acceptodds
Under review as a conference paper at ICLR 2027

SIMPLIFYING COMPLEXITY: STAGE ADAPTIVE EVIDENCE ROUTING FOR UAV VLN

Abstract

UAV vision-language navigation requires long-range search, target identification, and precise terminal control. However, existing methods typically use the same visual-language inputs throughout navigation, failing to account for the change in evidence requirements before and after target discovery. Before the target appears, navigation primarily relies on scene context and directional hints for exploration; once the target is reliably identified, the decision process should shift toward target-centered local control. To address this, we propose a stage-adaptive evidence-routing framework that divides navigation into exploration before target discovery and target-centered navigation afterward. The latter is further divided into approach, regrounding, and reacquisition. The framework combines hint-conditioned target identification with lightweight target memory to determine the current stage from the relationship between current and previously accepted target evidence. It then routes the corresponding visual and textual information to a shared action policy. On 3,005 TravelUAV test episodes, our method achieves success rates of 58.41%, 66.29%, and 40.71% on the Seen, Unseen Object, and Unseen Map splits, improving over AerialVLA by 10.45, 9.69, and 3.13 percentage points, respectively. These results show that adapting decision inputs to target discovery and changes in target evidence improves successful termination across diverse navigation conditions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Aerial Vision-and-Language Navigation, Vision-Language-Action Learning, Stage-Adaptive Decision Making, Multimodal Evidence Routing.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qDe6OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses UAV vision-language navigation and argues that the evidence an agent needs changes once the target has been found. Before discovery, the agent explores using scene context and the instruction. After discovery, the task becomes target-centered local control, split into three sub-stages: approach, regrounding, and reacquisition. The method identifies the target with an open-vocabulary detector and a vision-language verifier, and it keeps a lightweight target memory to decide the current stage. It then routes only the visual and language inputs suited to that stage into a single shared action policy. Experiments on the TravelUAV benchmark show solid improvements over existing methods, including on unseen objects and unseen maps.

Critical strengths
  • The motivation is clear and intuitive: once a static target is grounded, the task reduces to target-centered local control.
  • The stage logic is explicit and interpretable (Table 1). Separating presence verification from ranking, and allowing landing only when the target is currently localized, are sound design choices.
  • Gains are consistent across all three splits in the Full setting. The controlled comparison in Table 6 directly supports stage-adaptive routing.
  • The appendix gives detailed training data, schedules, and thresholds, which helps reproducibility.
Critical weaknesses
  • Heavy reliance on the oracle directional hint. At every step, including test time, the hint is computed from the ground-truth target position (Eq. 1), as the UAV-Need-Help protocol allows. Beyond the policy input, the method builds view routing and 10° geometric candidate filtering on top of this hint, so its target identification may depend on it much more than the baselines do. The paper doesn't say whether baselines were retrained to use the hint equally well. Results with a noisy or coarser hint would show how robust the method really is.
  • The efficiency claim isn't supported. The paper says uniform multimodal fusion adds unnecessary computation. Yet each step chains Grounding DINO, up to three Qwen3-VL-8B calls, and OpenVLA-7B, and the limitations section admits latency goes up. Per-step latency, memory, and parameter comparisons with the baselines should be reported.
  • The experimental analysis is incomplete. The method falls behind AerialVLA on some subsets, such as NE and OSR on Unseen Map - Easy and SPL on Seen Hard, and these aren't discussed. Stage-determination accuracy and a failure breakdown are missing, so it's hard to tell whether errors come from perception or control. No variance across seeds is reported, which matters for the +3.13 SR gain on Unseen Map.
  • No real-world experiments. Everything is evaluated in simulation, with no physical UAV deployment or sim-to-real validation. The approach stage drops background and language inputs and assumes a clear line of sight to a static target, so robustness to obstacles, occlusion, wind, and sensor noise is untested. At least a small real-world or perturbation study is needed.
Questions for authors

N/A

6 comments
AuthorsOpenReview ↗

Response part 1

Thank you for your comment and review. We find that navigation decisions require different evidence before and after target discovery: before discovery, they depend on the global scene and the full instruction; afterward, they shift to target-driven local control. We therefore propose multi-stage adaptive evidence routing and lightweight target memory. By selecting visual and language inputs by stage, a shared action policy can handle long-range search, target approach, regrounding, and recovery after target loss. This addresses the difficulty of adapting fixed inputs to every stage and avoids interference from substantial redundant information. Across all 3,005 official TravelUAV test trajectories, our method's SR exceeds that of the strongest baseline reported in the paper on Seen, Unseen Object, and Unseen Map. We respond in detail below to your concerns under Weaknesses and Questions.

W1: The method relies heavily on an oracle direction hint computed from the ground-truth target position at every step, including test time (Eq. 1), as permitted by the UAV-Need-Help protocol. The hint drives not only the policy input but also view routing and the 10° geometric candidate filter, so target recognition may depend on it more than in the baselines. The paper does not say whether baselines were retrained to use the hint equally effectively. Results with noisy or coarse hints would show the method's actual robustness.

Response: Thank you for your comment. Because UAV-Need-Help is a difficult long-range navigation task, all compared methods use the target-direction information provided by the task to assist search. Our method follows the same task protocol and uses relative bearing information without additionally providing target coordinates or a ground-truth path. It receives a quantized, coarse relative bearing rather than target coordinates, distance, or a path. Because the hint also participates in target recognition, we agree that its sensitivity to errors should be tested.

