Preprint · 2026

VoLN Vision-Only Long-Horizon Navigation—Paradigm, Benchmark, and Method

Jiabin Lou Haopeng Wang Yuanshuai Wang Xinyu Liu Xuxin Lv
Yuxin Guo Lei Huang Rongye Shi Wenjun Wu*

Beihang University, Beijing, China

* Corresponding author

Paper Code Dataset Environments Demo
Comparison between conventional language-guided VLN and the vision-only VoLN setting
VoLN replaces privileged language and route priors with visual evidence. The goal is specified by terminal views, while route intent must be recovered from in-scene semantic beacons, egocentric observations, and onboard sensing.

Abstract

Vision-and-Language Navigation (VLN) enables embodied agents to follow natural-language instructions. However, route-level instructions commonly encode spatial priors, such as orientation, distance, and layout, that are not explicitly available from onboard sensing at deployment in open, GPS-denied environments. Benchmark performance under such interfaces therefore jointly reflects visual navigation ability and the use of instruction-derived route structure, making perception-driven navigation difficult to isolate. To separate these factors, we propose Vision-Only Long-Horizon Navigation (VoLN), a task formulation in which goal views specify the destination and route-relevant guidance is presented as locally observable in-scene semantic beacons. During execution, the interface provides no externally supplied episode-level text, global maps, GPS, or shortest-path information. We instantiate VoLN for aerial navigation and introduce VoLN-UAV, a benchmark that combines long-horizon goal-directed flight, continuous 3D motion, large viewpoint changes, and context-dependent beacon selection. As an initial method, VoLN-MLLM aligns self-supervised visual features with a structured semantic space and predicts short-horizon waypoint segments from observation history, goal views, retrieved visual-semantic tokens, and proprioception. On Test-Unseen, it achieves success rates of 7.4%, 4.5%, and 1.8% on Easy, Normal, and Hard episodes, respectively. These results provide an initial evaluation of VoLN and reveal substantial remaining challenges in long-horizon evidence integration, cross-view goal matching, and closed-loop stability.

VoLN-UAV Benchmark

A scene-disjoint benchmark for long-horizon UAV navigation across diverse natural and built environments.

7,210Episodes
5,047Train
1,082Validation-Seen
1,081Test-Unseen
Representative natural and built environments in VoLN-UAV
Environment coverage. Representative scenes include desert, forest, mountain, urban canyon, night urban, campus, tunnel, industrial corridor, and open-site conditions.
VoLN-UAV trajectory collection, beacon augmentation, annotation, and data split pipeline
Dataset construction. Preset and operator-collected trajectories are augmented with route-aware beacons, synchronized step-level annotations, and scene-source splits for unseen-environment evaluation.

Difficulty: Easy < 300 m · Normal 300–450 m · Hard ≥ 450 m, measured by reference path length.

VoLN-MLLM

Visual-semantic alignment followed by closed-loop trajectory planning.

The two-stage VoLN-MLLM method framework
Method overview. Phase I maps frozen DINOv3 features into the CLIP image-embedding space through a lightweight adapter. Phase II retrieves visual-semantic tokens and predicts an executable local waypoint sequence from the aligned observation, terminal goal views, and proprioception.

Implementation. VoLN-MLLM uses a frozen DINOv3 ViT-B/16 visual backbone and a frozen Vicuna-7B-v1.5 planner. Top-8 category tokens are retrieved from the fixed semantic bank, rank-16 LoRA modules adapt Vicuna's attention and feed-forward projections, and learned heads output eight 3D waypoints plus a stop signal.

Benchmark Results

Test-Unseen performance under the unified visual-goal observation and waypoint-action protocol.

Method NE (m) ↓ SR (%) ↑ OSR (%) ↑ nDTW (%) ↑ SPL (%) ↑
Random 270.1 / 310.4 / 395.2 0.4 / 0.0 / 0.0 1.4 / 0.6 / 0.2 30.1 / 22.7 / 15.1 0.3 / 0.0 / 0.0
Seq2Seq-VG 208.6 / 254.8 / 309.9 1.0 / 0.4 / 0.1 4.8 / 2.5 / 0.9 28.9 / 21.4 / 13.0 0.7 / 0.3 / 0.0
CMA-VG 174.5 / 216.8 / 266.1 1.6 / 0.8 / 0.2 6.5 / 3.9 / 1.7 33.2 / 26.4 / 18.5 1.1 / 0.6 / 0.1
LAG-VG 122.4 / 158.3 / 206.7 2.3 / 1.2 / 0.4 6.4 / 3.8 / 1.7 28.1 / 20.5 / 14.0 1.5 / 0.7 / 0.2
VoLN-MLLM 97.1 / 131.4 / 176.8 7.4 / 4.5 / 1.8 14.6 / 10.1 / 4.5 53.1 / 41.2 / 28.0 5.7 / 3.2 / 1.3

Values are ordered as Easy / Normal / Hard and follow the current manuscript.

Qualitative Analysis

Active-beacon selection and passive-distractor rejection determine whether local decisions remain aligned with the route.

Aligned successful and failed rollouts for beacon discrimination
Beacon-discrimination case study. The successful rollout ignores passive distractors and re-identifies the goal. The failed rollout commits to a passive cue, produces an incorrect local update, and accumulates drift.

Simulation and Real-World Demonstrations

The physical testbed is a preliminary qualitative feasibility study rather than a large-scale quantitative real-world benchmark.

Physical indoor VoLN testbed with trajectory and beacon overlays
Physical indoor testbed. The overlay marks the UAV, executed trajectory, next visual beacon, and goal region under the same visual-goal interface.

Simulation

Open MP4

Physical flight

Open MP4

Resources

The navigation dataset provides sharded trajectories, RGB observations, annotations, and metadata. The simulator package provides the Unreal Engine and AirSim environments for online evaluation. Training, benchmark construction, VoLN-adapted baselines, and evaluation scripts are available in the code repository.

BibTeX

If you find this work helpful, please consider citing our paper.

@article{lou2026voln,
  title  = {VoLN: Vision-Only Long-Horizon Navigation---Paradigm, Benchmark, and Method},
  author = {Lou, Jiabin and Wang, Haopeng and Wang, Yuanshuai and Liu, Xinyu and Lv, Xuxin and Guo, Yuxin and Huang, Lei and Shi, Rongye and Wu, Wenjun},
  journal = {arXiv preprint arXiv:2607.21400},
  year   = {2026},
  eprint = {2607.21400},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url    = {https://arxiv.org/abs/2607.21400}
}