ManiSkill2
收藏资源简介:
ManiSkill2是由加州大学圣地亚哥分校和清华大学合作开发的下一代操作技能统一基准。该数据集包含超过2000个对象模型和400多万个演示帧,覆盖固定/移动基座、单/双臂、刚性/软体操作任务,并支持2D/3D输入数据。ManiSkill2定义了统一的接口和评估协议,支持包括经典感知-规划-行动、强化学习、模仿学习在内的广泛算法,以及点云、RGBD等视觉观察模式和多种控制器。此外,它还实现了渲染服务器基础设施,允许所有环境共享渲染资源,显著减少内存使用。数据集旨在推动可泛化的操作技能研究,解决研究者在利用基准进行操作技能研究时遇到的痛点。
ManiSkill2 is a unified next-generation benchmark for manipulation skills, jointly developed by the University of California, San Diego and Tsinghua University. This dataset contains over 2000 object models and more than 4 million demonstration frames, covering fixed/mobile bases, single/dual-arm, rigid/soft-body manipulation tasks, and supports 2D/3D input data. ManiSkill2 defines a unified interface and evaluation protocols, supporting a wide range of algorithms including classical perception-planning-action pipelines, reinforcement learning, and imitation learning, as well as visual observation modalities such as point clouds, RGBD, and various controllers. Furthermore, it has implemented a rendering server infrastructure that enables all environments to share rendering resources, significantly reducing memory footprint. This benchmark aims to advance research on generalizable manipulation skills, addressing the pain points encountered by researchers when carrying out manipulation skill research using this benchmark.