We tested navigation with zero-mean Gaussian noise added to the target's relative bearing at each step, then mapped the perturbed angle to a direction category. Mild, moderate, and severe noise use standard deviations of 10°, 30°, and 60°, respectively. “No Hint” retains the original full task instruction but removes the step-wise dynamic direction hint. The results are:

Hint conditionStage error rate (%)
No noise21.33
Mild noise27.54
Moderate noise40.38
Severe noise57.64
No Hint62.57
MethodHint conditionSR (%)OSR (%)NE (m)SPL (%)
TravelUAVNo noise14.2639.50115.5312.95
TravelUAVMild noise12.1235.93116.6510.76
TravelUAVModerate noise10.3817.54136.228.98
TravelUAVSevere noise2.617.86160.651.23
TravelUAVNo Hint1.354.12164.940.51
AerialVLANo noise46.4657.6765.4436.94
AerialVLAMild noise43.7654.6967.0630.15
AerialVLAModerate noise31.7137.5892.2424.23
AerialVLASevere noise8.4713.34129.266.04
AerialVLANo Hint1.755.45134.540.81
Our methodNo noise54.4271.4343.3540.48
Our methodMild noise50.4067.8846.0439.39
Our methodModerate noise34.9244.3480.8725.80
Our methodSevere noise8.9914.96125.216.39
Our methodNo Hint3.157.77131.231.04

As direction noise increases, our stage-error rate rises from 21.33% without noise to 27.54%, 40.38%, and 57.64%; removing the dynamic hint raises it to 62.57%. Thus, inaccurate direction information increasingly disrupts stage decisions. At the navigation level, SR and SPL decline for all three methods. Our method retains the highest SR, OSR, and SPL and the lowest NE in all five conditions. Hint quality affects both stage classification and navigation, but our method preserves a relative advantage under the tested perturbations.

W2: The efficiency claim is not fully supported. The paper argues that uniform multimodal fusion adds unnecessary computation, yet each step chains Grounding DINO, up to three Qwen3-VL-8B calls, and OpenVLA-7B. The limitations section acknowledges increased latency. Please compare per-step latency, memory, and parameter counts with baselines.

Response: Thank you for your comment.

To assess efficiency, we measured latency, GPU memory, and parameter counts on the same hardware and compared our method with the released baseline implementations. NavFoM and LongFly were excluded because their code was unavailable. The results are:

MethodMean per-step decision latencyMean navigation stepsMean cumulative decision time per navigation episodeLoaded parametersInference GPU memory
AerialVLA1.643 s169277.667 s7.77 B14.86 GiB
TravelUAV1.826 s171312.246 s9.28 B18.07 GiB
Our method4.669 s62289.478 s16.82 B22.49 GiB+15.07 GiB

We also measured module latencies and step proportions for each stage:

AuthorsOpenReview ↗

Response part 2

Decision stageStep proportionSemantic verificationCandidate rankingStage comparisonStage-decision subtotalAction policyComplete decision
Exploration32.25%0.755 s0.095 s0.361 s1.211 s0.677 s1.888 s
Approach20.33%3.173 s1.305 s2.605 s7.083 s0.670 s7.753 s
Regrounding19.95%4.037 s1.251 s3.724 s9.012 s0.663 s9.675 s
Reacquisition27.47%0.743 s0.105 s0.385 s1.233 s0.783 s2.016 s

Our method requires more computation per step, more parameters, and more GPU memory: its stage-weighted complete decision latency is 4.669 s per step, or 2.84 times that of AerialVLA and 2.56 times that of TravelUAV. Action-policy inference remains at 0.663–0.783 s per step across the four stages; the added time mainly comes from semantic verification and stage decisions, especially the more detailed checks in Approach and Regrounding. However, our method averages only 62 navigation steps, versus 169 for AerialVLA and 171 for TravelUAV. Mean cumulative decision time per episode is 289.478 s, only 4.25% above AerialVLA and 7.29% below TravelUAV. Overall navigation efficiency therefore does not decline substantially. Given the improved success rate of the full method, we consider this computational cost acceptable.

W3: The experimental analysis is incomplete. The method trails AerialVLA on some metrics and subsets, including NE and OSR on Unseen Map and SPL on Seen Hard, without discussion. Stage accuracy and failure decomposition are missing, making it hard to distinguish perception from control errors. There is no seed variability analysis, which matters for the +3.13-point SR gain on Unseen Map.

Response: Thank you for your comment.

The original manuscript reported performance on each subset, but did not analyze the local disadvantages or display the multi-seed results obtained earlier. Below, we first list the local differences across the relevant subsets, then report the aggregate results for seeds 41–50 to examine whether the full method's gain relative to AerialVLA is stable.

Subset and metricAerialVLAOur methodDifference (Ours − baseline)
Unseen Map Easy · NE ↓44.99 m53.19 m+8.20 m
Unseen Map Easy · OSR ↑58.47%57.14%−1.33 points
Unseen Map Easy · SR ↑41.89%45.81%+3.92 points
Seen Hard · SPL ↑40.26%38.48%−1.78 points
Seen Hard · SR ↑46.30%51.39%+5.09 points

On Unseen Map Easy, our SR is higher than AerialVLA's, but OSR is slightly lower and mean NE is higher. Our retrospective analysis indicates that unseen-map obstacles can make a coarse direction hint insufficient to identify a passable route; when a similar distractor is mistaken for the target, the policy can approach the wrong location. Both can prevent some trajectories from reaching the true target neighborhood, lowering OSR and raising NE. On Seen Hard, we observed that occlusion-triggered reacquisition can make successful trajectories backtrack, while the limited scene context during Approach can require detours or retreat around obstacles. Both increase flown path length and lower SPL.

