Lumen Labs is building the cognitive layer for physical AI. Nearly every major robotics effort today teaches machines by showing them thousands of hours of human teleoperation. We think that paradigm has a ceiling, and that the path to robots that genuinely adapt runs through a different kind of learning and we are trying to figure out what that is. We are primarily focused on nature inspired AI architectures for unstructured environments (construction sites, mining, oil pipelines, battlefield — places where no pre-training will help).
The founders are Harry (2x founder, Thiel Fellow, previous 1517 Fund) and Daniel (OSV Fellow, previously Wolfram Institute, The Residency) with a background in cognitive science, robotics and sensors. Pre-seed from top VC funds and operators at robotics/AI companies (Anthropic, Waymo, etc.).
Location: In-person in San Francisco.
You'll own the gauntlet: our test library for putting new architectures through their paces. The gauntlet is a Gymnasium-style interface on both sides, environment and model, so that anything built to spec is compatible with everything else we have. Today it runs mostly on MuJoCo and Gazebo with dm_control-style abstract tests. We are moving toward Isaac Sim, and it's early enough that if you have better ideas on sims or compute, it is not hard to pivot.
The job is building worlds and keeping them running. We test on a very wide variety of tasks: GPS-denied drone navigation, construction site navigation, abstract control problems, automated boat cleaning, automated forklifts. Sometimes that means adapting an existing sim, sometimes frankensteining something new from pieces that exist and fixing the researcher code along the way, and sometimes building from scratch. Our models mostly sim2real well, so a quick and dirty sim is often enough, but a higher quality one is always better.
If it has to be simulated for training or inference, it's yours: the physics, the sensors, digital twins of our real platforms, domain randomization, rendering. You'll work closely with our model engineer on what the gauntlet measures and with our hardware engineer on the sim2real gap.
This isn't a role where you run the experiments yourself. You're the one who makes it smooth and easy for everyone else to.
We provide competitive compensation (salary + equity), and we level and pay based on what you've built rather than what's on your résumé. We also have health insurance and provide relocation support.