跳到正文
arXiv cs.AI· Junru Song, Wenhao Zhang, Yang Yang, Xuekai Qiu, Feifei Wang, Weien Zhou, Tingsong Jiang, Ying Wen, Yang Li, Wen Yao·· 2 小时前AI 评分27

CoNav-UAV:基于 Stackelberg 学习的双高度协同无人机导航

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

AI 导读

研究者提出 CoNav-UAV,将目标导向空中视觉语言导航建模为高空领导者和低空跟随者之间的 Stackelberg 博弈,两架无人机仅依赖机载传感。其 Iterative Stackelberg Learning 方法让领导者通过基于记忆的 in-context learning 精进视觉语言推理,跟随者通过 DAgger 式专家蒸馏微调运动控制。

正文 · 原文

View PDF HTML (experimental)

Abstract:Aerial search missions often require an unmanned aerial vehicle (UAV) to locate a target from only a brief description of its appearance and surroundings, a task known as target-oriented aerial vision-and-language navigation (VLN). The task demands city-scale exploration and grounding as well as collision-free close-range approach, two interleaved processes that are difficult to reconcile within a single agent. Most existing methods adapt the ground VLN paradigm to a single UAV and rely on external assistance to compensate for its inefficiency. A recent attempt deploys two UAVs at complementary altitudes but still depends on privileged sensing and trains the two agents independently, precluding any mutual adaptation. We propose CoNav-UAV, which models the task as a Stackelberg game between a high-altitude leader and a low-altitude follower to capture their sequential decisions and asymmetric observations and utilities, with both agents using only onboard sensing. We introduce Iterative Stackelberg Learning, which tailors optimization to each agent's role: the leader's high-level vision-language reasoning builds on pretrained capabilities and is refined through memory-based in-context learning, while the follower's precise motion control is fine-tuned through DAgger-style expert distillation. Alternating the two updates lets each agent continually adapt to the other's evolving behavior. CoNav-UAV outperforms single- and dual-agent baselines across three high-fidelity urban scenes from the AerialVLN benchmark. It improves success rate by up to 30.8 points on the learning scene and by up to 9.0 points under cross-scene transfer, where the leader reuses its learned memory and the follower is adapted with 2.6-3.7 times less target-scene data. Extensive ablations confirm the contributions of key components and the benefit of mutual adaptation, with consistent gains across VLM backbones.
Comments: 27 pages, 10 figures. Revised manuscript with expanded analysis and experiments. Removed an AAAI copyright notice inadvertently included through the template in the previous version
Subjects: Artificial Intelligence (cs.AI); Robotics (cs.RO)
Cite as: arXiv:2608.01802 [cs.AI]
  (or arXiv:2608.01802v2 [cs.AI] for this version)
  https://doi.org/10.48550/arXiv.2608.01802

arXiv-issued DOI via DataCite

Submission history

From: Junru Song [view email]
[v1] Mon, 3 Aug 2026 07:10:38 UTC (3,741 KB)
[v2] Wed, 18 Nov 2026 11:47:41 UTC (17,529 KB)

来源:arXiv cs.AI · arxiv.org