UAV-ON
收藏资源简介:
UAV-ON 是一个大规模的空中物体目标导航数据集,用于评估无人机在开放世界环境中的自主导航能力。数据集包含了14个高保真度的Unreal Engine环境,这些环境具有多样化的语义区域和复杂的空间布局,涵盖了城市、自然和混合使用场景。数据集定义了1270个带有注释的目标对象,每个对象都有一个实例级别的指令,包含类别、物理足迹和视觉描述符,允许进行接地推理。这些指令作为语义目标,为空中代理引入了真实的模糊性和复杂的推理挑战。数据集和代码可以在提供的链接中获取。
UAV-ON is a large-scale aerial object navigation dataset designed to evaluate the autonomous navigation capabilities of unmanned aerial vehicles (UAVs) in open-world environments. The dataset includes 14 high-fidelity Unreal Engine environments, which feature diverse semantic regions and complex spatial layouts, covering urban, natural, and mixed-use scenarios. It defines 1270 annotated target objects, each paired with instance-level instructions that encompass category, physical footprint, and visual descriptors, enabling grounded reasoning. These instructions serve as semantic goals, introducing realistic ambiguities and complex reasoning challenges for aerial agents. The dataset and accompanying code are accessible via the provided link.
UAV-ON 数据集概述
数据集简介
- 名称: UAV-ON: A Benchmark for Open-World Object Goal Navigation with Aerial Agents
- 目标: 促进空中智能体在开放世界环境中进行大规模目标导航(ObjectNav)的研究
- 特点:
- 包含14个高保真Unreal Engine环境,涵盖城市、自然和混合用途场景
- 定义了1270个带注释的目标对象,每个对象配有结构化语义提示
- 语义提示包含类别、估计物理足迹和详细视觉描述符
- 旨在推动复杂现实环境中基于语义目标描述的可扩展无人机自主性研究
数据集内容
- 环境:
- 训练环境: 44.1G (可从Hugging Face获取)
- 测试环境: 26.8G (可从Hugging Face获取)
- 数据集文件: JSON格式 (可从Hugging Face获取)
使用指南
-
依赖安装: bash conda create -n uavon python==3.8 conda activate uavon pip install -r requirements.txt
-
环境准备:
-
下载并解压训练和测试环境
-
确保目录结构如下:
workspace/ ├── UAV_ON/ ├── DATASET/ ├── TRAIN_ENVS/ └── TEST_ENVS/
-
-
运行步骤:
-
启动AirSim环境服务器: bash python airsim_plugin/AirVLNSimulatorServerTool.py --port=30000 --root_path="your workspace path"
-
执行评估脚本: bash
AOA-F/V
bash scripts/eval_fixed.sh bash scripts/eval_unfixed.sh
CLIP-H
bash scripts/eval_cliph.sh bash scripts/metric.sh
-
补充信息
- 演示视频: YouTube链接
- 附录文档: Google Drive链接

- 1UAV-ON: A Benchmark for Open-World Object Goal Navigation with Aerial Agents哈尔滨工业大学(深圳) · 2025年



