR, OSR, and SPL are percentages.

SplitMethodCR ↓NE ↓SR ↑OSR ↑SPL ↑
Test SeenQwen3-VL Base74.96138.5215.5918.1214.25
Test SeenGeoRisk-IL44.01113.8928.1432.7224.97
Test Unseen ObjectQwen3-VL Base64.71138.3221.7825.6020.76
Test Unseen ObjectGeoRisk-IL42.29112.2027.9833.0726.00
Test Unseen MapQwen3-VL Base83.19140.169.1911.278.52
Test Unseen MapGeoRisk-IL43.42115.9222.9629.2320.45

Reviewer 6jGt — Part 2/3

W2: The paper is motivated by differences in the safety margins of expert trajectories, but does not establish how common low-clearance demonstrations are or how much they contribute to baseline collisions.

W2.1: How common are low-clearance demonstrations?

A: Low-clearance demonstrations are very common in the training set. We measured the minimum obstacle clearance of GT expert trajectories. As shown in Table R1-2, 85.18% of training episodes contain at least one frame with clearance lower than 2.0 m, which we believe can be considered as 'low clearance'.

Table R1-2. Clearance-based trajectory distribution on TravelUAV training set.

Episode-level minimum clearance<0.5 m<1.0 m<1.5 m<2.0 m
Percentage of episodes58.74%62.97%70.53%85.18%

W2.2: How much do low-clearance demonstrations contribute to baseline collisions? Collisions may also arise from perception failures, trajectory prediction errors, or error accumulation. The results support the effectiveness of the clearance penalty, but leave the link between demonstration risk and imitation-learning failures insufficiently established.

A: We feel this needs clarification. From our point of view, the core issue with imitation learning (IL) is that it cannot distinguish which trajectory is better suited to the current situation. In UAV navigation, one core issue is collision, and we thus believe the learning paradigm should reflect which trajectory is better in terms of obstacle clearance. We do not want to indicate or in any manner hint that using ‘safe’ GT trajectories alone can solve this issue with IL. (In L22 and L47, our words are 'vanilla imitation learning treats these trajectories as equally desirable supervision, which is not helpful for lowering collision rates' and 'Standard imitation learning treats every demonstrated trajectory as an equally valid target, ..., blindly following these expert demonstrations cannot enhance the collision avoidance ability of UAVs')

In fact, using a ‘safe’ GT trajectory (with a higher obstacle clearance) alone is not enough for improving collision avoidance capabilities. In Table R1-3, we substitute the low-clearance GT actions with a safe alternative, i.e., future trajectories whose obstacle clearance is larger than 2m. As the results indicate, this does not bring significant changes to the crash rate and success rate. In contrast, as long as we provide direct supervision on differentiating the ‘suitability’ of the trajectories, we see a strong and consistent improvement (77.29% vs. 64.03% success rate in Table R1-3).

As for the reasons for collisions, we believe that they mainly come from the training paradigm, where it is not possible to distinguish the 'safety' aspect of trajectories. The newly added experiments in Table R1-3 support this belief, where vanilla IL alone cannot help to lower the crash rate, even when the GT supervision is from high-clearance trajectories.

We will clarify this.

Table R1-3. Influence of using safe GT trajectories as targets for IL on OpenUAV Test Seen.

ModelLabelsCR (%) ↓SR (%) ↑
Qwen3-VL BaseOriginal Traj.30.4764.03
Qwen3-VL Base‘Safe’ Traj.30.3964.60
GeoRisk-ILOriginal Traj.17.9877.29

Note: All models in Table R1-3 were independently retrained and reevaluated for this experiment.

Reviewer 6jGt — Part 3/3

Q1: Could the authors analyze the main causes of baseline collisions and assess how much low-clearance expert trajectories contribute?

A: According to W2.1 and W2.2, low-clearance demonstrations are common, but replacing them with higher-clearance targets barely changes baseline collisions. We believe the main cause of baseline collisions lies in the training objective: vanilla IL matches GT trajectories without explicitly distinguishing predictions by their obstacle clearance. GeoRisk-IL addresses this limitation by supervising the safety of the model's own predicted trajectories.

Q2: Do limited safety margins amplify the collision risk caused by perception or prediction errors?

A: Limited safety margins amplify the collision risk from both perception and prediction errors. We evaluated Qwen3-VL Base in open loop on all expert demonstration frames in OpenUAV Test Seen. Perception noise consisted of Gaussian perturbations to XYZ positions (σ=1\sigma=1 m) and RGB channels (σ=100/255\sigma=100/255); prediction noise consisted of random trajectory perturbations with an RMS displacement of 1 m.

Table R1-4. Open-loop predicted-path collision rates (%) of Qwen3-VL Base on OpenUAV Test Seen.

Current clearanceNo noisePerception noisePrediction noise
0<c<10<c<1 m2.9164.33610.246
1≤c<21\le c<2 m0.3901.0250.820
c≥2c\ge2 m0.0240.0510.027

Q3: Besides, if possible, could author provide explicit evidence that the benefit of combining geometric alignment with risk supervision comes from better geometry representations?

A: During rebuttal, we added a linear probing experiment on depth estimation to verify that LLM features from GeoRisk-IL indeed offer better geometry representations, and this stronger geometric understanding positively correlates with the collision avoidance capabilities.

We examine this connection by comparing depth reconstruction and closed-loop navigation for the same four variants: Qwen3-VL Base, Geometry only, Penalty only, and GeoRisk-IL. Both evaluations use OpenUAV Test Unseen Map.

