Robotics Data Researcher

What we need

• Research depth in robot learning. A PhD, or a research record that does the same job, in robotics or machine learning. Close to finishing counts.
• You have trained and evaluated policies yourself, in simulation or on hardware. You know where robotics evaluations go wrong, and how to stop it.
• You move fast. You run the experiment early and write up what you found, including when it is the opposite of what you hoped.
• You can explain it. To a researcher, to an operations lead and to a client.
• AI native. You already work with AI tools and coding agents daily, and you build what you need without waiting for anyone to show you.
• Professional English, used daily with the team and with clients.

• Based in Barcelona, or moving here. The role is on-site with the team, and Sovrano helps with the move.

Physical AI needs its own data layer: the data, the benchmarks and the evidence that connect real-world skill to the models that run robots. Sovrano is building it. This role leads the research that tells Sovrano, and the field, what actually makes robots better, and then makes sure it gets built. Sovrano AI is a data and research lab in Barcelona. Sovrano brings real experts into building AI: they do the work, Sovrano captures their judgment, and turns it into the data, benchmarks and evidence that frontier AI labs, robotics companies and enterprise AI teams use to train and evaluate their models. Behind it sit a network of 400,000 or more professionals, 100 or more partner universities across Europe, EFMD certification, 23 languages and EuroExec, Sovrano's own benchmark. Language models got good because the internet already existed. Physical AI has no internet: every robot that learns useful work will learn it from data someone deliberately built, tested and proved. What data a policy needs, how to judge its quality and how to test a policy against real work are still open across the field. Europe has the experts and the research talent to answer them, and this is a founding seat on the research team that intends to. The week is about 80 percent research: training, evaluating, publishing. The other 20 percent is communication and coordination, with clients, with marketing and with the data team. You report to Pau Arnal, Founding Researcher, who leads the research team and sets its agenda, and you work directly with him and with the CEO. Said plainly: scope moves as results and clients come in. Some weeks you train models all week, some weeks you are on a call explaining why a data spec should change. Much of the research infrastructure you need is yours to build. The trade is ownership most researchers wait a decade for: a research agenda that is genuinely yours to shape, and results that change what the company does the following week.

What you own

• Proving the data works. Train and fine-tune policies on Sovrano's datasets and measure the lift. When the data is not pulling its weight, find out why and say so plainly, because that call decides where the company puts its next bet.
• The benchmarks. Design the benchmarks that tell physical AI how good a policy really is on real work. Evaluation for physical tasks is still an open problem, and Sovrano intends to define it.
• Training the models. Pick the method each question needs, and run it end to end.
• The direction of the data layer. Your results decide which tasks, which embodiments and what quality bar the data layer is built around.
• The next dataset spec. You work with the data team so your feedback shapes how data gets made, and you verify the quality of what ships.
• Explaining the evidence. When a robotics company asks what data its policies need, your results are the answer, explained to a researcher, an operations lead and a client.
• Publishing. Papers at top venues, open releases, blog posts. Your name goes on the work, and the work goes out in public.