Embodied AI Glossary中文Genesis
CommonA multi-physics GPU simulation platform open-sourced in late 2024, built for speed and unified rigid-body, soft-body, and fluid simulation.
Genesis (now branded Genesis World) is a physics simulation platform for robotics and embodied AI, open-sourced as an academic project in December 2024 under the Apache 2.0 license and now developed with backing from Genesis AI; it has over 30,000 GitHub stars. It unifies several solvers — rigid body, finite element method (FEM, for soft-body deformation), material point method (MPM, for granular materials like sand), and PBD/SPH particle methods (for cloth and fluids) — inside a single scene, and ships with a photorealistic renderer called Nyx. It is written in Python and compiled to CUDA, Metal, and other backends by Quadrants, a compiler forked from Taichi. The developers claim throughput 10 to 80 times that of GPU simulators such as Isaac Gym/Sim/Lab and MuJoCo MJX. It is commonly used for massively parallel reinforcement learning, such as training locomotion controllers for quadrupeds and humanoids.
ExampleThe official examples include a script that runs thousands of Unitree Go2 environments in parallel with Genesis and trains a walking policy with PPO, completing training on a single GPU.
- Also called
- Genesis World, Genesis Simulator, Genesis-Embodied-AI
- Related
- GPU-Accelerated Parallel Simulation · Physics Engine · Material Point Method · NVIDIA Isaac Lab · MuJoCo (Multi-Joint dynamics with Contact) · Generative Simulation
- Sources
- Genesis-Embodied-AI/Genesis (GitHub)
Genesis World Docs: What is Genesis - As of
- 2026-09