Role description
HubAbout Hub
Hub (Y Combinator 2026) collects real-world human and robot data to train robots, and works with some of the most recognised robotics and AI labs in the United States. Today Hub specialises in egocentric data collected at a scale and speed no lab could reach on its own: every contributor is identity-verified (KYC) and every contribution goes through several layers of quality control. Scalable QA of egocentric data, run by our Refinery team, remains a core company priority.
Hub Labs, headquartered in Toulouse, is where Hub turns that data into working robot behaviour. Within Hub Labs, the Brain team, led by Guillaume, works on robot learning across humanoids and other types of embodied machines.
The role
This is a hands-on physical robotics role. You take existing robot policies, make them run on our robots and data, and measure honestly what works. Your first goal is a reproducible baseline on real hardware, followed by one measured data or model experiment that tells us something we did not know.
Location: San Francisco, USA OR Toulouse, France. In-office and hands-on with physical robots; this is not a remote role.
Location terms: San Francisco-specific compensation and visa sponsorship arrangements are to be confirmed. The listed EUR range does not establish US employment terms.
What you will do
- Reproduce accessible existing policies and get them running on our robots.
- Adapt observation and action spaces to our embodiments and sensor setups.
- Build versioned datasets with clean train, validation and held-out splits.
- Run controlled, held-out trials on physical robots, with protocols others can repeat.
- Deliver a documented baseline, then design and run at least one measured data or model experiment against it.
- Report what the numbers show and what they do not, and catch the bugs that quietly distort a result.
- Work closely with research scientists, deployment engineers and the data teams who collect and QA our data.
Who we are looking for
- Hands-on experience training or adapting robot-learning policies (for example imitation learning, diffusion or transformer policies, VLAs, or RL), not only reading about them.
- Experimental discipline: clean splits, controlled comparisons, reproducible runs and honest reporting.
- Experience running learned policies on real robots: sensors, calibration, latency and the failures that only show up in the physical world.
- Excellent Python and PyTorch; comfortable with Linux, GPUs and robotics software.
- Experience with humanoids, manipulators, mobile robots or other embodied systems.
- Fluent English, written and spoken.
Nice to have
- Experience with LeRobot, MCAP or similar robot data formats.
- Experience with egocentric video, hand pose, depth or IMU data.
- Simulation experience (MuJoCo, Isaac or similar) used alongside physical trials.
- Open-source contributions or published work in robot learning.
Why Hub Labs
- Real-world data at a scale few labs have, and a team that can collect more when an experiment calls for it.
- A founding seat: you help set how the lab trains, evaluates and ships robot behaviour.
- Your work goes straight into the data, models and benchmarks Hub builds with leading AI and robotics labs.
How to apply
Send a short application, in your own words:
- Your CV
- The policy you trained or reproduced on a real robot that you are most proud of, and how you measured it
- Your GitHub and Hugging Face links