Most robotics simulators force a tradeoff: pick one tool for contact-rich rigid bodies, a separate solver for cloth or fluids, and yet another renderer — then glue them together. The bet here is that a single GPU-native, differentiable engine can host rigid bodies, MPM, FEM, fluids, and particles in one shared scene, scripted from plain Python. That unification, more than any benchmark, is why a 20+ lab academic collaboration spent two years building it.
What Sets It Apart
- Unified multi-physics: rigid, FEM, MPM, fluid, and particle solvers share one scene and state, so a robot can grasp a deformable object or pour a liquid without bolting on external libraries.
- Portable backend: the same code compiles to CUDA, AMD ROCm, Apple Metal, and Vulkan, scaling from a laptop to a datacenter.
- Differentiable and generative: gradients flow through the physics for optimization and RL, and a generative layer can author 4D scenes from text prompts.
- The viral "43 million FPS / 430,000x real-time" figure is real but narrow — it was measured on a single Franka arm with self-collision only. Independent benchmarks found typical contact-rich scenes far slower, and the team has since published more representative numbers.
Who It's For
Great fit if you train embodied-AI or RL policies and want one Python-first stack spanning rigid and soft-body physics, with differentiability and a permissive Apache-2.0 license. Look elsewhere if you need a battle-tested, paper-grade baseline today — MuJoCo and Isaac Sim have deeper validation — or if your plans hinge on the headline speed numbers, which swing sharply with scene complexity.
