Genesis World
Open-source multi-physics simulator for robot learning. Rigid, FEM, MPM, SPH and PBD solvers share one Python scene, with a photoreal renderer and heavy GPU batching.
Overview
Started as an academic project across CMU, Stanford and collaborators in December 2024, renamed Genesis World, and now developed with support from Genesis AI. The pitch is one simulator instead of four. Rigid-body dynamics, finite element soft bodies, material point method for granular flow, SPH fluids and position-based cloth all run inside a single scene description, which is genuinely unusual and the main reason to look at it over MuJoCo.
Everything is Python on top of Taichi and CUDA kernels, so reading and patching a solver is realistic for a team with a competent numerics person. The Nyx renderer covers photoreal image output for perception training, and the Quadrants compiler spreads the same code across CUDA, ROCm, Metal, Vulkan, x86 and ARM64. That AMD and Apple coverage is rare. Version 1.3.3 landed on 13 August 2026 and needs Python 3.10 or newer, below 3.14.
Treat the benchmarks with care. The 43M FPS figure quoted for a Franka manipulation scene comes from the project itself, on a scene chosen to suit the rigid solver, and community reproductions vary widely. Funding is not the risk: Genesis AI raised $105M in July 2025 and was reported in July 2026 to be raising again at a $3B pre-money valuation. Attention is the risk. The company shipped its own humanoid, Eno, in mid-2026, and an open-source simulator maintained alongside a hardware business can drift toward whatever that hardware needs.
Key Features
- ✓ Unified scene where rigid bodies, FEM soft bodies, MPM granular material, SPH fluids and PBD cloth coexist instead of living in separate engines
- ✓ Written in Python on Taichi and CUDA kernels, so solver internals are readable and patchable rather than buried in C++
- ✓ Nyx renderer for photoreal output aimed at perception training and synthetic data generation
- ✓ Quadrants cross-platform compiler targeting CUDA, ROCm, Metal, Vulkan, x86 and ARM64 from the same source
- ✓ Large-scale environment batching for policy training, with published rigid-body figures in the tens of millions of steps per second on one RTX 4090
- ✓ Apache 2.0 across the whole stack, renderer included
Where it holds
- • Multi-physics in one scene solves a real problem: deformables plus granular media normally means bolting two engines together and writing the coupling yourself
- • Python-first internals make modifying a solver practical instead of filing an issue and waiting a quarter
- • Runs on AMD and Apple silicon, which almost nothing else in this category does
- • Around 29k GitHub stars and steady releases through August 2026, so it is well past the demo phase
Where it breaks
- • The headline throughput numbers are the project's own, measured on scenes that favour the rigid solver, and independent reproductions have been inconsistent since the December 2024 launch
- • Sim-to-real track record is thin next to MuJoCo, which has years of published legged-robot transfers behind it
- • The API has churned across 1.x releases and documentation lags the code, so expect to read source
- • Development is funded by Genesis AI, a startup that pivoted to building its own humanoid in May 2026, so open-source priorities sit downstream of a commercial roadmap
My Take
Launched in December 2024 with throughput claims that sounded implausible, and the reaction split the robotics crowd down the middle. Eighteen months on it is version 1.3.3, released 13 August 2026, with solvers for rigid bodies, FEM, MPM, SPH and PBD sharing a single scene. That combination is the real argument, not the frames per second: simulating a gripper picking up sand or wet cloth normally means two engines and a hand-written coupling layer. Sim-to-real evidence remains thinner than MuJoCo's, and the roadmap now sits inside a company building its own humanoid, which is worth watching.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- Free under Apache 2.0. `pip install genesis-world`, version 1.3.3 dated 13 August 2026, requiring Python 3.10 or newer and below 3.14. Backends cover CUDA, AMD ROCm, Apple Metal, Vulkan, x86 and ARM64, so there is no NVIDIA lock-in and no hosted tier. Genesis AI funds development and sells robots, not the simulator.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Genesis World holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to Genesis World
MuJoCo
open-sourceGoogle DeepMind's rigid-body physics engine for robotics. Contact-rich dynamics, MJCF models, plus MJX and MuJoCo Warp for batched GPU rollouts during policy training.
NVIDIA Isaac Sim
freeGPU-accelerated robotics simulator for physics-accurate testing, synthetic data generation and large-scale policy training, with ROS 2 bridging and sim-to-real workflows.
BlenderProc
open-sourceGPL-3.0 Blender pipeline from DLR that renders photorealistic synthetic training images with RGB, depth, normals, segmentation and exact 6D pose, exporting to BOP and COCO.