To distinguish perception from control further, we checked the stage at each decision step and corrected erroneous stage decisions one at a time on failed trajectories. Each corrected rollout resumes from the same simulator state, replacing only the predicted stage with its ground-truth counterpart while holding target observations, policy weights, and the official success condition fixed. This tests whether stage correction improves navigation.

StageStage classification accuracyDescription
Exploration83.7%Global search before target discovery
Approach81.3%Local approach after target localization
Regrounding76.8%Reverification of current and remembered targets
Reacquisition70.5%Recovery after temporary target loss
StatisticDefinitionResult
Fraction of correction points yielding success after ground-truth stage substitutionCorrect one point at a time and rerun closed-loop; successful corrections ÷ all tested erroneous transitions47.89%
NE change at correction points that remained unsuccessfulPaired comparison of final-position error in the original and corrected trajectories84.6 m → 72.9 m

When an erroneous stage decision was replaced with its ground-truth stage from the same simulator state, the originally failed navigation succeeded for 47.89% of the tested correction points. For reruns that still failed, mean NE decreased from 84.6 m to 72.9 m. This shows that correct stage selection can improve subsequent action decisions and closed-loop navigation outcomes. Because changing the stage also changes the evidence received by the action policy, this improvement reflects the joint effect of stage routing and action decisions; their independent contributions cannot be separated. We further report multi-seed results to help assess the robustness of the full method's gain relative to AerialVLA.

AuthorsOpenReview ↗

Response part 3

To show seed variability, we summarize our results for seeds 41–50 on the same splits and under the same evaluation protocol as mean ± standard deviation, alongside the results reported for AerialVLA. The following tables retain NE, SR, OSR, and SPL from Tables 2–4.

Table 2. Seen split

SettingMethodNE (m) ↓SR (%) ↑OSR (%) ↑SPL (%) ↑
FullTravelUAV98.6617.4548.8715.76
NavFoM93.0529.1749.2425.03
LongFly60.0236.3965.8731.07
AerialVLA65.8847.9657.6938.54
Ours37.96±2.2358.41±1.3779.06±1.5540.62±0.85
EasyTravelUAV66.4020.2651.2318.10
NavFoM58.9832.9153.1627.87
LongFly38.1038.5271.9031.24
AerialVLA43.7649.3061.3037.14
Ours19.87±2.6264.10±2.0085.73±2.1442.36±1.43
HardTravelUAV138.0414.0245.9812.90
NavFoM143.8323.5843.4020.80
LongFly85.2033.9458.9430.88
AerialVLA93.1646.3053.2340.26
Ours60.26±3.8351.39±1.8670.83±2.3938.48±0.65

Table 3. Unseen Object split

Full: n = 629; Easy: n = 209; Hard: n = 420.

SettingMethodNE (m) ↓SR (%) ↑OSR (%) ↑SPL (%) ↑
FullTravelUAV118.1122.4246.9020.51
NavFoM108.0429.8347.9927.20
LongFly66.7443.8764.5638.39
AerialVLA61.4556.6064.8646.61
Ours39.09±2.4566.29±1.4681.83±1.6453.56±1.46
EasyTravelUAV86.1224.4049.2822.03
NavFoM70.5132.5450.7229.54
LongFly54.8438.0156.8431.36
AerialVLA45.7256.9464.1143.76
Ours46.04±0.2162.00±1.8372.90±2.0345.51±0.75
HardTravelUAV134.0321.4345.7119.75
NavFoM133.0128.0346.1825.64
LongFly57.0750.2574.1645.27
AerialVLA69.2756.4365.2448.03
Ours35.63±3.6168.43±2.0686.27±2.2557.57±2.12

Table 4. Unseen Map split

Full: n = 958; Easy: n = 561; Hard: n = 397.

SettingMethodNE (m) ↓SR (%) ↑OSR (%) ↑SPL (%) ↑
FullTravelUAV138.804.1820.773.84
NavFoM125.106.3018.955.68
LongFly108.3211.2730.279.32
AerialVLA67.4237.5852.9228.22
Ours54.13±2.4540.71±0.8853.32±0.4231.68±0.94
EasyTravelUAV102.944.6322.824.24
NavFoM102.416.7720.076.04
LongFly78.5612.9634.3110.32
AerialVLA44.9941.8958.4729.72
Ours53.19±3.2045.81±1.4657.14±0.4533.92±1.53
HardTravelUAV189.463.5317.883.28
NavFoM170.585.3615.714.97
LongFly148.109.0224.887.98
AerialVLA99.1131.4945.0926.11
Ours55.47±3.8233.50±0.7547.92±0.8428.52±0.80

Across the three tables, the full method's mean SR on Seen, Unseen Object, and Unseen Map Full exceeds the published AerialVLA results by 10.45, 9.69, and 3.13 percentage points, respectively. The cross-seed standard deviations shown are 1.37, 1.46, and 0.88 percentage points. Positive SR gains also appear on the Easy and Hard subsets of all three splits. These multi-seed summaries provide supplementary evidence for the robustness of the full method's gain relative to AerialVLA.

W4: There are no real-world experiments. All evaluation is in simulation, with no physical-drone deployment or sim-to-real validation. During Approach, the policy drops background and language inputs and assumes a clear view of a static target. Robustness to obstacles, occlusion, wind, and sensor noise remains untested. At least a small real-world or perturbation study is needed.

Response: Thank you for pointing out the importance of real-world validation. Published UAV-Need-Help comparisons mainly rely on reproducible simulation tasks, and there is not yet a unified real-world evaluation protocol. UAV-Need-Help requires a UAV to search for, recognize, and land near a target that is initially invisible; the task itself is very difficult. Consequently, related methods have conducted their evaluations in simulation. We follow the official simulation splits and success criteria to align with these methods. Following your suggestion, we also added a small physical-drone experiment to test whether stage routing can support closed-loop navigation under real obstacles, occlusion, and wind disturbance.

