You'll help define the simulation substrate Luma uses to train general-purpose robot policies — a faithful, controllable simulation of the world built on our generative video and 3D models. You'll sit at the boundary between generative models and classical physics simulation, and decide where each one earns its keep.
It's a role at the edge of generative rollouts and physics engines, building environments and evaluation harnesses trainable at scale. It fits someone with real simulation-systems experience who's shipped tools other people used. If you want a pure-research or pure-graphics seat, this lives in the hybrid between them.
What You'll Own
Design simulation environments that are visually rich, physically plausible, and trainable at scale — a hybrid of generative rollouts and physics-engine scenes.
Build the evaluation harness that shows whether the world model is good enough to train robots on (sim-to-real gap, physical consistency, long-horizon coherence).
Develop differentiable and GPU-accelerated simulation pipelines where they unlock new training signal.
Drive the asset, scene, and task generation pipelines, including using Luma's own generative stack to bootstrap diversity.
Collaborate with world-model researchers upstream and policy-learning researchers downstream.
First 90 Days
One way the first 90 could unfold.
Days 1–30 — Immerse & Diagnose: Learn the generative stack and where physics simulation and generative rollouts each fit.
Days 30–60 — Ship & Validate: Build a first hybrid simulation environment and an evaluation harness for it.
Days 60–90 — Scale & Systemize: Scale the asset, scene, and task pipelines and make the eval a reliable training signal.
What You Bring
A strong background in robotics simulation, computer graphics, physics-based modeling, or generative 3D, by degree or practical record.
Fluency in Python and C++, and deep familiarity with at least one production physics engine (Isaac Sim/Lab, MuJoCo, Bullet, PhysX, Drake).
A track record building simulation systems other people actually used.
Nice to Have
Research on sim-to-real transfer, domain randomization, differentiable simulation, or neural-rendering-based simulation.
A game engine, CGI, animation, or photogrammetry background.
Publications at top venues (CoRL, RSS, ICRA, NeurIPS, SIGGRAPH).
About Luma: Luma's mission is to build unified general intelligence that can generate, understand, and operate in the physical world. We believe multimodality is critical for intelligence — the next step beyond language models comes from vision. Luma is an equal opportunity employer.