mani-skill/maniskill is an open-source framework for simulating and learning manipulation tasks where robots handle objects. It bundles GPU parallel simulation and benchmarks, and the latest release is v3.0.1. However, no changes are given for this version, so what was added cannot be confirmed.
The underlying engine is SAPIEN. It collects visual data in parallel on the GPU and simulates it, and it provides parallel environment simulation that combines different scenes and objects, plus a task configuration API for defining tasks directly. The README describes this repository as a robot simulation and learning framework focused on manipulation tasks. It includes examples with multiple robots and tasks, real2sim environments that port real environments into simulation, sim2real examples that apply simulation strategies to real robots, and learning baselines for several robots. It is written in Python and licensed under Apache-2.0. The repository was created on August 2, 2022 and has 3436 stars.
That said, comparative material is limited. No speed or performance figures against other simulators are presented. In the evaluation scores, only momentum comes out low at 3.3 (out of 10), while the novelty, technical, and quality items are empty. So it is hard to judge which simulator is better right now. The repository is worth trying if you want to build a manipulation learning environment yourself, but it is not one to choose on the basis of a performance advantage.