ManiSkill2

ManiSkill2 is the second generation of the SAPIEN ManiSkill benchmark, developed by Jiayuan Gu and collaborators to address critical pain points in generalizable manipulation skill research. Published in 2023, it provides the embodied AI community with a comprehensive evaluation platform. The benchmark includes 20 manipulation task families with over 2,000 object models and more than 4 million demonstration frames. It covers stationary and mobile-base manipulation, single-arm and dual-arm tasks, as well as rigid-body and soft-body manipulation, all simulated by fully dynamic physics engines with 2D and 3D input data. ManiSkill2 was designed to overcome limitations of earlier benchmarks: lack of object-level topological and geometric variations, insufficient dynamic simulation, and inadequate support for diverse manipulation types. It provides a unified interface and evaluation protocol that simplifies reproducibility across different research groups. The benchmark is closely tied to SAPIEN, a rich virtual environment for embodied AI, and supports multiple task categories including articulation, grasping, and insertion tasks that test different aspects of manipulation skill generalization. ManiSkill2 has become a standard evaluation platform for manipulation skill learning, complementing other benchmarks by focusing on object-level generalization — testing whether policies can adapt to novel object geometries, poses, and physical properties at test time.