Robotics ML and Research Recruitment

The people doing this work are spread thinly across a handful of academic labs and industrial research groups, and they do not share a job title. The same person is a Research Scientist in one organisation, a Robotics Engineer in the next and an ML Engineer in the one after that. Searching either community by title finds whoever sits squarely inside it, which is rarely the person you want.

The separator we screen on is whether somebody has had a policy running on real hardware. Plenty of people can train one in simulation and show you a convincing video, and the current wave of vision-language-action models has made an impressive demo cheaper to produce than it has ever been. Far fewer have crossed into a physical system, where the contact dynamics are wrong, the sensors are noisy and the failure modes look nothing like the ones in the simulator. That crossing is where robot learning actually gets hard, and it is the clearest signal of whether someone has done the work or read about it.

Data is the other constraint. Robot data is expensive to collect and most of it is never shared, so the number of people who have trained on it at any real scale is small and heavily concentrated in the places that own it. Where a brief needs that specifically, the realistic pool is measured in dozens. The same applies to VLA work: the people who have trained one on real robot trajectories rather than fine-tuned somebody else’s checkpoint are a much smaller group than the job titles suggest.

We work the learning side of robotics rather than the hardware side, and we will say so at kickoff if what you have described is really a mechanical or electrical engineering search. Getting that wrong costs a month neither of us gets back.