BGC-Nav: Cross-Altitude Multi-UAV Collaborative Navigation via Belief-Guided Coupling

Jiuwu Hao; Yueyang Wu; Liguo Sun; Haolin Song; Zao Zhang; Pin Lv

cs.ROOct 10, 2026

BGC-Nav: Cross-Altitude Multi-UAV Collaborative Navigation via Belief-Guided Coupling

Authors: Jiuwu Hao, Yueyang Wu, Liguo Sun, Haolin Song, Zao Zhang, Pin Lv

Organizations: Institute of Automation, Chinese Academy of Sciences · School of Artificial Intelligence, University of Chinese Academy of Sciences

Abstract

Despite recent advances in Aerial Vision-and-Language Navigation (VLN), single Unmanned Aerial Vehicle (UAV) search-and-rescue in outdoor scenarios remains challenging due to partial observability, making cross-altitude collaboration a promising direction. However, existing collaborative navigation methods rely on precise target localization and rigid goal-dispatching paradigms, ignoring high-altitude observation uncertainty and suppressing low-altitude exploration autonomy. To address these issues, we introduce BGC-Nav, a training-free cross-altitude collaboration framework that enables robust navigation via Belief-Guided Coupling (BGC). Specifically, the high-altitude UAV maintains a Global Belief Memory encoding task-conditioned priors, which the low-altitude agent incorporates into decision-making via Local Exhaustion Memory (LEM), Local Attraction Augmentation (LAA), and Directional Belief Prior (DBP) as an advisory bias rather than a rigid command. Extensive experiments on ESARBench show that BGC-Nav consistently outperforms state-of-the-art single-agent and collaborative methods. Code is available at https://github.com/uestchjw/BGC-Nav.

Figures & tables

Figure 1: Overview of BGC-Nav. Following role-specific instruction decomposition, the high-altitude UAV maintains a Global Belief Memory encoding task-conditioned priors. The low-altitude UAV queries the Global Belief Memory to extract a local belief map, discounts it via a Local Exhaustion Memory, and modulates its attraction map using Local Attraction Augmentation and Directional Belief Prior for subsequent decision-making.

Mode	Method	ESAR-short			ESAR-long
Mode	Method	SR $\uparrow$	OSR@40 $\uparrow$	RS $\uparrow$	SR $\uparrow$	OSR@40 $\uparrow$	RS $\uparrow$
Single-UAV	FBE [ 17 ]	9.86	9.72	11.30	4.34	5.03	10.03
	VLFM [ 19 ]	12.82	14.74	12.87	5.70	6.56	9.97
	SPF [ 4 ]	12.19	11.37	9.09	3.31	3.01	8.24
	UniGoal [ 18 ]	4.65	6.46	8.47	3.13	1.89	7.14
	APEX [ 22 ]	12.61	18.20	14.68	4.52	2.58	10.93

Table 1: Performance comparison on ESARBench across ESAR-short (

<

200 m) and ESAR-long (

\geq

200 m) subsets, split by start-to-target distance.

Reliability	Method	SR $\uparrow$	RS $\uparrow$	OSR@30 $\uparrow$	OSR@60 $\uparrow$
Low	APEX [ 22 ]	3.70	10.60	3.70	7.41
	AeroDuo [ 14 ]	1.85 1.85↓	8.71 1.89↓	0.10 3.60↓	0.93 6.48↓
	BGC-Nav	3.09 0.61↓	10.89 0.29↑	4.32 0.62↑	10.19 2.78↑
Mid	APEX [ 22 ]	13.87	16.60	3.65	14.37
	AeroDuo [ 14 ]	11.08 2.79↓	15.17 1.43↓	3.15 0.50↓	14.59 0.22↑
	BGC-Nav	12.03 1.84↓	15.76 0.84↓	7.66 4.01↑	16.53 2.16↑

Table 2: Performance comparison under varying levels of high-altitude guidance reliability.

Panel of Figure 2 — Figure 2: Qualitative comparison of navigation trajectories. Unlike the dispatch-and-reach collaborative method (AeroDuo) that rigidly enforces target waypoints, BGC-Nav maintains robust target search performance by balancing global semantic guidance and exploration autonomy.

Belief Source	Method	SR $\uparrow$	RS $\uparrow$	OSR@40 $\uparrow$
CLIPSeg [ 10 ]	AeroDuo [ 14 ]	6.67	12.70	8.30
CLIPSeg [ 10 ]	BGC-Nav	12.82 6.15↑	15.06 2.36↑	8.97 0.67↑
Claude Sonnet 5 [ 1 ]	AeroDuo [ 14 ]	10.95	13.97	11.82
Claude Sonnet 5 [ 1 ]	BGC-Nav	13.24 2.29↑	15.49 1.52↑	9.85 1.97↓

Table 3: Performance comparison across different belief-map generation models.

