ManiSkill is a GPU-parallelized robotics simulator and benchmark framework for robot manipulation skill learning and training.
The framework addresses the challenge of efficiently collecting large-scale synthetic training data for robot learning by leveraging GPU parallelization across its entire pipeline. It provides GPU-accelerated simulation, visual data collection at rates exceeding 30,000 FPS on high-end hardware, and support for heterogeneous parallel environments where each simulated instance can have completely different scenes and objects. The tool abstracts away complex GPU memory management through an object-oriented task API, making it accessible for researchers building custom manipulation tasks.
ManiSkill suits teams developing robot learning systems that need high-throughput synthetic data generation or want to benchmark manipulation policies across diverse robot embodiments and task types. The framework includes example tasks spanning humanoids, mobile manipulators, and single-arm robots, with applications ranging from tabletop manipulation to dexterous control. It provides real2sim evaluation capabilities for testing real-world policies in simulation at accelerated speeds, along with sim2real deployment examples. The tool integrates with established reinforcement learning baselines such as PPO and SAC, imitation learning approaches including behavior cloning and diffusion-based methods, and large vision-language-action models.
The project maintains active engagement with its community through multiple channels including GitHub issues, discussions, and a Discord server for announcements. Development includes comprehensive documentation covering task building, sim2real workflows, and baseline implementations. The codebase is structured to support both new users through quick-start guides and advanced practitioners implementing custom environments and training pipelines.
