Genesis World — Simulation platform for general-purpose robotics & embodied AI learning

TL;DR

Genesis World is the multi-physics simulator at the centre of Genesis AI’s open-source robotics stack — Apache 2.0, pip-installable as genesis-world, supporting Linux/macOS/Windows on CPU + NVIDIA/AMD/Apple GPUs. It unifies rigid, FEM, MPM, SPH, PBD, IPC (via libuipc), and SAP solvers behind a single Pythonic scene API, with three interchangeable couplers swappable via a one-line change. It plugs into Nyx — GPU path-traced renderer plugin for Genesis World for rendering, runs on top of Quadrants — high-performance multi-platform compiler for physics simulation for GPU compilation, and was originally an academic project (named “Genesis”) started in Dec 2024 before becoming the formal foundation of the The Role of Simulation in Scalable Robotics, Genesis World 1.0, and the Path Forward platform.

Key claims

Unified multi-physics: Rigid, FEM (linear corotated, Implicit FEM with Newton+CG), MPM (granular/elasto-plastic), PBD (cloth, position-based liquids), SPH (fluids), libuipc-extended IPC (intersection-free contact for delicate deformables), and SAP all share one scene and one state [README §What is Genesis World?].
Three interchangeable couplers — fast general-purpose, Drake-style Semi-Analytic Primal (SAP) with hydroelastic contact, and IPC with intersection-free contact — switch via a one-line API call without changes to assets, sensors, or policies [README §Catalogue / Physics].
Asset support spans URDF (including xacro), MJCF, OBJ, GLB, USD, with MuJoCo general actuators, compound/mimic joints, and equality/weld constraints [README §Catalogue].
Sensor suite includes RGB and depth cameras, IMU, lidar (teleop demo), contact force, tactile sandbox, surface distance (shadowhand demo), and temperature grid — covering perception modalities normally hard to obtain on real hardware [README §Simulation Interface].
Multi-robot embodiment support: examples ship for Franka, Go2, Shadow Hand, drones, and the Genesis-internal Wuji / Sharpa / Pika gripper / Tianji arm / G1 humanoid embodiments shown in the launch video [README §Catalogue + blog post embodiment list].
Multi-platform: pip install on Python 3.10–3.13, Docker images for both NVIDIA and AMD (ROCm/HIP), with gs.init(backend=gs.amdgpu) selecting the AMD path [README §Quick Installation, §Docker].
Heterogeneous parallel simulation: mixes different objects, kinematic trees, and scene layouts in a single batched environment [README §Catalogue / Simulation Interface — heterogeneous_simulation.py].
Public APIs cover vertex manipulation, kinematic/potential energy queries, forward kinematics, Jacobian-at-point, and mass-matrix access — broader than typical user-facing simulator APIs [blog post §Coverage].
Origin and license: originally released as “Genesis” in Dec 2024 as an academic project; renamed and now officially supported by Genesis AI under Apache 2.0. The README cites both a 2024 GitHub bibtex and a 2026 blog-post bibtex [README §License and Acknowledgments, §Citation].
Acknowledged dependencies: Taichi (compiler ancestry of Quadrants), libuipc (IPC backend), MuJoCo, libccd, PyRender, LuisaCompute/LuisaRender, Madrona/Madrona-mjx, and reference solver implementations from FluidLab, SPH_Taichi, and PBF3D [README §License and Acknowledgments].

Method

The repo is a four-layer stack delivered as a single Python package: Simulation Interface → Physics → Render → Compiler. The Simulation Interface exposes the user-facing API (asset parsing, entity accessors, controllers, sensors, parallel and heterogeneous environments, plus a built-in GUI with ImGui joint control, mesh point picker, mouse interaction). The Physics layer integrates the multi-solver engine sharing one scene + one state, with three couplers. Rendering plugs in as camera sensors — Nyx (separate pip-installable extra gs-nyx), Luisa, or Pyrender. The Compiler (Quadrants, bundled automatically — no extra install) lowers Python kernels to CUDA/ROCm/Metal/Vulkan/x86/ARM64 via LLVM and carries autodiff, GPU graphs, and fastcache.

Installation paths cover pip, uv, and Docker (separate NVIDIA and AMD Dockerfiles). Optional extras: pip install pyuipc for the IPC backend (Linux/Windows x86, NVIDIA GPU), and pip install gs-nyx for the renderer plugin. Editable installs (pip install -e ".[dev]") are recommended for contributors.

The catalogue ships runnable demos across three sections — Physics (solvers and multi-solver coupling: franka cube, collision tower, FEM hard/soft constraint, MPM sand wheel, SPH+MPM coupling, PBD cloth/liquid, IPC robot cloth teleop, water wheel, flush cubes, SAP grasp), Rendering (follow entity, animated camera, Nyx hello/attached camera/PBR/light types/Gaussian splat/object picking/multi-cam multi-env), and Simulation Interface (controllers, GUI, heterogeneous envs, domain randomization, sensors including depth/IMU/lidar/tactile/contact force/surface distance/temperature grid, Diff-IK controller, batched IK, drone hover, advanced worm).

Results

This is a software release, not a paper — no benchmark numbers in the README itself. Performance claims (sim-to-real correlation, throughput, contact-solver speedups) are reported in The Role of Simulation in Scalable Robotics, Genesis World 1.0, and the Path Forward rather than in the repo documentation. Adoption signal: the README shows a PyPI package, ReadTheDocs documentation, GitHub Issues/Discussions, and a Discord — standard infrastructure for a maintained open-source release. Citation data is not yet meaningful at filing time (Dec 2024 academic version + May 2026 platform release).

Why it’s interesting

This is the core open-source artifact behind the Genesis World 1.0 announcement. Within World Foundation Models, it represents the explicit-physics-engine bet — contrast with neural-WFM analogs like HY-World 1.5 (WorldPlay): A Systematic Framework for Interactive World Modeling with Real-Time Latency and Geometric Consistency or PAN: A World Model for General, Interactable, and Long-Horizon World Simulation that learn dynamics from video. Within RL Environment Platforms, it occupies the physical-robotics niche that the existing terminal/coding env entries don’t cover. The Taichi-fork ancestry of its compiler also makes it the most concrete recent counterpoint to “Python physics simulation is a developer-experience deadweight” — Quadrants demonstrates that Python-frontend GPU kernels can compete with hand-rolled CUDA at production scale.