acceptodds
Under review as a conference paper at ICLR 2027

: Towards Efficient Closed-Loop Aerial Vision-Language Navigation via Visual Compression

Abstract

Aerial vision-language-action (VLA) models incur substantial computational costs from visual feature extraction and downstream processing of dense visual token sequences, limiting navigation decision rates on resource-constrained unmanned aerial vehicles (UAVs). Existing approaches reduce these costs through encoder compression or token pruning, but optimizing either stage alone leaves the other source of computation largely unchanged. Moreover, token selection based solely on current-frame information can overlook cues needed for subsequent navigation, as flight actions change the viewpoint and available observations. We introduce Aerial Closed-Loop Navigation via Visual Compression (), a framework that addresses both computational costs while using navigation context to guide visual information retention. Complementary-teacher distillation transfers spatial and semantic features from two teacher encoders into a single lightweight student, which replaces both teachers during inference. A state-adaptive token allocator then uses navigation context and recent control history to select visual tokens, adjusts the token budget according to observation uncertainty, and applies role-aware retention constraints. The VLA backbone processes only the selected tokens, reducing downstream multimodal computation. We evaluate on the Seen, Unseen Object, and Unseen Map splits of TravelUAV. Compared with the dense VLA baseline, reduces FLOPs by 40.54% and mean end-to-end latency by 27.72% while also improving navigation performance. On the Seen split, it achieves a success rate (SR) of 51.48% and success weighted by path length (SPL) of 41.02%, representing improvements of 3.95% and 3.18%.

Primary area:
applications to robotics, autonomy, planning
Keywords:
UAV, vision-language-action model, robot learning

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_6P3YOpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper introduces AeroC2AeroC^2, a visual compression framework for efficient closed-loop aerial vision-language navigation. The framework combines complementary-teacher distillation, which transfers spatial and semantic knowledge into a lightweight visual encoder, with state-adaptive token allocation, which dynamically retains visual evidence relevant to navigation and subsequent control. Experiments on TravelUAV show that AeroC² maintains strong navigation performance across seen and unseen environments while reducing FLOPs by 40.54% and mean end-to-end latency by 27.72% compared with the dense VLA baseline.

Critical strengths

The paper achieves strong empirical results, consistently improving navigation performance across the Seen, Unseen Object, and Unseen Map splits while substantially reducing model parameters, FLOPs, and inference latency. The two-stage framework is reasonably innovative and well integrated: complementary-teacher distillation reduces visual encoding cost, while state-adaptive token allocation decreases downstream multimodal computation. The overall system is complete and practical, and the component-level ablations generally support the effectiveness of its main modules. Overall, AeroC2AeroC^2 provides a favorable trade-off between navigation accuracy and computational efficiency.

Critical weaknesses

Insufficient support for future-control-aware visual compression. The paper argues that existing frame-local token compression overlooks the “future observability induced by each action” and proposes to retain visual evidence relevant to “subsequent control.” However, AeroC2AeroC^2 does not explicitly model how an action changes the next viewpoint or which visual tokens will benefit subsequent decisions. Its action-conditioned evidence only adjusts spatial importance using recent actions, while temporal memory accumulates previous evidence scores and retention decisions. Moreover, these history-dependent components are disabled during shuffled-frame Stage-II training and introduced as deterministic heuristics during inference. The experiments in Table 4 demonstrate the benefits of multi-cue evidence scoring, adaptive budgeting, and role-aware retention, but do not separately evaluate action-conditioned evidence or temporal memory. Figure 3 only visualizes the changing token budget KtK_t. Therefore, the results support state-adaptive token allocation, but do not sufficiently validate the claimed consideration of future observability and subsequent control.

The multi-property distillation objective is insufficiently validated. Although the proposed objective combines teacher-branch features, projected features, feature scale, and spatial relations, the ablation in Table 3 only compares different teacher configurations. It demonstrates the benefit of dual-teacher supervision, but does not isolate the contribution of the four losses. Therefore, the necessity of the four-part distillation objective remains unclear.

Table 3 does not sufficiently justify the use of DINOv2 as an additional spatial teacher. Although the dual-teacher configuration outperforms the SigLIP-only variant, the latter distills the DINOv3 student solely toward the SigLIP feature space. It does not evaluate the more natural alternative of preserving the native DINOv3 representation while using SigLIP only to provide complementary semantic supervision. Therefore, the observed performance gap may result from the loss of DINOv3’s native spatial information under SigLIP-only distillation, rather than demonstrating that DINOv2 contributes indispensable spatial knowledge. The experiment only establishes that the proposed dual-teacher objective is better than single-teacher SigLIP distillation under the authors’ formulation.