We used a controlled 28 m × 28 m indoor arena with two kinds of static-target tasks, both with the target initially out of view. In the first, few obstacles surround the target and the drone can approach directly after discovery. The second includes box obstacles and visually similar distractors, so detours can temporarily hide the target. For each fixed layout, we tested ten distinct targets once each, giving 20 trials with a 70-decision-step limit. The platform uses a Jetson Orin NX 16GB, a MicoAir743v2 flight controller, and five D435 RGB-D cameras facing forward, backward, left, right, and down. RGB supports recognition and high-level action prediction; depth and fused pose support local mapping and collision avoidance.

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

This paper proposes stage-adaptive evidence routing for UAV vision-language navigation. It divides navigation into exploration, approach, regrounding, and reacquisition, maintains lightweight target evidence across stages, and selects stage-appropriate visual and textual inputs for a shared action policy. On 3,005 TravelUAV episodes, it reports SR improvements over AerialVLA across Seen, Unseen Object, and Unseen Map splits, supported by stage ablations and a controlled fixed-input comparison.

Strengths

  1. The motivation for using different evidence before and after target discovery is intuitive and relevant to long-horizon aerial navigation.

  2. The four-stage design is clear, and the paper provides component ablations and a controlled comparison against fixed multimodal inputs in Tables 5 and 6.

  3. The experimental evaluation covers all 3,005 test episodes and reports results across seen and unseen targets and environments.

Weaknesses

  1. The overall computational benefit is not yet demonstrated. Although routing reduces the inputs presented to the action policy, target recognition and stage determination use Grounding DINO and Qwen3-VL-8B, with repeated semantic auditing (Section 4.1). The paper acknowledges increased decision latency in Section 4.6, but no end-to-end runtime or model-call analysis is reported.

  2. The direct-approach assumption would benefit from qualification. Section 1 motivates target-centered control partly through an unobstructed viewing ray, but visual line of sight does not always guarantee a collision-free straight-line flight segment in three-dimensional scenes. A discussion or analysis of failures around obstacles and partial occlusion would be useful.

  3. The relative contribution of individual stages is less clear in the more difficult unseen-map setting. In Table 5, removing regrounding changes SR only from 40.71% to 39.81%, while the overall SR improvement over AerialVLA is 3.13 points (Table 4). Reporting variability or examples of incorrect stage transitions would help judge the stability of these gains.

Critical strengths
  1. The central observation that evidence requirements change after target discovery is sensible and relevant to aerial VLN. The stage-based formulation provides a clear system design for separating long-range search from target-centered control.
  2. The method is evaluated on the complete TravelUAV test set, including unseen objects and maps, with component ablations. These results support the practical value of stage-adaptive evidence selection.
Critical weaknesses
  1. The claimed simplification is not fully supported by an end-to-end efficiency analysis. The system relies on Grounding DINO and Qwen3-VL-8B for target grounding and stage decisions, and Section 4.6 acknowledges increased decision latency. Without overall inference time or model-call counts, the computational benefit of reducing policy inputs remains uncertain.
  2. The individual contributions of later navigation stages are not equally clear on unseen maps. Table 5 shows that removing regrounding changes unseen-map SR from 40.71% to 39.81%, while the full improvement over AerialVLA on that split is 3.13 points (Table 4). A more focused analysis of stage-transition errors would help distinguish a consistent mechanism from small empirical fluctuations.
Questions for authors
  1. What is the end-to-end decision latency, including grounding, semantic verification, stage decisions, and action-policy inference, relative to the fixed-input baseline?
  2. How frequently do incorrect stage transitions occur in unseen maps, and what fraction of unsuccessful episodes can be attributed to these transitions?
4 comments
AuthorsOpenReview ↗

Response part 1

Thank you very much for recognizing the motivation, the four-stage design, and the evaluation on the complete test set. Your meticulous and insightful comments impressed us and conveyed your expectation that we further improve this work.

As stated in the General Response, we supplemented the decision-latency, stage-transition, and navigation analyses to examine the benefits and costs of our method, particularly its system cost and stability on Unseen Map. Beyond its performance on the present task, we believe selecting different evidence before and after target discovery while using a shared action policy may also inspire other long-range navigation tasks. Other reviewers likewise recognized this core observation, the controlled comparison, and the interpretability of the stage rules. Thank you again for your detailed comments and suggestions; they are important to improving this paper. We respond to your concerns point by point below.

W1 & Q1: Relative to the fixed-input baseline, what is the end-to-end decision latency, including grounding, semantic verification, stage decisions, and action-policy inference?

Response: Thank you for your comment.

Using the code available for the open-source baselines (NavFoM and LongFly were not included because their code was unavailable), we measured complete decision time from the current observation to the final action. The measurement includes target grounding and semantic verification, candidate ranking, stage comparison, and action-policy inference. The results are:

MethodMean per-step decision latencyMean simulation stepsMean cumulative decision time per episode
AerialVLA1.643 s169277.667 s
TravelUAV1.826 s171312.246 s
Ours4.669 s62289.478 s

Our method determines target state and navigation stage before action inference, so its mean per-step latency is 4.669 s, compared with 1.643 s for AerialVLA and 1.826 s for TravelUAV. However, stage-adaptive inputs reduce the mean number of simulation steps to 62, versus 169 and 171 for the two baselines. Mean cumulative decision time per navigation episode is therefore 289.478 s, only 4.25% above AerialVLA and 7.29% below TravelUAV, placing the methods in a similar range of overall decision-time efficiency.

