We're building a physics-grounded world model as the shared "brain layer" for general-purpose robot fleets — hardware-agnostic, cross-embodiment. The core bet: a learned world model that recalibrates itself from real rollouts plus physics priors, trained through self-play, beats approaches capped by human demonstration data. You'd own a large piece of the research that decides whether that bet holds.
_ Define and drive research on world-model architectures — dynamics models, physics priors, self-play curricula for robotic control.
_ Design and run self-play training loops that scale, and diagnose where they break.
_ Own scaling experiments and ablations end-to-end, from hypothesis to written-up result.
_ Work directly with real rollout data from our robot fleet access (Unitree G1/B2W, ROSbot 3, Z1) to close the sim-to-real loop.
_ Collaborate with our academic network (TUM, MIRMI), publish where it makes sense, and help shape the research agenda as one of the first hires.
_ Strong research background in reinforcement learning, world models, or model-based control — PhD or equivalent industry research experience.
_ Hands-on experience with self-play, model-based RL, or learned dynamics models (video-generation and physics-informed learning backgrounds also welcome — we care more about depth than the exact subfield).
_ Comfortable owning an open-ended research problem without a lot of hand-holding; this is a founding-stage team, not an established lab.
_ Can read and reason about robotics or physics simulation code, even if that's not your primary focus.
_ First-author publications on world models, model-based RL, self-play, or video/3D generative modeling.
_ Experience with SE(3)-equivariant architectures or other structured/geometric priors.
_ Track record of mentoring junior researchers or informally leading a small research effort.
_ Prior sim-to-real experience on real robot hardware.
