Organizations: Tianmushan Laboratory, Beihang University, Hangzhou 311115, China · Hangzhou International Innovation Institute, Beihang University, Hangzhou 311115, China · Foshan Graduate School of Innovation, Northeastern University, Foshan, China · Faculty of Robot Science and Engineering, Northeastern University, Shenyang 110819, China · School of Aeronautic Science and Engineering, Beihang University, Beijing 100191, China · qingniaoAI
Existing aerial Vision-Language Navigation (VLN) methods predominantly adopt a detection-and-planning pipeline, which converts open-vocabulary detections into discrete textual scene graphs. These approaches are plagued by inadequate spatial reasoning capabilities and inherent linguistic ambiguities. To address these bottlenecks, we propose a Visual-Spatial Reasoning (ViSA) enhanced framework for aerial VLN. Specifically, a triple-phase collaborative architecture is designed to leverage structured visual prompting, enabling Vision-Language Models (VLMs) to perform direct reasoning on image planes without the need for additional training or complex intermediate representations. Comprehensive evaluations on the CityNav benchmark demonstrate that the ViSA-enhanced VLN achieves a 70.3% improvement in success rate compared to the fully trained state-of-the-art (SOTA) method, elucidating its great potential as a backbone for aerial VLN systems.
Figures & tables
Fig. 1: Overall architecture of ViSA . The system iterates through three phases per time step: (1) VPG transforms raw observations into SoM-annotated visual representations; (2) VM performs Three-Stage Verification reasoning and emits task primitives; (3) Executor translates semantic decisions into UAV motion. Closed-loop feedback from VM to VPG enables adaptive detection across iterations.
Method
Easy
Medium
Hard
NE ↓
SR ↑
OSR ↑
SPL ↑
NE ↓
SR ↑
OSR ↑
SPL ↑
NE ↓
SR ↑
OSR ↑
SPL ↑
Random
340.62
0.00
0.00
0.00
548.30
0.00
0.00
0.00
654.26
0.00
0.00
0.00
Greedy
105.63
1.55
3.15
0.97
103.00
0.00
4.88
0.00
78.80
0.00
2.71
0.00
GPT-4o
301.45
0.00
7.50
0.00
327.25
0.00
6.25
0.00
401.63
0.00
0.00
0.00
Qwen3-VL-PLUS
838.33
0.49
7.54
0.03
836.11
0.00
2.49
0.00
876.23
0.24
1.65
0.17
GeoNav [ 7 ]
59.86
26.53
73.47
12.05
53.80
22.92
39.58
17.06
68.90
16.67
22.92
12.49
TABLE I: Performance comparison of zero-shot methods on the CityNav dataset across three difficulty levels. Bold indicates best, underline indicates second-best.
Method
NE ↓
SR ↑
OSR ↑
SPL ↑
Random
208.80
0.00
1.44
0.00
Seq2Seq
174.50
1.73
8.57
1.69
CMA
179.10
1.61
10.07
1.57
MGP
93.80
6.38
26.04
6.08
FlightGPT [ 22 ]
76.20
21.20
35.38
19.24
ViSA-enhanced VLN (Ours)
45.73
36.11
43.37
27.31
TABLE II: Performance comparison of trained methods on the CityNav Test-Unseen split. Bold indicates best, underline indicates second-best.
Fig. 2: Visualization of the VLM’s explicit reasoning trace during a navigation episode in the CityNav environment.ViSA demonstrates robust spatial reasoning by accommodating a flawed spatial preposition (“underneath” instead of “on”). The sequence illustrates the logical rejection of an initial false-positive red car ① located in front of the Tram Depot ③, followed by closed-loop guided search that successfully locates and verifies the correct target positioned behind the depot.
Method
V
R
D
E
NE ↓
SR ↑
OSR ↑
SPL ↑
ViSA
✓
✓
✓
✓
47.75
30.19
38.39
19.35
w/o R
✓
✓
✓
77.68
20.14
29.03
13.02
w/o V
✓
✓
✓
54.30
20.83
33.19
10.44
w/o R+V
✓
✓
63.24
10.83
24.17
5.90
w/o D
✓
✓
✓
57.48
20.56
22.50
17.74
w/o E
✓
✓
✓
88.90
9.51
10.73
8.99
TABLE III: Ablation study on Val-Seen Easy. ✓ indicates the component is included. Abbreviations: V : Visual Prompting, R : Verification Reasoning, D : Dual-stage VLM, E : Executor.
Foshan Graduate School of Innovation, Northeastern University, Foshan 528311, China · College of Information Science and Engineering, Northeastern University, Shenyang 110819, China · Faculty of Robot Science and Engineering, Northeastern University, Shenyang 110819, China +3