Socially CompliAnt Navigation Dataset (SCAND)
收藏资源简介:
SCAND是一个大规模的第一人称视角社会导航演示数据集,包含8.7小时、138条轨迹、25英里的社会合规性人类远程操作驾驶演示。数据集涵盖了多种数据流,包括3D激光雷达、操纵杆命令、里程计、视觉和惯性信息,由四名不同的人类演示者在室内和室外环境中收集。SCAND旨在通过模仿学习解决移动机器人在人类环境中导航的挑战,特别是在确保安全和舒适的人机共存方面。数据集不仅包含丰富的多模态现实世界数据,还包含每个轨迹上自然发生的社交互动的标签,适用于研究社会导航策略和学习社会合规的本地和全局导航策略。
SCAND is a large-scale first-person perspective social navigation demonstration dataset, containing 8.7 hours, 138 trajectories, and 25 miles of socially compliant human teleoperated driving demonstrations. The dataset covers multiple data streams, including 3D LiDAR, joystick commands, odometry, visual and inertial data, collected by four distinct human demonstrators in both indoor and outdoor environments. SCAND aims to address the challenge of mobile robot navigation in human-centric environments via imitation learning, particularly in ensuring safe and comfortable human-robot coexistence. In addition to rich multimodal real-world data, the dataset also includes labels for naturally occurring social interactions on each trajectory, making it suitable for researching social navigation strategies and learning socially compliant local and global navigation strategies.



