UAV-Track VLA
收藏资源简介:
UAV-Track VLA是由北京理工大学、中国科学院自动化研究所等机构联合构建的首个面向城市复杂场景的无人机视觉-语言-动作追踪基准数据集。该数据集基于CARLA仿真平台构建,包含89.2万帧多模态轨迹数据,覆盖85类动态目标(如车辆、行人)和176项细粒度追踪任务,支持自然语言指令输入与四自由度连续运动控制。数据通过专家演示与人工势场算法混合采集,涵盖动态天气、全距离追踪及目标运动学多样性,旨在解决无人机在语义级动态追踪中的跨模态对齐与实时控制问题,为城市交通监控、紧急搜救等场景提供算法训练基础。
UAV-Track VLA is the first benchmark dataset for unmanned aerial vehicle (UAV) vision-language-action tracking in complex urban scenarios, jointly constructed by institutions including Beijing Institute of Technology and the Institute of Automation of the Chinese Academy of Sciences. Built on the CARLA simulation platform, this dataset contains 892,000 frames of multimodal trajectory data, covering 85 categories of dynamic objects (e.g., vehicles, pedestrians) and 176 fine-grained tracking tasks, and supports natural language instruction input and four-degree-of-freedom continuous motion control. Collected via a hybrid approach combining expert demonstrations and artificial potential field algorithms, the dataset covers dynamic weather, full-range tracking, and diverse target kinematics. It aims to address the cross-modal alignment and real-time control issues in semantic-level dynamic tracking for UAVs, providing a foundational resource for algorithm training in scenarios such as urban traffic monitoring and emergency search and rescue.

- 1UAV-Track VLA: Embodied Aerial Tracking via Vision-Language-Action Models北京理工大学; 中国科学院·自动化研究所; 三亚大学; 北京邮电大学; 湖南大学; 北京航空航天大学 · 2026年