Method	SR $\uparrow$	OSR@30 $\uparrow$	OSR@60 $\uparrow$	CDS $\uparrow$	RS $\uparrow$
FBE [ 17 ]	4.34	2.01	8.94	3.10	10.03
+ Ours	12.82 8.48↑	4.03 2.02↑	17.31 8.37↑	6.42 3.32↑	13.21 3.18↑
VLFM [ 19 ]	5.70	4.19	8.76	4.10	9.97
+ Ours	6.93 1.23↑	8.42 4.23↑	9.49 0.73↑	4.41 0.31↑	13.42 3.45↑
APEX [ 22 ]	4.52	2.20	7.35	2.36	10.93
+ Ours	13.24 8.72↑	6.21 4.01↑	14.05 6.70↑	4.05 1.69↑	15.49 4.56↑

Table 4: Integration of our training-free belief-guided coupling mechanism into existing exploration methods.

Method	SR $\uparrow$	OSR@30 $\uparrow$	OSR@60 $\uparrow$	CDS $\uparrow$	RS $\uparrow$
Ours w/o LAA	7.31	6.08	11.29	3.98	13.56
Ours w/o DBP	7.69	6.21	7.95	2.00	12.77
Ours w/o LEM	8.39	4.17	10.42	4.02	13.42
Ours	13.24	6.21	14.05	4.05	15.49

Table 5: Ablation study of LAA (Local Attraction Augmentation), DBP (Directional Belief Prior), and LEM (Local Exhaustion Memory) components.

UAV Navigation

Unmanned aerial vehicle (UAV) navigation develops methods for locating and guiding drones safely through known, unknown, and changing environments.

See all 272 papers

CoNav-UAV: Cooperative Dual-Altitude Aerial Navigation via Stackelberg Learning
Aug 3, 2026Junru Song, Wenhao Zhang, Yang Yang +7UAV Navigation Stackelberg Games
1Shanghai Jiao Tong University · 2Intelligent Game and Decision Laboratory · 3Renmin University of China +1
Air-Ground Collaborative Vision-and-Language Navigation via Shared Bird's-Eye Maps
Sep 3, 2026Shuning Zhang, Liang Li, Yunheng Wang +3Robot Navigation UAV Navigation
Robotics and Autonomous Systems Thrust, Systems Hub, The Hong Kong University of Science and Technology (Guangzhou), China · School of Communications and Information Engineering, Nanjing University of Posts and Telecommunications, China
Beyond Matching to Tiles: Bridging Unaligned Aerial and Satellite Views for Vision-Only UAV Navigation
Mar 23, 2026Kejia Liu, Haoyang Zhou, Ruoyu Xu +3UAV Navigation Cross-View Geo-Localization
College of Computer Science and Technology, Zhejiang University · State Key Laboratory of Blockchain and Data Security, Zhejiang University · Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security

Aug 3, 2026J/K move · Enter open · S save

CoNav-UAV: Cooperative Dual-Altitude Aerial Navigation via Stackelberg Learning

Junru Song, Wenhao Zhang, Yang Yang, Xuekai Qiu+6

1Shanghai Jiao Tong University · 2Intelligent Game and Decision Laboratory · 3Renmin University of China+1

UAV Navigation Stackelberg Games Vision-Language Navigation Dataset Aggregation Aerial VLN

Open paper AskPDF arXiv

Target-oriented vision-and-language navigation (VLN) on aerial platforms is attracting growing attention for missions such as disaster rescue, infrastructure inspection, and security patrol. In this task, an unmanned aerial vehicle (UAV) needs to locate targets given only a concise description of their appearance and surroundings. This requires global exploration and grounding as well as collision-free close-range approach, two interleaved processes difficult to reconcile within a single agent. Most existing methods transfer the ground VLN paradigm to a low-altitude UAV and compensate for its inefficient exploration with external assistance. A recent attempt deploys two UAVs at complementary altitudes yet still relies on privileged information and trains its two agents independently, precluding any mutual adaptation essential for cooperation. Here we propose CoNav-UAV, which explicitly models the task as a Stackelberg game between a high-altitude leader and a low-altitude follower, with the system operating on onboard visual and linguistic inputs alone. To solve this game, we introduce Iterative Stackelberg Learning. The leader's high-level vision-language reasoning is refined via memory-based in-context learning, while the follower's precise motion control is updated via DAgger-style expert distillation. The alternation drives both agents toward a Stackelberg equilibrium. CoNav-UAV consistently outperforms single- and dual-agent baselines across three high-fidelity urban scenes from the AerialVLN benchmark. Success rate improves by up to 30.8 points on the learning scene, and 9.0 points under cross-scene transfer while using about 3x less adaptation data. Further analyses validate the complementary gains of the leader and follower updates and reveal robust gains yet distinct learning dynamics across VLM backbones.

Abstract

UAV Navigation

Explore similar work

CoNav-UAV: Cooperative Dual-Altitude Aerial Navigation via Stackelberg Learning

Air-Ground Collaborative Vision-and-Language Navigation via Shared Bird's-Eye Maps

Beyond Matching to Tiles: Bridging Unaligned Aerial and Satellite Views for Vision-Only UAV Navigation