ManiSkill#

teaser

Sample of environments/robots rendered with ray-tracing. Scene datasets sourced from AI2THOR and ReplicaCAD

Downloads Open In Colab PyPI version Docs status Discord

ManiSkill is an open-source framework for robot simulation and training powered by SAPIEN, with a strong focus on manipulation skills. Among its features include:

  • GPU parallelized visual data collection system. On the high end you can collect RGBD + Segmentation data at 30,000+ FPS on a 4090 GPU

  • GPU parallelized simulation, enabling high throughput state-based synthetic data collection in simulation

  • GPU parallelized heterogeneous simulation, where every parallel environment has a completely different scene/set of objects

  • Example tasks cover a wide range of different robot embodiments (humanoids, mobile manipulators, single-arm robots) as well as a wide range of different tasks (table-top, drawing/cleaning, dexterous manipulation)

  • Flexible and simple task building API that abstracts away much of the complex GPU memory management code via an object oriented design

  • Real2sim environments for scalably evaluating real-world policies 100x faster via GPU simulation.

  • Sim2real examples for deploying policies trained in simulation to the real world

  • Many tuned robot learning baselines in Reinforcement Learning (e.g. PPO, SAC, TD-MPC2), Imitation Learning (e.g. Behavior Cloning, Diffusion Policy), and large Vision Language Action (VLA) models (e.g. Octo, RDT-1B, RT-x)

For more details we encourage you to take a look at our paper, published at RSS 2025.

Please refer to our documentation to learn more information from tutorials on building tasks to sim2real to running baselines. If you find any bugs or have any feature requests please post them to our GitHub issues or discuss about them on GitHub discussions. We also have a Discord Server through which we make announcements and discuss about ManiSkill.

Users looking for the original ManiSkill2 can find the commit for that codebase at the v0.5.3 tag