Per-stage module latency and step proportions are:

Decision stageStep proportionSemantic verificationCandidate rankingStage comparisonStage-decision subtotalAction policyComplete decision
Exploration32.25%0.755 s0.095 s0.361 s1.211 s0.677 s1.888 s
Approach20.33%3.173 s1.305 s2.605 s7.083 s0.670 s7.753 s
Regrounding19.95%4.037 s1.251 s3.724 s9.012 s0.663 s9.675 s
Reacquisition27.47%0.743 s0.105 s0.385 s1.233 s0.783 s2.016 s

Weighted by each stage's actual proportion of decision steps, the full method's mean end-to-end decision latency is 4.669 s. Action-policy inference remains stable at 0.663–0.783 s across the four stages. The added compute mainly comes from target verification and stage decisions performed as needed. Approach and Regrounding require more stringent comparisons between the current and historical targets and therefore take longer; Exploration and Reacquisition use lighter decision processes. Considering per-step latency, decision-step count, and total latency per episode together, the stage-adaptive method uses more complete evidence per decision while retaining overall navigation efficiency close to that of the fixed-input methods. Thank you again for your suggestion. We will add these experimental results and detailed analysis in a future version.

W2 & Q2: How frequently do erroneous stage transitions occur on unseen maps, and how many failed episodes can be attributed to them?

Response: Thank you for your comment.

To measure transition errors directly, we checked the stage at each decision step. We constructed a reference stage from true-target visibility and target-memory identity, then compared the predicted stage change between adjacent decisions with the reference. An incorrect switch or a missed switch was counted as a transition disagreement.

Our analysis of the full Unseen Map test set yields a stage-transition disagreement rate of 20.58%.

To test whether correcting the stage can improve failed trajectories, we intervened at each erroneous stage point separately: we resumed navigation from the same simulator state and replaced only that predicted stage with the ground-truth stage, leaving target observations, policy weights, and the official success condition unchanged. Across all failed episodes, 51.74% turned into successes after correction, showing that correct stage transitions can improve navigation outcomes.

AuthorsOpenReview ↗

Response part 2

However, stage classification determines which visual and language evidence reaches the action policy, so routing and action selection act together. Their individual contributions to failure cannot be separated from this result. We therefore provide multi-seed results as complementary evidence for the robustness of the full method's gain over AerialVLA.

To show seed variability, we summarize our method over seeds 41–50 as mean ± standard deviation under the same split and protocol, alongside the AerialVLA results reported in the literature:

Table 4. Unseen Map split

Full: n = 958; Easy: n = 561; Hard: n = 397.

SettingMethodNE (m) ↓SR (%) ↑OSR (%) ↑SPL (%) ↑
FullTravelUAV138.804.1820.773.84
NavFoM125.106.3018.955.68
LongFly108.3211.2730.279.32
AerialVLA67.4237.5852.9228.22
Ours54.13±2.4540.71±0.8853.32±0.4231.68±0.94
EasyTravelUAV102.944.6322.824.24
NavFoM102.416.7720.076.04
LongFly78.5612.9634.3110.32
AerialVLA44.9941.8958.4729.72
Ours53.19±3.2045.81±1.4657.14±0.4533.92±1.53
HardTravelUAV189.463.5317.883.28
NavFoM170.585.3615.714.97
LongFly148.109.0224.887.98
AerialVLA99.1131.4945.0926.11
Ours55.47±3.8233.50±0.7547.92±0.8428.52±0.80

As shown above, on Unseen Map Full, the full method's mean SR, OSR, and SPL exceed the AerialVLA results by 3.13, 0.40, and 3.46 percentage points, respectively. The corresponding cross-seed standard deviations are 0.88, 0.42, and 0.94 percentage points. We will add these experimental results and the corresponding discussion in a future version. Thanks again!

We hope these clarifications and additional results address your concerns and strengthen confidence in our conclusions. Thank you again for your careful and fair evaluation. We would be pleased to clarify any remaining questions.

Thank you for the response. The authors have adequately addressed most of my concerns.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you for your insightful feedback and valuable comments. We will incorporate all your suggestions into the next revision of the manuscript. Thanks again!

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

This paper proposes a stage-adaptive evidence routing system for vision-language navigation with drones. Using hint-conditioned target identification and a lightweight target memory, each decision step is classified into one of four stages (exploration, approach, regrounding, reacquisition). Visual and textual inputs adapted to the stage are fed into a shared policy based on OpenVLA-7B. The evaluation covers all 3005 test episodes of the UAV-Need-Help task in TravelUAV, compared against the best baseline AerialVLA, the method improves the success rate (SR) by 10.45/9.69/3.13 points on the Seen/Unseen-Object/Unseen-Map splits, respectively. To validate the routing contribution, a controlled comparison with a fixed-input variant in Table 6 and an ablation that disables stage-wise feeding in Table 5 were carried out.

Critical strengths

1.Table 6 presents a controlled comparison, not just an end-to-end score jump. Even when the policy receives fixed inputs consisting of 5 viewpoints and the complete instructions at every stage, the version with routing remains superior on all splits, confirming the paper's hypothesis that tuning the input to the navigation stage is effective. In addition, Table 5 disables routing in each stage individually, falling back to fixed inputs whenever that stage occurs. However, the training protocol for the fixed-input variants requires clarification (see the questions).

2.The evaluation covers the full official test set and four families of relevant baselines, all receiving the same directional hint, and Table 4 explicitly indicates the cells where AerialVLA dominates on NE and OSR in the Unseen Map Easy subset. The full stage rules in Table 1 and Equations 5-11 are documented, making it possible to reproduce the system.

