GENESIS

World's fastest physics engine: 43M FPS for embodied AI learning

License

Apache-2.0

Overview

The world's fastest physics engine delivering 43M FPS simulation speeds. A generative world for general-purpose robotics & embodied AI learning, built from a 24-month collaboration of 20+ research labs. Supports diverse physics solvers (rigid body, MPM, SPH, FEM, PBD, Stable Fluid) and can train real-world transferable robot policies in just 26 seconds. 100% Python, cross-platform, with photo-realistic rendering.

GENESIS is the world's fastest physics engine, delivering unprecedented 43 million FPS simulation speeds. It's a generative world platform for general-purpose robotics and embodied AI learning, born from a 24-month collaboration of over 20 research labs worldwide. As a universal physics engine, GENESIS supports diverse physics solvers including rigid body, MPM (Material Point Method), SPH (Smoothed Particle Hydrodynamics), FEM (Finite Element Method), PBD (Position Based Dynamics), and Stable Fluid simulations. It can train real-world transferable robot policies in just 26 seconds and features 100% Python implementation with cross-platform support and photo-realistic rendering capabilities. The current release, v1.4.6 (8 November 2026), makes differentiable simulation work natively with PyTorch and about three times faster, extends it to FEM entities with the implicit solver, and runs the rigid and implicit FEM solvers several times faster for small batches on GPU. It replaces the built-in floor of the SAP and FEM solvers with the rigid ground plane, so scenes that relied on that floor need one.

Why it matters

GENESIS is a paradigm shift in physics simulation — 43 million FPS is orders of magnitude faster than existing engines. This speed unlocks training regimes that were previously computationally infeasible, making large-scale robot learning practical on a single GPU.

Problem it solves

Training robot policies in simulation requires billions of environment steps, but existing simulators are too slow. GENESIS delivers simulation speeds that make massively parallel training practical, reducing training time from days to hours.

Key takeaways

• 43 million FPS — the fastest physics simulation engine available
• Supports rigid body, soft body, fluid, and cloth simulation in one engine
• Built-in differentiable simulation for gradient-based optimization
• Generative framework can create scenes, objects, and camera trajectories
• Open-source with Python API for easy integration

Who it is for

• Robot learning researchers needing massive simulation throughput
• Teams training reinforcement learning policies at scale
• Researchers in deformable object manipulation
• Labs exploring generative simulation environments

Limitations

• Relatively new project with evolving API and documentation
• Custom physics engine may behave differently from MuJoCo or Bullet
• Generative features are experimental and require significant GPU memory
• Community and ecosystem smaller than established simulators