acceptodds
Withdrawn from ICLR 2027

ForeFly: A Dual-Horizon World Action Model for Aerial Vision-Language Navigation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Aerial Vision-Language Navigation (AVLN) requires UAVs to maintain reliable instruction following over long trajectories in complex 3D environments. However, existing AVLN approaches are predominantly reactive or limited to single-horizon prediction, overlooking complementary future cues across different temporal horizons. To address this limitation, we propose ForeFly, a dual-horizon latent world action model that predicts both a proximal future for local continuity and an adaptive route-critical future for long-range guidance. Horizon-specific foresight queries are primed with recent and route-critical visual memories, providing history-aware context for future prediction. To exploit their distinct roles in action generation, we introduce Foresight-Guided Action Refinement (FGAR), which asymmetrically exploits proximal foresight for local action enhancement and route-critical foresight for feature-wise correction and route-level guidance. Experiments on the TravelUAV and UAV-ON benchmarks show that ForeFly consistently outperforms strong baselines across seen and unseen settings, validating the effectiveness of dual-horizon foresight and FGAR learning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Aerial Vision-language Navigation, Vision-Language Navigation, Embodied AI

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_t7ziOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes ForeFly, a dual-horizon latent world action model for Aerial Vision-Language Navigation. Instead of relying solely on reactive policies, ForeFly jointly predicts a proximal future latent for local motion continuity and a route-critical future latent for long-range route guidance. It incorporates a dual-memory foresight query priming mechanism and a Foresight-Guided Action Refinement module to integrate these futures into waypoint generation. Experiments are conducted on the TravelUAV and UAV-ON benchmarks.

Critical strengths

None

Critical weaknesses
  1. Misattribution and Omission of Related Work In Section 2, the authors cite works such as WorldVLA, WAM-Flow, and DreamZero, claiming them as algorithms applying WAMs to the VLN domain. However, these papers primarily target autonomous driving, general robotics, or general VLA frameworks rather than vision-language navigation. This misattribution misleads readers regarding the actual landscape of VLN research. Conversely, the authors largely overlook true recent world model approaches tailored for VLN, such as the VISTA series, SparseVideoNav, NavDreamer, WorldVLN, WNM-3D .... The related work section requires a major overhaul to remove out-of-domain citations and properly engage with genuine VLN world model literature. The core motivation and the third paragraph of the introduction bear a striking resemblance to recent works like SparseVideoNav in their conceptual narrative. Simply transferring future-prediction narratives from ground-based or general video navigation to aerial settings without distinct conceptual advancements diminishes novelty. It also sounds like the sparse future can cover both the Proximal Future Latent and the Route-critical Future Latent to some extent.

  2. Insufficient Justification from Ablation Studies and Absence of Real-world experiment In Table 3, Proximal Only and Route-Critical Only exhibit tradeoffs across seen/unseen splits, but the paper lacks deeper analysis into failure modes when anchor thresholds (τp, τg) drift or fail. In Table 5, the marginal differences between Direct Dual-Future and the full FGAR on certain metrics (e.g., OSR) suggest that the complex feature-wise transformation (γt,βt) and additive branches (Btc) may introduce redundancy or overfitting without proportional performance gains. As for the overfitting concern, the authors should present real-world experiment to showcase their performance and demonstrate the edge compared to the recent works like navdreamer.

Overall, reading through this manuscript gives a strong impression of an AI-generated submission rather than a rigorously crafted human research work, primarily characterized by superficial narratives and sloppy bibliographic grounding. Combined with the severe literature misattribution (falsely claiming non-VLN WAMs/VLAs as VLN while omitting true VLN world models), the heavy conceptual overlap in its motivation and introduction, the unconvincing justification of its complex architectural designs in the ablations, and the complete lack of real-world validation, this paper falls well below the acceptance bar of ICLR. Therefore, I strongly recommend a direct rejection.

Questions for authors

See Weakness

Reviewer_EbWFOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ForeFly, a dual-horizon latent world action model for aerial vision-language navigation. Instead of predicting actions reactively, ForeFly predicts two future latents: a proximal future (the next step) for local continuity, and a route-critical future (the next turning or distance-based anchor) for long-range guidance.

Critical strengths
  1. The paper is well-motivated. Distinguishing proximal and route-critical futures, and predicting them in latent space rather than pixel space, is intuitive for long-horizon aerial navigation.

