ManiSkill

Common

An open-source robot manipulation simulation framework and benchmark from a UC San Diego team, built for GPU parallelism.

ManiSkill is a robot manipulation simulation framework and benchmark developed by Hao Su's group at UC San Diego together with Hillbot and others, built on the same team's SAPIEN engine and now in its third generation. The latest version, ManiSkill3 (RSS 2025), runs both physics simulation and camera rendering in parallel on the GPU; its paper claims 10 to 1000 times the speed of other platforms and 2 to 3 times less GPU memory use, generating over 30,000 segmented RGB-D frames per second on a single RTX 4090. It covers 12 task categories, including tabletop manipulation, mobile manipulation, humanoids, dexterous hands, and soft bodies (soft-body tasks don't support batched parallelism), plus digital-twin environments built to match real-world setups for quickly evaluating real-robot policies in simulation. Its interface follows Gymnasium, and it exited beta with version 3.0.1 in April 2026.

ExampleManiSkill3's built-in PickCube-v1 task, where an arm has to pick up a cube, can be run as a thousand parallel environments at once to train a camera-conditioned grasping policy with PPO.

Also called: ManiSkill3, ManiSkill2, ManiSkill 1
Related: SAPIEN (SimulAted Part-based Interactive ENvironment) · GPU-Accelerated Parallel Simulation · Simulation-Based Evaluation · SimplerEnv · Digital Twin · Benchmark
Sources: haosulab/ManiSkill (GitHub)
ManiSkill3: GPU Parallelized Robotics Simulation and Rendering for Generalizable Embodied AI (arXiv 2410.00425)
mani-skill (PyPI)
As of: 2026-04