System Identifies 3D Parts Without Retraining the AI Vision Model

Published on March 19, 2026 | Translated from Spanish

A team from KU Leuven, Materialise, and Iristick has developed a method to recognize parts manufactured with 3D printing. The proposal avoids the usual step of retraining an artificial intelligence model every time a new component is added. The solution relies on the original CAD models to create visual references.

A robotic arm inspects real 3D parts alongside their CAD models on screen, identifying them without needing to retrain the artificial vision system.

From CAD to visual prototype: the identification process 🤖

The system generates prototype representations of each object from multiple rendered views of its CAD file. When an operator, equipped with smart glasses, captures an image of a physical part, the artificial vision model compares it with that bank of prototypes. It assigns the part to the class with the greatest similarity. This few-shot learning approach only requires the digital model, without new data collection or training.

Goodbye to "And who do you belong to?" in the parts drawer 🕵️

This solves the classic workbench dilemma: that printed part which, separated from its manufacturing ticket, becomes a mysterious plastic artifact. The system acts as a companion with a photographic memory that never forgets a face (or a CAD geometry). It promises to end meetings of anonymous parts asking each other which project they belong to.