// the find
mani-skill/ManiSkill
SAPIEN Manipulation Skill Framework, an open source GPU parallelized robotics simulator and benchmark
ManiSkill is a GPU-parallelized robotics simulator built on SAPIEN, designed for training manipulation policies at scale. It handles the full pipeline from simulation to real-world deployment, with built-in support for PPO, SAC, Diffusion Policy, and VLA models like RDT-1B. The target audience is robotics ML researchers who need high-throughput data collection without standing up their own simulation infrastructure.
GPU-parallelized heterogeneous simulation is the real differentiator — every parallel environment can have a completely different scene and object set, which most simulators can't do. The 30,000+ FPS RGBD collection claim on a 4090 is credible given the SAPIEN backend, and that throughput changes what's practical for imitation learning datasets. Real2sim environments for policy evaluation are a genuine time-saver — running 100x faster in sim before deploying to hardware is a meaningful engineering win. The baseline integrations are pre-tuned and actually cited, not just wrappers around reference implementations.
GPU simulation is Linux-only with an NVIDIA GPU — Windows and Mac users get CPU sim at best, which defeats most of the throughput argument. The asset license (CC BY-NC 4.0) blocks commercial use, so if you're building a product rather than publishing a paper, you need to audit every environment you use. The framework is deeply coupled to SAPIEN, which has had historically slow release cycles and limited community debugging; when SAPIEN breaks something, you wait. Custom task authoring requires understanding the GPU memory management abstractions the framework wraps, and those abstractions leak when something goes wrong.