Data engine for physical AI
Foundation models already understand the world. nablaX provides the missing action data that teaches robots how to act in it.
We record how the physical world responds when people handle it: what they look at, how they move, how much force they apply, what changes — and turn that into training data robotics teams can license.
Why now
There is no shortage of data about how the world looks. There is very little about what happens when you reach into it.
Perception, language and reasoning are no longer the constraint. What is missing is the bridge from understanding to acting, and that bridge is made of data.
Humanoids and capable manipulators are moving from lab demos to procurement decisions. Fleets are being ordered. Those fleets need training data that does not yet exist at the required quality.
Not "more data" — a particular kind: egocentric, gaze-aligned, hand-pose-accurate, tactile, with object state change labelled. Abundant for how the world looks; effectively absent for how it responds.
Machines that have felt the world, not only watched it. That is what nablaX builds the data for.
“Everything that moves will be autonomous.”
What we deliver
We are building the data engine for that gap: contact-rich datasets for humanoids and robotic manipulation — collected, annotated and validated, as a licence, as a commissioned collection programme, or both. The first collections run against pilot clients' task specifications.
Contact-rich multimodal datasets you can train on and licence commercially, with provenance attached to every capture.
Continuous data collection captured to your task specification — your objects, your environments, your manipulation problem.
Annotation, validation and evaluation designed to answer one question: did the data move your model, and how do you know?
Consented capture, privacy-preserving by construction, audit-ready provenance — built for a procurement review, not retrofitted to one.
| Criterion | Generic video datasets | nablaX |
|---|---|---|
| Signal | How the world looks | How it feels and changes through interaction |
| Provenance | Scraped, unclear provenance | Consented, privacy-preserving, audit-ready |
| Metric | Volume as the metric | Signal density as the metric |
| Licensing | Research-grade licensing ambiguity | Enterprise-licensable by construction |
As twin brothers, we have always enjoyed building and competing together. One of those competitions was the robo.innovate hackathon at TUM, where our eleven-person team won with GrowBot, a tree-care robot.
What stuck with us was a technical insight: perception is only half the problem. The real challenge is enabling robots to interact with the real world precisely and reliably. nablaX was created to solve exactly that challenge.
Co-founder, M.Sc. in Nanophysics
“After studying physics, 5 years of building Data & AI systems professionally and building robots as a side obsession, I realized I wasn't pursuing three separate interests. I was building toward nablaX. I am excited to build at the intersection of intelligence and automation and help shape the future.”
Co-founder, Ph.D. in Economics
“I have spent a lot of time analysing how economies respond to change. Robots that can reliably operate in the physical world will be extremely transformative. I'm very excited to work on building that future and contribute to making it a reality.”
Currently open to
Founding team
We are adding robotics depth to the founding team — as a co-founder or as an advisor. Production robotics or perception experience.
Introduce yourselfPilot clients
Robotics teams with a manipulation problem and no data for it. We will collect against your task specification.
Describe your taskData partnerships
Organisations with access to real work environments where contact-rich capture is possible.
Tell us about your siteHiring
Not yet formally. We still want to meet people who would want to build this.
Say helloLeave your address and we will be in touch.