Contact: Daniel Wälchli, [email protected]
🎯 Challenge Overview
Track: GenAI in Manufacturing
→ Problem Description:
Manufacturing companies still rely heavily on 2D technical drawings (PDF) to define how a part must be produced and inspected, even when a 3D CAD model exists. Critical product and manufacturing information (PMI) such as material, dimensions, tolerances, surface roughness, and notes are typically embedded in these drawings as visual annotations rather than machine-readable data.
As a result, machine programmers and quality engineers must manually interpret drawings, match the annotations to 3D geometry, and translate them into manufacturing and inspection instructions. This process is time-consuming, error-prone, and difficult to scale. It creates bottlenecks, drives up costs, and increases the risk of misinterpretation.
The core challenge is to automatically understand technical drawings and their product definition to the correct features in a 3D model. This requires combining document understanding and geometric reasoning to transform unstructured data into machine-usable product intelligence.
GenAI is a prime candidate for solving this problem, which would enable a shift from document-based workflows to true digital manufacturing pipelines, unlocking automation across the value chain.
→ Primary Objective:
Participants are asked to build a GenAI-powered solution that links technical drawing annotations to the correct features in a 3D model, with a specific focus on drilled and threaded holes.
In 3D CAD geometry, a cylindrical hole is typically represented in the same way regardless of whether it is a simple drilled hole, a tapped hole, or a threaded feature. The manufacturing intent is not visible from geometry alone. It is defined in the drawing through annotations such as thread callouts (e.g., M10 × 1.5), tolerance classes (e.g., 6H), fit specifications (e.g., H7 according to ISO 286), depth indications, and simplified hole/thread notes.
The objective is to automatically:
The goal is not to solve the entire drawing-to-model problem, but to demonstrate a technically sound and scalable approach for one high-impact and non-trivial manufacturing case where semantic interpretation is essential.
→ Datasets (type, format, access):
Detailed technical drawings and associated CAD models used in industry are typically proprietary to manufacturers. For this reason, no industrial datasets can be shared. This reflects a realistic constraint in manufacturing AI: solutions must work in environments where annotated, high-quality training data is limited.
To support experimentation and benchmarking, participants can use the following publicly available resource:
NIST CAD Models and STEP Files with PMI: https://catalog.data.gov/dataset/nist-cad-models-and-step-files-with-pmi-3bb54
Participants are encouraged to generate synthetic drawing–CAD pairs to scale their training data if needed.
→ Available APIs & Services: