DexCap Data

该数据集名为DexCap，提供高质量的三维动作捕捉数据，记录了人类手部在机器人操作空间内的运动，非常适合用于训练灵巧型机器人的技能。DexCap的设计理念是快速收集数据，其速度大约是传统远程操作方法的三倍，并且对遮挡具有鲁棒性。在前三个任务中，DexCap数据收集了30分钟，分别产生了201、129和82个演示；在野外环境中，DexCap收集了1小时的数据，提供了96个包装任务的演示；剪刀剪裁和泡茶准备任务各自收集了1小时的数据，分别产生了104和55个演示。这些任务包括海绵拾取、球体收集、盘子擦拭、包装、剪刀剪裁和泡茶准备等多种灵巧操作任务。

This dataset, named DexCap, provides high-quality 3D motion capture data that records human hand movements within robotic manipulation workspaces, making it ideally suited for training dexterous robotic manipulation skills. The core design goal of DexCap is to enable fast data collection, with a throughput approximately three times that of conventional teleoperation-based data collection methods, while maintaining robustness against occlusions. For the first three tasks, 30 minutes of data collection was carried out for each, yielding 201, 129, and 82 demonstrations respectively. In unconstrained real-world environments, DexCap collected 1 hour of data and generated 96 demonstrations for packaging tasks. For the scissor cutting and tea preparation tasks, 1 hour of data collection was performed for each, generating 104 and 55 demonstrations respectively. These tasks encompass a diverse set of dexterous manipulation tasks including sponge picking, sphere collection, plate wiping, packaging, scissor cutting, and tea preparation.