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

Kunhui Wang; Xintong Zhang; Junyu Gao; Changsheng Xu

cs.CVSep 27, 2026

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

Authors: Kunhui Wang, Xintong Zhang, Junyu Gao, Changsheng Xu

Organizations: State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China · School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences, Beijing, China · Division of Natural and Applied Sciences, Duke Kunshan University, Suzhou, China · Peng Cheng Laboratory, ShenZhen, China

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. The code is available at: https://github.com/kunhuiW/ForeFly

Figures & tables

Figure 1: (a) Reactive policies predict actions without future modeling. (b) Single-horizon methods predict a single future latent, overlooking complementary navigation roles for local continuity and route-level anticipation. (c) ForeFly jointly predicts proximal and route-critical futures, while FGAR integrates them for locally smooth and strategically consistent navigation.

Figure 2: Overview of ForeFly. ForeFly primes dual-horizon foresight with complementary historical memories, predicts proximal and route-critical future latents, Foresight-Guided Action Refinement (FGAR) asymmetrically uses proximal foresight for local enhancement and route-critical foresight for feature-wise correction and route-level guidance.

Method	Seen			Unseen
Method	SR $\uparrow$	OSR $\uparrow$	SPL $\uparrow$	SR $\uparrow$	OSR $\uparrow$	SPL $\uparrow$
AerialVLN Liu et al. (2023)	4.55	16.84	4.24	3.51	20.77	3.41
AOA-F Xiao et al. (2025)	7.87	17.98	4.21	6.45	16.09	3.98
OpenFly Gao et al. (2026)	12.29	26.20	6.83	11.82	25.71	5.64
APEX Zhang et al. (2026a)	12.36	19.10	9.13	16.13	22.58	13.03
Navid Zhang et al. (2024)	14.17	30.08	7.19	10.70	28.51	5.99

Table 4: Results on the UAV-ON dataset.

Method	Seen			Unseen
Method	SR $\uparrow$	OSR $\uparrow$	SPL $\uparrow$	SR $\uparrow$	OSR $\uparrow$	SPL $\uparrow$
AerialVLN Liu et al. (2023)	4.55	16.84	4.24	3.51	20.77	3.41
AOA-F Xiao et al. (2025)	7.87	17.98	4.21	6.45	16.09	3.98
OpenFly Gao et al. (2026)	12.29	26.20	6.83	11.82	25.71	5.64
APEX Zhang et al. (2026a)	12.36	19.10	9.13	16.13	22.58	13.03
Navid Zhang et al. (2024)	14.17	30.08	7.19	10.70	28.51	5.99

Table 4: Results on the UAV-ON dataset.

Method	NE $\downarrow$	SR $\uparrow$	OSR $\uparrow$	SPL $\uparrow$
MLP Head	64.41	33.08	62.91	27.99
Direct Dual-Future*	62.03	34.00	62.50	28.73
FGAR w/o $\mathbf{B}_{t}^{c}$	54.61	38.35	68.73	32.55
FGAR	48.74	41.75	70.25	34.75

Table 5: Ablation on Foresight-Guided Action Refinement (FGAR) Design.

Figure 3: Qualitative Results. (a) Bird’s-eye view of predicted trajectories. (b) 3D trajectory comparison showing improved route consistency of ForeFly. (c) Step-by-step navigation, where route-critical foresight helps ForeFly anticipate the upcoming turn, while the baseline deviates. The corresponding future visualization procedure is described in Sec. B.2 .

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Visualization of route-critical anchor construction. Representative UAV trajectories of different lengths are shown. Green and black stars mark the start and end states, red stars indicate key frames at turning points, and yellow diamonds denote distance anchors. The color scale represents the Euclidean distance from the starting position.

Figure 5: Results on TravelUAV Unseen Map and Unseen Object subsets. The top row compares ForeFly with existing methods on the Unseen Map subset, and the bottom row reports results on the Unseen Object subset.

Figure 6: Qualitative visualization of dual-horizon future prediction. Each row shows two sampled waypoints from the same navigation trajectory. For each waypoint, we visualize the current observation together with the ground-truth and predicted proximal and route-critical futures.

Explore similar work

DreamFly: Causal Memory and Receding-Horizon Diffusion Planning for Aerial Vision-Language Navigation
Aug 12, 2026Yan Deng, Fei XuAerial Vision-Language Navigation Perception Pipeline
School of Electronic Information Engineering, Xi’an Technological University, Xi’an, China · School of Computer Science and Engineering, Xi’an Technological University, Xi’an, China
Think Like a Pilot: Fine-Grained Long-Horizon UAV Navigation
Jun 5, 2026Xiangyi Zheng, Xiangyu Wang, Qinan Liao +6Aerial Vision-Language Navigation Pilot
Colab, Beihang University · National University of Singapore · Meituan
ImagineUAV: Aerial Vision-Language Navigation via World-Action Modeling and Kinodynamic Planning
May 31, 2026Xuchen Liu, Jiawei Huang, Shihao Xia +3Aerial Vision-Language Navigation Vision-Language Navigation
Pengcheng Laboratory, Shenzhen, Guangdong, China · School of Computer Science and Cyber Engineering, Guangzhou University, Guangzhou, Guangdong, China · Southern University of Science and Technology, Shenzhen, Guangdong, China

Aug 12, 2026, cs.CVJ/K move · Enter open · S save

DreamFly: Causal Memory and Receding-Horizon Diffusion Planning for Aerial Vision-Language Navigation

Yan Deng, Fei Xu

School of Electronic Information Engineering, Xi’an Technological University, Xi’an, China · School of Computer Science and Engineering, Xi’an Technological University, Xi’an, China

Aerial Vision-Language Navigation Perception Pipeline Navigation Temporal Reasoning

Open paper AskPDF arXiv

Aerial vision-language navigation (VLN) requires an embodied agent to integrate visual evidence over time, plan future actions, and determine when it has reached a navigation goal under partial observability. Although recent VLA models offer a promising perception-to-action paradigm, adapting them to aerial navigation remains challenging due to limited historical context, short planning horizons, and unreliable implicit termination. To address these challenges, we propose DreamFly, a diffusion-based aerial VLN framework built on Dream-VLA. DreamFly introduces a causally aligned historical memory that augments the current visual representation using only observations preceding the current decision step, enabling temporal reasoning without future information leakage. We further formulate navigation as receding-horizon diffusion planning, where the policy predicts a

K

-step action chunk but executes only the first action before replanning. This plan-

K

, execute-one strategy uses future actions as auxiliary planning targets while preserving closed-loop visual feedback. Finally, LiteStop estimates the stop probability directly from action logits at the initial all-mask state, decoupling explicit termination from action generation. Experiments on the OpenFly benchmark demonstrate consistent improvements in seen and unseen environments. DreamFly achieves 32.04%/29.46% SR and 28.22%/23.54% SPL on the test-seen/test-unseen splits, respectively, outperforming all compared methods on both metrics while attaining the lowest navigation error. These results demonstrate the effectiveness of jointly modeling historical context, future action structure, and explicit termination for aerial VLN.

Abstract

Explore similar work

DreamFly: Causal Memory and Receding-Horizon Diffusion Planning for Aerial Vision-Language Navigation

Think Like a Pilot: Fine-Grained Long-Horizon UAV Navigation

ImagineUAV: Aerial Vision-Language Navigation via World-Action Modeling and Kinodynamic Planning