Genesis-World: Core Physics Engineer

What we're building

Robots will learn in simulation before they hit the factory. Genesis-World is our bet on that future.

Genesis-World is an open-source, general-purpose simulation platform for physical AI from Genesis AI. One unified multi-physics engine: rigid bodies, FEM, MPM, particles, cloth, fluids. A robot arm can pour water onto sand, grasp a deformable object, or cut a soft body, all in the same simulation. Nyx, our in-house renderer, may be the most promising renderer for robotics out there: real-time photo-realistic rendering, advanced features like depth of field, and state-of-the-art techniques never seen before.

Sensors of every kind: cameras, lidar, IMU, contact forces, temperature, plus arguably the most advanced tactile simulation available (paper). And the engine keeps growing: we are developing internally the most comprehensive and fastest Incremental Potential Contact (paper) solver for deformable body dynamics we know of, soon to be open-sourced. It powers real business applications, from full-fledged box packaging with labelling machine and all, to wire harnessing and lab automation, without any physics hack or compromise.

Everything is Python-first and runs anywhere. Kernels are written once, and Quadrants, our in-house JIT compiler, lowers them to CUDA, AMD ROCm, Apple Metal, Vulkan, x86, and ARM
64. A single laptop or a datacenter. Massively batched GPU simulation for learning at scale, and complex non-batched scenes where CPU wins outright.

This is at the core of Genesis AI's strategy. Evaluation is the bottleneck of scalable robotics: real hardware caps iteration at wall-clock time, but simulation turns it into a compute problem. Ours already runs two orders of magnitude faster than hardware (tens of thousands of episodes in half an hour instead of 200+ hours), while correlating with on-hardware rollouts at 89%.

The north star: physical AI that improves at the speed of compute.

You push the physics of Genesis-World forward. The mandate is clear: ship production-ready simulation capabilities that matter for the company's internal needs. Research applied end-to-end, from algorithm to merged, tested, documented code that real robot-learning pipelines depend on. Occasional groundbreaking research happens, notably through academic collaborations. But the core of the job is making the engine measurably better along five axes:

• Speed. Algorithms that are not only faster but also smart enough to spend compute only where it matters across both time and space: larger stable timesteps, selective fidelity (adaptive across scales or simply hand-set), structure-aware solvers.

• Completeness. No physics off limits: water, human animation, air flow, gravel, tendons, even body organs. Whatever the next use case needs, the engine grows to cover it.

• Fidelity. More realistic models: contact, friction, deformation, energy, actuation, materials…

• Versatility. Extensible multi-physics without compromise on realism: all solvers in the scene coupled together at once, two-way and constraint-based. Write your own solver and it joins the scene like a native one, growing into an open solver ecosystem.

• Scalability. From workstation to factory scale, and one day, city scale: thousands of interacting entities, batched across environments, without losing physical soundness.

Our ambition is to establish Genesis-World as the go-to simulator for physical AI, from companies and research labs to individuals.

• Every fidelity for every physics. The same physics at every point of the speed-accuracy spectrum, from heavily batched training with XPBD or VBD to final validation with IPC. Same scene, same API, pick your tradeoff.

• Invent physics level-of-detail (LOD). Rendering has had LOD for decades, physics is still waiting. Simulate at full fidelity what agents interact with and see, coarsely what they do not.

• Heterogeneous environments. Every parallel world can hold a completely different model: different bodies, joints, and collision geometries.

• Adaptive timesteps per island. Error-based control with Runge-Kutta Dopri5, and Time-of-Impact stepping during collision detection, as done in…