FurnitureBench
收藏资源简介:
FurnitureBench是一个用于测试真实机器人复杂长时域操作任务的数据集。数据集聚焦于家具组装这一复杂长时域操作任务,其任务层次结构长,涉及家具部件的选择、抓取、移动、对齐和连接等步骤,平均任务时长在 60 - 230 秒(600 - 2300 低层级步骤)。任务要求机器人具备多种复杂技能,如精确抓取(不同家具部件抓取姿态各异)、部件重定向(通过拾取放置或推动实现)、路径规划(避免碰撞已组装部件)、插入和拧紧(精确对齐并重复操作)等。
FurnitureBench is a dataset for testing complex long-horizon manipulation tasks on real robots. This dataset focuses on furniture assembly, a complex long-horizon manipulation task with a lengthy hierarchical task structure involving steps such as furniture component selection, grasping, moving, alignment, and fastening. The average task duration ranges from 60 to 230 seconds, corresponding to 600 to 2300 low-level steps. Such tasks require robots to possess a range of complex skills, including precise grasping (with distinct grasping postures for different furniture components), component redirection (implemented via pick-and-place or pushing), path planning to avoid collisions with already assembled components, and insertion and tightening operations that demand precise alignment and repeated actions, among others.




