ManiSkill

Why we included this project

Training robot manipulation policies usually means waiting on slow, single-scene simulators. ManiSkill sidesteps that by running many environments in parallel on one GPU, so you can generate visual or state-based training data far faster than stepping through one scene at a time. It comes with a wide set of ready-made tasks, from table-top and dexterous manipulation to mobile and single-arm robots, and its task-building API keeps most of the GPU memory management out of your way. The bundled baselines for reinforcement learning, imitation learning, and vision-language-action models mean you don't have to implement those comparisons from scratch. If you care about sim-to-real transfer, the included real2sim and sim2real examples give you a quick way to check whether a policy trained in simulation behaves sensibly before you put it on physical hardware.