Questions for authors
  1. How is “future observability induced by each action” concretely modeled in AeroC²? The action-conditioned evidence uses previous actions rather than predicting how the current action will change the next viewpoint, while the remaining cues are primarily computed from the current observation.

  2. Temporal memory updates token-wise scores at corresponding grid indices across frames. Since UAV motion continuously changes the viewpoint, how are visual tokens spatially aligned between consecutive observations? Without explicit geometric correspondence or motion compensation, could the same token index refer to unrelated scene regions?

  3. Why are the action-conditioned and temporal-memory components disabled during shuffled-frame Stage-II training but activated during sequential inference? Does this introduce a training-inference mismatch, and can the authors separately quantify the contributions of these two history-dependent components?

  4. What are the individual contributions of Lbranch\mathcal{L}_{branch}, Lproj\mathcal{L}_{proj}, Lscale\mathcal{L}_{scale}, or Lrel\mathcal{L}_{rel}? Can the authors provide loss-level ablations to justify the complete multi-property distillation objective?

  5. Why is DINOv2 required as an additional spatial teacher for a pretrained DINOv3 student? In particular, how does the proposed dual-teacher design compare with preserving DINOv3’s native representation while using SigLIP only as complementary semantic supervision?

1 comment
Reviewer_6P3YOpenReview ↗

Response to the Authors

Thank you for the detailed response and the substantial additional experiments. The history-term ablations address my concern about the contributions of action-conditioned evidence and temporal memory. The loss-level ablations also validate the four components of the distillation objective. In addition, the explanation of the OpenVLA visual interface and the Native DINOv3 plus SigLIP comparison provide much stronger justification for the dual-teacher design.

My main remaining concern is the interpretation of future observability induced by each action. At step t+1t+1, the action from step tt has already been executed and the resulting observation has already been acquired. Conditioning token retention on the previous action therefore appears to be history-conditioned selection of current-frame tokens rather than prospective modeling of how a candidate action will affect future viewpoints or which visual evidence has long-term value for subsequent control. The new ablations convincingly show that these history-based heuristics improve navigation, but they do not fully substantiate the stronger claim of explicitly modeling future observability.

Overall, the rebuttal substantially improves my confidence in the empirical soundness of the method. My remaining concern mainly relates to the conceptual positioning and methodological novelty, since the framework still combines multi-teacher distillation with a deterministic allocator composed of manually designed cues, coefficients, thresholds, and retention safeguards. Nevertheless, given the additional evidence and clarifications, I am inclined to raise my score while retaining this concern about the future-control-aware claim and the incremental nature of the overall design.

Reviewer_z65AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AeroC$^2$ to reduce visual computation in aerial vision-language navigation through encoder distillation and adaptive token selection. DINOv2 and SigLIP features are distilled into a smaller DINOv3 encoder. A rule-based allocator then selects tokens using language, spatial cues and recent navigation history. Experiments on three TravelUAV splits report improved navigation performance, with 40.54% fewer FLOPs and 27.72% lower mean inference latency than the dense baseline.

Critical strengths
  1. The motivation is clear. Reducing both visual encoding cost and downstream token processing is useful for aerial VLA deployment.

  2. The model improves SR and SPL over the dense baseline on all three splits while reducing measured inference cost. Evaluation covers the full benchmark routes, including unseen objects and maps.

  3. The component comparisons are reasonably controlled. Table 2 separates the two compression stages, and Table 4 compares adaptive allocation with a fixed budget at the same average token count.

Critical weaknesses
  1. There is no comparison with existing token-pruning methods, such as VLA-Pruner[1] or LightVLA[2]. VLA-Pruner already uses temporal action information, so it is particularly relevant to the claimed advantage over frame-local pruning. Without this comparison, the value of the proposed aerial-specific rules is hard to judge.

  2. The temporal memory averages scores at fixed image locations without accounting for UAV motion (Eqs. 24–25). After a turn, the same location can correspond to entirely different content. The paper does not isolate whether this memory actually helps preserve useful evidence beyond the static spatial priors and view quotas.

  3. The adaptive budget relies on heuristic uncertainty signals whose meaning is not well established. For example, front and down views naturally contain different information, so different semantic responses do not necessarily indicate uncertainty. The fixed-budget comparison is useful, but does not show whether the proposed uncertainty score identifies decisions that benefit from more tokens.

