Build physical intelligence with us.

Physical AI is the next frontier of AI, and it is unsolved. Models that can reason, write, and see still can't act in the real world and solve physical tasks. The intelligence that moves the physical world hasn't been built yet.

We believe it gets built by learning the way humans learn: data-efficient, compute-efficient, and adaptive. Show it a new task, give it a few corrections, and it does the job the same day, with the dexterity real work demands, not just pick-and-place. Today's robot learning does the opposite: fleets of teleoperators and data centers of compute, for one task at a time.

We don't believe this is just an architecture problem. It's the whole system: what data you capture, how you train, how the policy runs on the robot, and what you learn when it fails. So we push the frontier end to end, world models at the core, developed closed-loop and tested in the real world, not in lab demos.

Who we are

Oktonex was founded this year by Katrin (researcher from MPI, Uni Tübingen and Wayve) and Frederik (ex founding engineer at a robotics startup), backed by Andreas Geiger (Prof. & CEO KE:SAI), Christoph Lassner (Co-Founder World Labs), Samuel Albanie (DeepMind) and many more. Part of the NVIDIA Inception program.

We're based in Tübingen (Cyber Valley) — KE:SAI, MPI-IS, ELLIS and the university around the corner, factories an hour away.

Who thrives here

• High agency. You don't wait to be told what to do. You decide with incomplete information and pull things over the line.
• Mission first. You're here because the intelligence that moves the physical world matters to you.
• Multiplier, not soloist. You make the whole team faster.
• Comfortable with ambiguity. At this stage, every job description is just a starting point.

A small team with a high bar and short paths. Decisions take minutes, not meetings. Hard problem, small team, real robots, no excuses.

Open Roles

Questions? Reach us at [email protected].