AerialAgent-Ego10k, CyberAgentEgo500k, SkyAgent-Scene3k, SkyAgent-Reason3k, SkyAgent-Nav3k, SkyAgent-Plan3k, SkyAgent-Act3k
收藏资源简介:
本研究由中国科学院空天信息创新研究院主导,构建了首个大规模的无人机视角图像-文本预训练数据集AerialAgent-Ego10k和虚拟图像-文本-姿态对齐数据集CyberAgentEgo500k,以及五个下游任务指令数据集。这些数据集涵盖了从真实到虚拟的多种城市环境,旨在提升无人机在复杂环境中的感知、认知和行动能力。数据集的创建过程结合了专业的无人机操作和数据标注,经过严格的质量控制,适用于无人机智能代理的预训练和微调,以解决无人机自主智能的关键问题。
This study was led by the Aerospace Information Research Institute, Chinese Academy of Sciences. The research team constructed the first large-scale UAV egocentric image-text pre-training dataset AerialAgent-Ego10k, the virtual image-text-pose aligned dataset CyberAgentEgo500k, and five downstream task instruction datasets. These datasets cover diverse urban environments spanning real-world and virtual scenarios, aiming to improve the perception, cognition and operational capabilities of UAVs in complex environments. The development of these datasets integrates professional UAV operations and data annotation, with rigorous quality control, and is tailored for pre-training and fine-tuning of UAV AI Agents to address core challenges in UAV autonomous intelligence.