  4. The deployment motivation is not fully tested. Latency is measured on an RTX 4090, with no onboard or real-flight evaluation, and it is unclear whether inference delay affects UAV motion during the navigation experiments. The results therefore do not establish how the measured speedup changes continuous-flight performance.

[1] Liu et al., VLA-Pruner: Temporal-Aware Dual-Level Visual Token Pruning for Efficient Vision-Language-Action Inference. arXiv:2511.16449 [2] Jiang et al., The Better You Learn, The Smarter You Prune: Towards Efficient Vision-language-action Models via Differentiable Token Pruning. arXiv:2509.12594

Questions for authors
  1. Can you compare with VLA-Pruner or LightVLA using the same encoder and policy, at comparable token budgets or measured latency?

  2. What happens when action-history and memory cues are disabled at inference, as they are during training? How does the fixed-grid memory behave after large viewpoint changes?

  3. Does the uncertainty score predict the benefit of allocating additional tokens? Would a simpler search/landing budget schedule perform similarly at the same average token count?

  4. Does AirSim continue evolving while inference runs? Do success rate or flight time improve when the actual inference delay is included, and are any onboard timings available?

1 comment

Thank you for the detailed rebuttal and additional experiments, which address several of my concerns.

However, I still have concerns about the fairness of the baseline comparisons. In particular, LightVLA performs notably much worse than random selection, which raises questions about its adaptation and training. The actual token budgets and latency of the compared methods are also not reported.

The memory ablation shows an overall benefit, but does not address its reliability under large viewpoint changes.

Given these remaining concerns, I will maintain my rating of weak rejection.

Reviewer_CVKnOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies visual computational efficiency for closed-loop aerial vision-language-action (VLA) navigation. It introduces AeroC2, which combines complementary-teacher distillation from DINOv2 ViT-L/14 and SigLIP ViT-So400M/14 into a DINOv3 ViT-B/16 student with a parameter-free, state-adaptive token allocator. The allocator combines seven hand-crafted evidence cues, an observation-uncertainty budget schedule, and role-aware front-view and down-view retention quotas before fine-tuning the projector and OpenVLA-7B LoRA adapters. Across the Seen, Unseen Object, and Unseen Map splits of TravelUAV, the paper reports 40.54% lower FLOPs, 27.72% lower mean latency, and 3.52% to 4.80% higher success rates than AeroVLA.