Critical weaknesses

1.No data is given on inference latency or memory usage, and everything appears to come from a single run. Every step calls Grounding DINO, up to 3 Qwen3-VL-8B adapters, and then the 7B policy. Section 4.6 mentions increased latency, but no per-step latency or memory figures are given, no runtime details for the modules (let alone the training and evaluation hardware), and no table contains random seeds or confidence intervals. The small difference on the Unseen Map split (SR +3.13, n=958) is difficult to assess without a paired statistical test at the episode level.

2.Issues related to the hint and the policy's input. The hint is computed in Equation 1 from the target's ground-truth position. All methods receive this hint, but in the proposed method it additionally drives the routing of viewpoints and geometric filtering, the hint's role in the stage decision is never ablated, so the conclusions are limited to hint-guided navigation. In the approach and regrounding stages, the policy receives only the crop of the target, the camera ID, and the hint defined in Equation 11. No history, no position, and no obstacle memory. Module-level ablations (presence verification, temporal comparison, detection threshold) are also missing, although Table 1 defines a reacquisition stage for lost targets, neither the frequency nor the success rate of this transition is reported.

Questions for authors

1.How was the fixed-input variant in Table 6 trained? Was it trained separately with adapted data and updates, or was the already trained model evaluated with fixed inputs? In the latter case, part of the difference may stem not from the routing itself but from the mismatch between training-time and test-time inputs. Please also explain the protocol for the variants in Table 5.

2.Section 4.1 states that all baselines receive the same directional hint. Please confirm that the values in Tables 2-4 were recomputed with this protocol, and specify the checkpoints and configurations used for each baseline.

3.Please report the per-module latency at each decision stage, the share of steps per stage, and the frequency and success rate of reacquisition transitions.

4.What happens to the stage decision if the direction computed from ground-truth data is replaced by a noisy estimate or the instructions alone?

3 comments
AuthorsOpenReview ↗

Response part 1

Thank you for your comment. We find that navigation decisions require different evidence before and after target discovery: before discovery, they depend on the global scene and the full instruction; afterward, they shift to target-driven local control. We therefore propose multi-stage adaptive evidence routing and lightweight target memory. By selecting visual and language inputs by stage, a shared action policy can handle long-range search, target approach, regrounding, and recovery after target loss. This addresses the difficulty of adapting fixed inputs to every stage and avoids interference from substantial redundant information. Across all 3,005 official TravelUAV test trajectories, our method's SR exceeds that of the strongest baseline reported in the paper on Seen, Unseen Object, and Unseen Map. We respond in detail below to your concerns under Weaknesses and Questions.

Q1: How was the fixed-input variant in Table 6 trained? Was it trained separately with adapted data and updates, or was an already-trained model evaluated with fixed inputs? In the latter case, some differences might reflect train-test input mismatch rather than routing. Please also explain the intervention protocol for the Table 5 variants.

Response: Thank you for your attention to the fairness of the fixed-input comparison and the implementation procedure for the stage-wise ablations. These two issues affect how the performance difference brought by stage routing should be interpreted. Our response follows.

The fixed-input variant in Table 6 was trained independently; we did not switch the trained stage-adaptive model to fixed inputs only at test time. Both variants use the same OpenVLA-7B initialization, training data, action supervision, and training budget. Every fixed-input training example contains five views (front, back, left, right, and down), the full instruction, and the current direction hint. The policy input does not vary by stage. We fine-tuned the language backbone using LoRA (r=64r=64, α=128\alpha=128, dropout=0.05\mathrm{dropout}=0.05). The visual encoder was frozen while the visual projector was trained. We used AdamW (weight decay=0.03\mathrm{weight\ decay}=0.03, gradient clipping=1.0\mathrm{gradient\ clipping}=1.0) and a cosine learning-rate schedule with a peak learning rate of ηmax⁡=2×10−4\eta_{\max}=2\times10^{-4} and warmup=5%\mathrm{warmup}=5\%; training used global batch size=64\mathrm{global\ batch\ size}=64, BF16 precision, and 5 epochs5\ \mathrm{epochs}. Evaluation retains exactly the fixed input used in training. Table 6 therefore compares two separately trained models with fixed versus stage-adaptive inputs; its difference does not arise from a test-only input substitution.

Table 5 uses the same trained weights as the full model for all stage-wise ablations. Target detection, semantic verification, stage classification, and target memory remain unchanged in every variant; only the input to the shared action policy changes. Thus, 'removing a stage' means disabling its stage-specific evidence route, not deleting the stage or changing its trigger rules. Specifically, 'w/o Exploration' replaces only the Exploration policy input with the fixed 'five views + full instruction + direction hint' input, while retaining the original routes for the other three stages. 'Exploration only' retains the Exploration route and uses fixed inputs for Approach, Regrounding, and Reacquisition. 'w/o Approach,' 'w/o Regrounding,' and 'w/o Reacquisition' replace only the policy input in the named stage. We will explain these interventions in greater detail in the next version. Thanks again!

Q2: Section 4.1 states that all baselines receive the same direction hint. Please confirm whether the numbers in Tables 2–4 were recomputed under this protocol, and specify the checkpoint and configuration used for each baseline.

Response: Thank you for your comment.

Regarding direction hints, all baseline models receive information about the target's relative bearing, although it is expressed differently. TravelUAV includes it in the task description and assistance instructions; NavFoM and LongFly receive it through the full task instruction; AerialVLA and our method use discrete direction categories. These differences mainly concern the language template and encoding form; each conveys the target's direction relative to the UAV.Policies based on large language models are robust to semantically equivalent descriptions and can interpret them as spatial directions. Our method receives no additional target distance, absolute coordinates, ground-truth waypoints, or obstacle-avoidance actions, and therefore gains no advantage from such extra hint information.

