ManiSkill

ManiSkill (SAPIEN Manipulation Skill Framework) is an open-source, GPU-parallelized robotics simulator and benchmark platform developed by the SAPIEN team (Jiayuan Gu, Fanbo Xiang, et al.). With over 3,020 GitHub stars, it provides the embodied AI community with standardized environments and evaluation protocols for manipulation skill learning. The platform builds on SAPIEN, a rich 3D virtual environment for embodied AI, offering GPU-accelerated parallel simulation that enables efficient training of reinforcement learning and imitation learning policies. It supports diverse manipulation tasks including articulation, grasping, insertion, and tool use. ManiSkill provides procedurally generated training tasks with infinite variations, addressing the critical need for diverse training data in generalizable manipulation. It includes both the original ManiSkill benchmark and the newer ManiSkill2, which expanded to 20 task families with 2,000+ object models and 4 million demonstration frames. The benchmark is characterized by its focus on object-level generalization — testing whether learned policies can handle novel object geometries, poses, and physical properties. It supports multiple input modalities including vision, depth, and proprioception. ManiSkill is widely used in robotics research for benchmarking manipulation policy learning. Its GPU-parallelized architecture makes it particularly suitable for large-scale policy training with modern deep RL and imitation learning algorithms.