We build the operating system that trains, deploys, and scales humanoid robots in physical retail in the EU. Hardware is converging. The software layer is what wins.
Demographics will shrink the EU labour pool by roughly 1 million potential workers every year. Night-shift shelf replenishment is one of the hardest retail roles to fill — and the demographic pressure is structural, it will not reverse.
Industry average out-of-stock rate is 8%. Physical inventory accuracy sits at just 63–65% while the product is physically in the back room. Lost sales, frustrated customers, manual root-cause analysis.
More than €30 million of food is binned in EU stores every single day — poor rotation logic, inconsistent shelf checks, no real-time expiry visibility. By 31 December 2030, EU member states must cut per-capita food waste 30% across retail, food service and households combined.
For the first time, three forces meet simultaneously: humanoid hardware reaching commercial readiness, the deepest labour shortage in decades, and AI/computer vision mature enough for real-time retail operations. The window is now.
Like Android operating different manufacturers' phones, Roboshelf AI runs any humanoid robot in any retail store — training it, deploying it, and managing its entire operation.
We don't bet on one robot manufacturer. The same model and recipe has been retrained on three manufacturers' humanoids in simulation — Unitree (80%) and Booster Robotics (86%) clear the bar, Fourier Intelligence is still in tuning (20%). The platform is the moat, not the metal. When the hardware market consolidates, we win either way.
A hard-to-replicate asset base: a 3D product database with grasp parameters per SKU, and safety compliance every humanoid needs before it can enter an EU store. Together they are what no robot manufacturer delivers — a store-ready system.
AI infrastructure built on European soil sits inside the EU's tech sovereignty priority — a structural tailwind across innovation funding, retail partnerships, and policy alignment. GDPR compliance is native to our stack, not retrofitted. We operate in the same market as our customers.
UnifoLM-VLA-0 fine-tuned on 1,000 scripted-expert demonstrations of a push task — moving an object into a target zone. Evaluated over 50 independent MuJoCo simulation episodes. Achieved 80% task success rate — 10× improvement over the v1 baseline in a single iteration cycle, at €0 infrastructure cost. The same model and recipe — unchanged — was then retrained on two further manufacturers' humanoids: a Booster Robotics T1 (86%, beating the G1) and a Fourier GR1T1 (full pipeline runs, policy still tuning at 20%). Two of three manufacturers clear the acceptance bar so far — the platform is not tied to one robot. This is one manipulation primitive, not full shelf restocking — grasping and physical transfer come next.
The pre-seed round is aimed at hiring a robotics engineer and AI training expert to build Milestone 3.