AuthorsOpenReview ↗

Response part 2

The baseline results in Tables 2–4 were not recomputed by us under a unified codebase. They come from the results reported for the same official TravelUAV splits in ‘AerialVLA: A Vision-Language-Action Model for UAV Navigation via Minimalist End-to-End Control.’ The reported configurations are TravelUAV with Vicuna-7B and five views, NavFoM with Four Views and Qwen2-7B, the full LongFly model with Qwen2.5-3B, and the full AerialVLA model with OpenVLA-7B. Because these are published results and some works did not provide downloadable final checkpoints or unique identifiers, we cannot specify exact checkpoint filenames. We will state each result's source and model configuration in the revision.

W1, W2 & Q3: Report module latency by decision stage, the proportion of steps spent in each stage, reacquisition frequency and success, and results across random seeds.

Response: Thank you for your review. Following your suggestion, we recently supplemented our evaluation with complete decision-latency measurements from observation input to action output, using released baseline code. NavFoM and LongFly were not included in this comparison because their code was unavailable. The results are as follows:

MethodMean per-step decision latencyMean simulation stepsMean cumulative decision time per episode
AerialVLA1.643 s169277.667 s
TravelUAV1.826 s171312.246 s
Ours4.669 s62289.478 s

Our method performs semantic verification, candidate ranking, and stage classification before each action, increasing per-decision latency. However, it averages 62 simulator steps, compared with 169 for AerialVLA and 171 for TravelUAV. Its mean cumulative decision time per episode is only 4.25% higher than AerialVLA's and 7.29% lower than TravelUAV's; the three methods therefore have comparable episode-level decision time. The added stage decisions do not create a large end-to-end time disadvantage.

The following table gives the mean per-step module latency and the fraction of steps in each stage:

Decision stageStep proportionSemantic verificationCandidate rankingStage comparisonStage-decision subtotalAction policyComplete decision
Exploration32.25%0.755 s0.095 s0.361 s1.211 s0.677 s1.888 s
Approach20.33%3.173 s1.305 s2.605 s7.083 s0.670 s7.753 s
Regrounding19.95%4.037 s1.251 s3.724 s9.012 s0.663 s9.675 s
Reacquisition27.47%0.743 s0.105 s0.385 s1.233 s0.783 s2.016 s

Weighted by the observed stage-step proportions, our mean complete decision latency is 4.669 s. Approach and Regrounding require more careful target confirmation and candidate comparison and therefore take longer; Exploration and Reacquisition rely mainly on direction evidence and target memory and have lower overhead. Action-policy inference ranges from 0.663 to 0.783 s across the four stages. The main added cost is stage-dependent semantic reasoning.

We assessed reacquisition against the dataset-labeled true final target and the official success condition:

Reacquisition metricProportion
Reacquisition steps as a fraction of all decision steps27.47%
Navigation episodes with at least one reacquisition89.81%
Episodes reaching the true target after triggering reacquisition68.79%

The target can temporarily leave view during Exploration, Approach, or Regrounding because of occlusion or viewpoint changes, so most trajectories trigger Reacquisition. Under the official success criterion, 68.79% of trajectories that enter Reacquisition ultimately reach the annotated true target. This indicates that reacquisition can restore target-directed navigation after temporary target loss.

To show seed variability, we summarize our results for seeds 41–50 on the same splits and under the same evaluation protocol as mean ± standard deviation, alongside the results reported for AerialVLA. The following tables retain NE, SR, OSR, and SPL from Tables 2–4.

Table 2. Seen split

SettingMethodNE (m) ↓SR (%) ↑OSR (%) ↑SPL (%) ↑
FullTravelUAV98.6617.4548.8715.76
NavFoM93.0529.1749.2425.03
LongFly60.0236.3965.8731.07
AerialVLA65.8847.9657.6938.54
Ours37.96±2.2358.41±1.3779.06±1.5540.62±0.85
EasyTravelUAV66.4020.2651.2318.10
NavFoM58.9832.9153.1627.87
LongFly38.1038.5271.9031.24
AerialVLA43.7649.3061.3037.14
Ours19.87±2.6264.10±2.0085.73±2.1442.36±1.43
HardTravelUAV138.0414.0245.9812.90
NavFoM143.8323.5843.4020.80
LongFly85.2033.9458.9430.88
AerialVLA93.1646.3053.2340.26
Ours60.26±3.8351.39±1.8670.83±2.3938.48±0.65
AuthorsOpenReview ↗

Response part 3

Table 3. Unseen Object split

Full: n = 629; Easy: n = 209; Hard: n = 420.

SettingMethodNE (m) ↓SR (%) ↑OSR (%) ↑SPL (%) ↑
FullTravelUAV118.1122.4246.9020.51
NavFoM108.0429.8347.9927.20
LongFly66.7443.8764.5638.39
AerialVLA61.4556.6064.8646.61
Ours39.09±2.4566.29±1.4681.83±1.6453.56±1.46
EasyTravelUAV86.1224.4049.2822.03
NavFoM70.5132.5450.7229.54
LongFly54.8438.0156.8431.36
AerialVLA45.7256.9464.1143.76
Ours46.04±0.2162.00±1.8372.90±2.0345.51±0.75
HardTravelUAV134.0321.4345.7119.75
NavFoM133.0128.0346.1825.64
LongFly57.0750.2574.1645.27
AerialVLA69.2756.4365.2448.03
Ours35.63±3.6168.43±2.0686.27±2.2557.57±2.12

Table 4. Unseen Map split

Full: n = 958; Easy: n = 561; Hard: n = 397.