Critical strengths
  1. The paper addresses a practical deployment bottleneck in aerial VLA navigation: both dual-encoder feature extraction and dense visual token sequences limit onboard control rates. Targeting both stages within a single pipeline is a sensible systems direction.
  2. Partitioning retention quotas across front-view and down-view observations aligns with UAV sensing geometry. Reserving view-specific coverage and terminal down-view capacity provides a structured way to prevent single-view collapse during approach and landing.
  3. The evaluation covers all 3,005 routes across the Seen, Unseen Object, and Unseen Map splits of TravelUAV in closed-loop AirSim simulation. Reporting both navigation accuracy and per-decision efficiency metrics broadens the empirical comparisons.
  4. The submission documents the distillation losses, cue equations, quota rules, and hyperparameter tables in Appendices A and B. Including qualitative budget schedules and difficulty breakdowns in Appendix D.2 adds useful context.
Critical weaknesses
  1. The navigation gains are confounded by a stronger student backbone and higher input resolution in Stage I. In Table 2, Visual distillation only (retaining all 256 tokens without token allocation) already increases Unseen Map SR from 37.16% (Dense VLA) to 42.69%, exceeding the complete AeroC2 model (42.38%), whereas Token allocation only reaches 37.47%. Appendix B.1 indicates that while the DINOv2 and SigLIP teachers process 224x224 inputs, the student switches to a newer DINOv3 ViT-B/16 backbone pretrained on LVD-1689M and increases the input resolution to 256x256. Consequently, the headline generalization improvement over Dense VLA on Unseen Map is driven by the upgraded student backbone and higher input resolution rather than by the proposed token allocation mechanism.
  2. The closed-loop token allocation claim suffers from a train–inference mismatch and lacks cue-level ablations. Section 3.3 and Appendix B.2 state that during Stage-II training on shuffled frames, all history-dependent components—temporal memory, previous-motion inputs, motion uncertainty, and motion terminal priors—are set to zero. As a result, the VLA backbone is trained exclusively under static frame-local priors and encounters closed-loop cues only at test time. Furthermore, Table 4 does not ablate the seven scoring cues or compare inference with and without the history-dependent terms, leaving the actual contribution of closed-loop state feedback unverified.
  3. The allocator's benefit is not isolated from extensive manual tuning or compared against existing token-pruning methods. Appendices A.3–A.8 and Table 6 specify more than 40 fixed weights, Gaussian spatial priors, thresholds, and manually chosen grid coordinates (such as the four extra anchor positions in Section A.8), but the manuscript does not describe a validation split or sensitivity analysis supporting these choices. In addition, although Sections 1 and 2 repeatedly critique frame-local token pruning methods (e.g., Fang et al., 2026; FastV; ToMe; VisionZip; LightVLA), Table 4 compares the allocator only against naive random selection.
  4. The efficiency reductions in Table 2 are not internally consistent across stages. Because Stage-I feature extraction and multimodal projection always operate on all 256 tokens before Stage-II token selection (Appendix C, Eqs. 43–44), replacing the dual teachers with the student pathway should yield a constant latency and FLOPs reduction regardless of the downstream token budget. However, Stage-I distillation reduces mean latency by 62.01 ms (472.94 to 410.93 ms) and FLOPs by 0.32 T (5.18 to 4.86 T) at 256 tokens, but reduces latency by only 12.51 ms (354.37 to 341.86 ms) and FLOPs by 0.40 T (3.48 to 3.08 T) when combined with token allocation at 126 tokens. Appendix D.1 also estimates adaptive FLOPs by replaying bucket representatives rather than profiling full rollouts.
  5. Figures 1 and 3 have readability issues. In Figure 1, the text inside the bottom-left sub-panels illustrating compact visual encoding and state-adaptive token allocation is very small and difficult to read. In Figure 3, the axis labels, legends, and instruction text in the rollout panels are similarly cramped; Enlarging the typography and simplifying the inset labels would make the visual presentation clearer.
Questions for authors
  1. How does Dense VLA perform when using DINOv3 ViT-B/16 at 256x256, or how does Stage-I distillation perform when using a same-generation ViT-B/14 student at 224x224? Clarifying this is necessary to separate distillation from the backbone and resolution upgrade.
  2. Since all history-dependent cues (temporal memory, previous-motion inputs, motion uncertainty, and motion terminal priors) are zeroed out during Stage-II training, what are the test results when these history-dependent components remain disabled during inference? Please also provide ablations for the individual scoring cues in Eq. (6).
  3. On what data split were the 40+ constants and thresholds in Table 6 selected, and how sensitive is performance to them? At a matched budget of K = 126, how does the allocator compare against existing frame-local pruning baselines (such as FastV, ToMe, or top-K selection using only instruction relevance and view quotas)?
  4. Why does replacing the dual teacher encoders with the fixed 256-token student pathway reduce latency by 62.01 ms (and 0.32 T FLOPs) without token allocation, but by only 12.51 ms (and 0.40 T FLOPs) when token allocation is enabled downstream of the projector? Please provide a component-wise latency and FLOPs breakdown across Table 2.
1 comment

Thank you for the detailed rebuttal and additional results. The controlled comparisons with existing token-pruning baselines, the cue-level ablations, and the clarification on the DINOv3 distillation setup resolve my main concerns. Although the token allocator remains fairly heuristic-heavy, the overall efficiency and navigation improvements are convincing. I am raising my rating to 3.

Comments on the paper

1 comment
Reviewer_6P3YOpenReview ↗

Response to the Authors

Thank you for the additional clarification. I now better understand that the intended contribution is action-conditioned visual retention within the closed-loop action-to-observation process. The previous action provides motion context for selecting tokens from the resulting observation before the VLA predicts the next action, while temporal memory and terminal protection preserve evidence across successive decisions.

I still view this mechanism as reactive action-conditioned retention rather than prospective modeling of future observability. The allocator does not predict the observation resulting from a candidate action or estimate the long-term value of current visual evidence before that action is executed. Nevertheless, I agree that the clarification provides a more precise interpretation of the claimed closed-loop mechanism, and the additional history-term ablations demonstrate that incorporating the preceding action and temporal scores provides consistent empirical benefits.

The explanation of the system-level novelty is also clearer. Although the individual components remain largely based on established distillation, token selection and manually designed retention principles, their integration addresses both visual encoding and downstream token-processing costs within an aerial VLA. The new perturbation analysis further reduces my concern that performance depends strongly on finely tuned allocator constants.

Considering the complete set of additional ablations, alternative comparisons and clarifications provided during the discussion, my original soundness concerns have been substantially addressed. I have therefore updated my score accordingly, while retaining a minor concern about the future-observability terminology and the primarily systems-oriented nature of the methodological contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.