  2. The method is relatively novel, particularly the dual-horizon foresight design and the asymmetric use of the two horizons in FGAR.

  3. The experiments are extensive. ForeFly improves performance on both the Seen and Unseen splits of TravelUAV, with especially clear gains on the Hard subset, and further improvements on UAV-ON.

  4. The paper is well written and easy to follow.

Critical weaknesses
  1. It is unclear whether the gains come from foresight or from the FGAR module itself. The largest single improvement in Table 3 comes from adding FGAR (Seen SR from 33.08% to 41.75%). In contrast, directly fusing the two future latents improves only marginally over the MLP head (Table 5). Since FGAR adds extra parameters and a second training stage, its gains may stem from a stronger action head rather than from the predicted futures. A control that keeps FGAR but removes the foresight supervision (L_W), or replaces the predicted latents with random or shuffled ones, is needed. This matters because the paper’s central claim, that dual-horizon foresight improves navigation, depends on this attribution.

  2. The ablation setup is inconsistent. The Unseen numbers of the “Baseline” in Table 3 exactly match TravelUAV in Table 1, but its Seen numbers differ from TravelUAV in Table 2 (21.73% vs. 16.10% SR). In addition, many ablation results are multiples of 0.25 (e.g., 17.50, 19.25, 21.75), which suggests that the ablations were evaluated on a smaller subset, while the full model’s Unseen result (28.86%) was not. If different evaluation sets were used, the gains across rows are not directly comparable. No variance across runs is reported either. This undermines the reliability of the component-wise conclusions.

Questions for authors

Please see the weaknesses above.

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

Summary This paper proposes ForeFly, a dual-horizon world action model for Aerial Vision-Language Navigation (AVLN). The method predicts both short-term future visual representations and longer-term representations of route-critical events to support local action decisions and long-range navigation. The authors also introduce a dual-memory foresight query priming mechanism and Foresight-Guided Action Refinement (FGAR) to integrate future information from different time horizons into action prediction. Experiments on TravelUAV and UAV-ON show promising improvements over existing baselines, and ablation studies demonstrate the effectiveness of the main components.

Critical strengths

Clear motivation and a meaningful research problem. The paper addresses the limitations of existing AVLN methods in modeling future states at different time horizons. Combining short-term motion prediction with long-range route guidance is a reasonable idea that directly addresses the challenges of long-horizon aerial navigation.

Competitive experimental results with a clear framework. ForeFly achieves strong performance on both TravelUAV and UAV-ON under seen and unseen settings, with particularly noticeable improvements on long-distance navigation tasks. The paper is also well organized and provides ablation studies and qualitative visualizations to support its main design choices.

Critical weaknesses

Limited methodological novelty. The main components, including dual-horizon future prediction, turning- and distance-based keyframe selection, and the attention, FiLM, and residual operations used in FGAR, are largely based on existing techniques. The contribution mainly comes from combining and adapting these components to AVLN. However, the paper does not clearly explain what makes this design substantially different from existing future prediction and world action models.

Insufficient experimental evidence for the key design choices and future prediction capability. Although the paper evaluates the overall effectiveness of dual-horizon prediction and FGAR, several important comparisons are missing. For example, comparing the adaptive route-critical horizon with a fixed long-term prediction horizon would help justify the proposed design. The paper also lacks quantitative evaluation of future representation prediction quality and analysis across different prediction horizons. The qualitative visualizations alone are not enough to demonstrate whether the route-critical branch reliably captures future states that are useful for long-range navigation. In addition, the extra computational cost introduced by the proposed modules is not reported.

Questions for authors

How does the adaptive route-critical horizon compare with a fixed long-term prediction horizon under the same model architecture and training settings? Could the authors also provide an ablation of the turning-based and distance-based anchor selection strategies?

Could the authors provide quantitative evaluation of the predicted future representations, such as prediction errors at different horizons? This would help clarify whether the route-critical branch can reliably predict useful future information rather than simply improving downstream navigation performance.

What is the additional inference latency and computational overhead introduced by the dual-horizon prediction and FGAR modules compared with the baseline? This would help assess the practical efficiency of the method for UAV navigation.

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

Related papers

Discussion (0)

Sign in to comment.