SettingMethodNE (m) ↓SR (%) ↑OSR (%) ↑SPL (%) ↑
FullTravelUAV138.804.1820.773.84
NavFoM125.106.3018.955.68
LongFly108.3211.2730.279.32
AerialVLA67.4237.5852.9228.22
Ours54.13±2.4540.71±0.8853.32±0.4231.68±0.94
EasyTravelUAV102.944.6322.824.24
NavFoM102.416.7720.076.04
LongFly78.5612.9634.3110.32
AerialVLA44.9941.8958.4729.72
Ours53.19±3.2045.81±1.4657.14±0.4533.92±1.53
HardTravelUAV189.463.5317.883.28
NavFoM170.585.3615.714.97
LongFly148.109.0224.887.98
AerialVLA99.1131.4945.0926.11
Ours55.47±3.8233.50±0.7547.92±0.8428.52±0.80

Across the three tables, our Full-split mean SR exceeds AerialVLA's reported result by 10.45, 9.69, and 3.13 percentage points on Seen, Unseen Object, and Unseen Map, respectively; the corresponding across-seed standard deviations are 1.37, 1.46, and 0.88 points. All Easy and Hard subsets also show positive mean SR gains. Thus, the multi-seed results provide supporting evidence for the robustness of the full method's gains over AerialVLA.

W2 & Q4: What happens to stage decisions when the direction computed from ground truth is replaced with a noisy estimate or the instruction alone?

Response: Thank you for your comment.

In response to your concern, we additionally tested how direction noise affects navigation. At each step, we added zero-mean Gaussian noise to the target's relative bearing and mapped the perturbed angle to a direction category. Mild, moderate, and severe noise have standard deviations of 10°, 30°, and 60°, respectively. The 'No Hint' condition retains the original task instruction but removes the dynamic direction hint at each step. The results are shown below.

Hint conditionStage error rate (%)
No noise21.33
Mild noise27.54
Moderate noise40.38
Severe noise57.64
No Hint62.57
MethodHint conditionSR (%)OSR (%)NE (m)SPL (%)
TravelUAVNo noise14.2639.50115.5312.95
TravelUAVMild noise12.1235.93116.6510.76
TravelUAVModerate noise10.3817.54136.228.98
TravelUAVSevere noise2.617.86160.651.23
TravelUAVNo Hint1.354.12164.940.51
AerialVLANo noise46.4657.6765.4436.94
AerialVLAMild noise43.7654.6967.0630.15
AerialVLAModerate noise31.7137.5892.2424.23
AerialVLASevere noise8.4713.34129.266.04
AerialVLANo Hint1.755.45134.540.81
Our methodNo noise54.4271.4343.3540.48
Our methodMild noise50.4067.8846.0439.39
Our methodModerate noise34.9244.3480.8725.80
Our methodSevere noise8.9914.96125.216.39
Our methodNo Hint3.157.77131.231.04

As direction noise increases, our stage-error rate rises from 21.33% without noise to 27.54%, 40.38%, and 57.64%; without dynamic hints, it reaches 62.57%. Noisy directions increasingly disrupt stage decisions. At the navigation level, SR and SPL decline for all three methods, yet ours retains the highest SR, OSR, and SPL and the lowest NE in all five conditions. Hint quality affects stage classification and navigation, but our method maintains a relative advantage under these perturbations.

We appreciate your detailed comments, which helped us improve the experimental analysis and presentation. In response, we clarified the training protocol for the fixed-input variant and the provenance of baseline results, and added analyses of stage-wise latency, reacquisition, and noisy direction hints.

We respectfully ask you to reconsider the paper's core contribution: we turn the observation that navigation requires different evidence before and after target discovery into testable stage decisions and stage-specific evidence selection. The design retains a shared action policy while delivering appropriate inputs during search, approach, and recovery after target loss, offering a practical route toward long-range UAV navigation. The added results further support this contribution while clarifying computational cost and the dependence on direction hints. We welcome further questions and sincerely thank you for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comment

We thank all three reviewers for their careful reading and constructive comments. We propose a target-discovery-driven framework for stage-adaptive evidence routing. Once a target has been reliably identified, navigation should shift from environmental exploration to target-centered control; stage-specific inputs to the action policy also reduce distracting redundant information during approach and target relocation. Our method achieves the strongest reported performance on the Full setting of all three TravelUAV test splits. Reviewers YFvw, Uwuz, and qDe6 recognized the rationale for stage-adaptive evidence selection and the value of full-test-set evaluation and controlled comparisons. We summarize their main concerns below.

Reviewer YFvw: Asks about the training protocol for the fixed-input and stage-ablation variants, consistency of direction hints, and the source of baseline results. The reviewer also requests module latency, memory and hardware details, stage-step proportions, reacquisition frequency and success, module ablations, justification for compact policy inputs, stage decisions under noisy or instruction-only hints, and variability across seeds, confidence intervals, and episode-level paired tests.

Reviewer Uwuz: Asks for end-to-end latency and model-call costs relative to the fixed-input baseline; whether clear visual line of sight guarantees collision-free straight-line approach; and failure cases with obstacles and partial occlusion. The reviewer also asks whether the modest gain on Unseen Map is stable and how frequently erroneous stage transitions occur and contribute to failure.

Reviewer qDe6: Asks about reliance on ground-truth direction hints, fairness of baseline training, and robustness to noisy or coarse hints. The reviewer requests comparisons of latency, memory, and parameter counts; explanations for weaker metrics in some subsets; stage accuracy, failure analysis, and seed variability; and real-world or perturbation experiments.

We address these points reviewer by reviewer below. We again thank the reviewers for their valuable suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.