ManiSkill
maniskill.ai
- Free plan
- Windows
- macOS
- Linux
7.6 Club score Visit maniskill.ai
Plans
Every plan and priceOpen source
Free
Free plan
What the maker publishes
- Docker
- A Docker image named maniskill/base is provided on Docker Hub, with a corresponding Dockerfile. source
- GPU rendering speed
- The site says ManiSkill/SAPIEN can simulate and render RGB-D in parallel at up towards 30,000+ FPS. source
- Gymnasium interface
- ManiSkill is a SAPIEN-based simulator with a standard Gym/Gymnasium interface for reinforcement learning and imitation learning, and it supports GPU and CPU simulation plus parallelized rendering. source
- Heterogeneous simulation
- Its GPU simulator can run different objects and articulations in each parallel environment at the same time. source
- Learning baselines
- It provides tuned reinforcement-learning and imitation-learning baselines, including PPO, TD-MPC2, and Diffusion Policy. source
- Multi-GPU limit
- On multi-GPU servers a process can select a device with CUDA_VISIBLE_DEVICES, but exposing multiple visible CUDA devices to one process is not properly supported because of rendering bugs. source
- Physics and sensors
- It simulates low-level physics such as gravity and friction, as well as various sensors, in a realistic virtual environment. source
- Pre-built tasks
- It offers 40+ skills/tasks across 10+ categories, 2,000+ ready-to-use objects, millions of demonstration frames, and dense reward functions. source
- Real2sim and sim2real
- The docs list real2sim environments for evaluating real-world policies about 100x faster with GPU simulation, plus sim2real examples for deploying policies trained in simulation. source
- Rendering requirement
- State-based simulation needs no extra dependencies, but rendering requires a GPU with the Vulkan driver installed. source
- Support
- Bug reports and feature requests go to GitHub issues or discussions, announcements and discussion are on Discord, and general inquiries use [email protected]. source
- System support
- Linux with an NVIDIA GPU supports CPU simulation, GPU simulation, and rendering; Windows supports CPU simulation and rendering but not GPU simulation; MacOS supports CPU simulation and rendering but not GPU simulation; WSL supports CPU simulation only. source
- Team
- The about page lists current key team members Hao Su, Fanbo Xiang, and Stone Tao, each affiliated with UC San Diego. source
- What it is
- ManiSkill is an open-source, GPU-accelerated robotics simulation platform built on the SAPIEN engine for designing, simulating, testing, and training AI-based robots. source
- Who it is for
- It targets developers training and researching robot policies, and the about page says the interface and documentation make it accessible to developers of all experience levels. source
Checked against the maker's pages on 14 Nov 2026.