For depth probing, we froze each navigation model, including LoRA, and trained an identical single-layer fully connected probe on its final LLM-layer image-token states. The per-token predictions form separate 8×88\times8 depth maps for the front and downward views, supervised by spatially averaged ground-truth depth using an L1 loss. Each probe was trained on 20,000 training frames for 20 epochs (batch size 128, learning rate 10−310^{-3}), then evaluated on 10,000 randomly sampled frames from Test Unseen Map.

Table R1-5 reports these probe errors alongside the closed-loop CR and SR of the corresponding navigation models from the four-way ablation on the same split.

Geometric alignment improves depth reconstruction both without risk supervision (9.976 → 6.287 m) and with it (9.637 → 6.268 m, a 34.96% reduction). These comparisons show that geometric alignment makes depth information more accessible in the navigation representations.

The closed-loop results reveal how this representation improvement relates to navigation. Adding geometric alignment increases SR by 1.56 percentage points over Qwen3-VL Base, compared with 8.15 percentage points over Penalty only. In the latter comparison, CR also decreases from 41.75% to 38.94%. Thus, geometric alignment improves depth representations in both settings, while yielding a larger navigation success gain when risk supervision is present. Together, these results support the interpretation that better geometric representations help the policy learn from clearance supervision.

Table R1-5. Depth-probe accuracy and closed-loop navigation performance on OpenUAV Test Unseen Map. MAE is reported as mean ±\pm standard deviation over three independent probe-training and evaluation runs.

ModelDepth MAE (m) ↓CR (%) ↓SR (%) ↑
Qwen3-VL Base9.976 ±\pm 0.01163.2623.49
Geometry only6.287 ±\pm 0.03260.4425.05
Penalty only9.637 ±\pm 0.00641.7536.74
GeoRisk-IL6.268 ±\pm 0.07238.9444.89

Comments on the paper

1 comment
AuthorsOpenReview ↗

Message to AC and Reviewers

Dear Area Chair and Reviewers,

We sincerely thank the reviewers for their constructive feedback and for recognizing the value of our work.

In light of the recent changes, we present a summary of our rebuttal and revisions. We are pleased that all reviewers appreciate our simple-yet-effective design. Following their constructive feedback, we have conducted extensive additional experiments to address every concern, strengthening the quality of our paper.

1. Consensus on Strengths

Reviewers highlighted the following strengths of our work:

  • Strong performance (Reviewers 6jGt, nvy5, Tuyq): All reviewers noted our empirical performance improvement over previous methods as ‘substantial’.
  • Simple design (Reviewers 6jGt, nvy5): Reviewers view the simplicity of our method as one of its core strengths. We believe being simple yet effective is a key characteristic that readers would find worth their attention and worth sharing with others.
  • No inference-time overhead (Reviewers nvy5, Tuyq): Unlike existing collision-risk-based designs, the proposed method only uses this loss during training.

2. Addressed Weaknesses & Additional Experiments

We have conducted extensive new experiments to resolve every raised question:

  • More thorough statistical study and added experiments (Response to nvy5, Tuyq): We added 5 repeated runs to establish statistical significance (Table R2-3), and extended ablation studies to more data splits (Table R3-3), verifying consistent improvements under different environments.
  • Clarification of the experimental settings (Response to 6jGt, nvy5, Tuyq): We detailed training-clearance statistics (Table R1-2), all baseline configurations (Table R2-1), our baseline design (Table R2-2), voxel size, and depth history window.
  • Experiments without the L1 assistant (Response to 6jGt): We evaluated without any oracle assistance (Table R1-1), demonstrating consistent performance improvements without any assistance.
  • Generalization under noisy inputs (Response to 6jGt): We examined perception and prediction noise (Table R1-4), confirming the strong generalization capabilities.
  • Physical flight validation (Response to Tuyq): We respectfully point out that physical UAV flight tests are not available at this moment due to regulatory reasons. Nevertheless, we conducted simulated experiments to verify the robustness to real-world flight perturbations (Tables R3-1 and R3-2).

3. Contribution and Novelty Concern

We respectfully disagree with Reviewers nvy5 and Tuyq.

We believe a meaningful research contribution need not rely on complex new architectures or building blocks. Instead, manuscripts that spot an underexplored issue with the most popular paradigm and deliver a practical, easy-to-apply fix can also be of interest to many readers and thus worth sharing with the community. Specifically, our work and newly added analyses establish three concrete contributions:

(i) Identifying imitation learning (IL)’s potential issue of treating all GT trajectories uniformly

(ii) A simple-yet-effective collision penalty + geometric supervision approach to vanilla IL to counteract this issue

(iii) Identifying the synergy between the collision penalty and the geometric supervision

4. Summary of Revisions

Based on the feedback, we have updated the paper as follows:

  • Section 4.3, Table 3: Extended ablation study.
  • Section 4.3, Table 4: Depth-probing and predicted-trajectory clearance analysis.
  • Appendix A.3: Voxel map clarification.
  • Appendix A.7, Table 6: Experiments with re-planned GT trajectories.
  • Appendix A.8, Table 7: Experiments without any assistant.
  • Appendices A.9–A.10, Tables 8–9: Trajectory clearance statistics and robustness to input noise.
  • Appendix A.11, Tables 10–11: Clarification on our baseline design.
  • Appendix A.12, Table 12: Experiments with 5 repeated runs.
  • Appendix A.13, Tables 13–14: Simulation under real-world flight noise.

With these comprehensive revisions, we have provided detailed responses and extensive new evidence for every point raised by the reviewers. We believe these updates significantly solidify the contribution and robustness of our method. We hope this summary facilitates a smooth review process for the Area Chair.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.