Research on world models and self-play only moves as fast as the data and infrastructure underneath it. We need someone who can turn "we have robot fleet access" into a working pipeline: collecting rollouts, closing the sim-to-real loop, and making experiments reproducible and fast to run. You'd own that layer.
_ Build and maintain the data pipeline connecting real robot rollouts to training infrastructure.
_ Own sim-to-real transfer — closing the gap between simulated training and real hardware performance.
_ Build tooling for large-scale training experiments: logging, evaluation harnesses, reproducibility, fast iteration loops.
_ Work closely with our research scientists to translate architecture and algorithm ideas into running systems.
_ Help shape engineering standards as one of the first hires — there's no legacy codebase to inherit or work around.
_ Strong software engineering background with real experience in robotics, ML infrastructure, or simulation systems.
_ Hands-on experience with at least one of: ROS/ROS2, robot simulation (Isaac Sim, MuJoCo, or similar), or large-scale ML training infrastructure.
_ Comfortable working close to real hardware — debugging when something breaks on an actual robot, not just in simulation.
_ Can move between "quick and dirty prototype" and "this needs to be reliable" depending on what the moment calls for.
_ Experience with reinforcement learning pipelines specifically (not just supervised/imitation training infra).
_ Background in sim-to-real transfer research or robot learning benchmarks.
_ Experience standing up ML infrastructure at a very early-stage team (few or no existing systems to build on).
_ Familiarity with physics simulators beyond a single ecosystem.
