Manufacturing accounts for roughly a fifth of German gross value added and employs some 5.5 million people, yet the fastest-growing part of it cannot be automated economically. Individualisation has pushed producers into high-mix, low-volume work: small batches, frequent changeovers, hundreds of part variants.

Conventional automation does not pay there. Every gripper is built for one part, every programme is written for one sequence. Where parts arrive unsorted, differ from piece to piece and require process-specific force, fixed automation amortises only over long runs. So machine tending, bin picking and line feeding remain manual in one of the world's highest-wage locations, precisely where labour is scarcest.

Robot hands would solve it. Until now the software to control them reliably in production has not existed.
HandOS, an adaptive manipulation layer that teaches multi-finger robot hands to grasp, reorient and place parts autonomously. The policy is trained in simulation on NVIDIA Isaac and generalises across parts, so a new variant requires neither new tooling nor new programming. Object pose estimation, sensor fusion and force control handle the uncertainty that defeats conventional cells.

The product is software. It runs on a platform assembled from standard industrial robots, force-torque sensors and vision systems, which removes any dependence on humanoid hardware timelines, and it is ready for humanoid platforms as they mature. Commercially it is sold either as a turnkey cell: hardware once, software on subscription, or a pure SaaS license that can be purchased by integrators.

Where we stand: a spin-out from Fraunhofer IPA, eleven people, around EUR 650k raised from business angels, lab benchmarks running, paid pilot with a major industrial customer, first purchase commitment targeted for this year and two on-site deployments planned for 2027.