The transition from Computer-Aided Design (CAD) to robotic manufacturing is a bottleneck in modern industry, often relying on manual intervention and expert knowledge, which hinders efficiency and introduces variability. This thesis presents an automated two-stage deep learning pipeline to generate robot code for laser-cutting operations directly from B-rep CAD models, preserving full geometric fidelity. The first stage employs an adapted BRepNet, a graph neural network architecture, to perform edge classification directly on the topological structure of the CAD model. By learning from geometric and topological features, this model effectively identifies manufacturing-relevant cut paths without resorting to precision-losing discretization methods. The second stage frames robot code generation as a sequence-to-sequence translation task. An encoder-decoder Transformer architecture is developed to translate the identified geometric primitives (lines and arcs) into a sequence of machine-executable instructions. This model utilizes a multi-head prediction system to simultaneously generate the heterogeneous components of each command. Experimental results on a dataset of structural I-beams demonstrate the pipeline's effectiveness. The BRepNet-based feature recognizer achieved 100% classification accuracy on the test set. The Transformer model successfully learned the complex syntax and implicit manufacturing logic, generating valid and logically coherent robot code. A final "Snapper" post-processing algorithm was introduced to align the model's continuous outputs with the deterministic geometric entities from the source CAD file, reducing coordinate errors by over 70% and achieving sub-millimeter precision. By successfully automating the CAD-to-robot code workflow, this research demonstrates a data-driven approach to enhance precision, speed, and autonomy in robotic fabrication.
La transizione dal Computer-Aided Design (CAD) alla produzione robotizzata costituisce un collo di bottiglia nell’industria moderna, poiché dipende spesso da interventi manuali che riducono l'efficienza e introducono variabilità. Questa tesi presenta una pipeline automatizzata di deep learning a due stadi per generare codice robot per il taglio laser direttamente da modelli CAD B-rep, preservando la totale fedeltà geometrica. Il primo stadio impiega una versione adattata di BRepNet, una rete neurale a grafi, per classificare gli spigoli basandosi sulla struttura topologica del modello. Apprendendo da caratteristiche geometriche e topologiche, il sistema identifica i percorsi di taglio senza ricorrere a discretizzazioni che ne ridurrebbero la precisione. Il secondo stadio modella la generazione del codice come una traduzione sequence-to-sequence: un’architettura Transformer encoder-decoder traduce le primitive geometriche (linee e archi) in istruzioni macchina, utilizzando un sistema multi-head per generare simultaneamente i componenti eterogenei di ogni comando. I risultati su travi strutturali a I confermano l'efficacia del metodo: BRepNet ha raggiunto un'accuratezza del 100%, mentre il Transformer ha appreso con successo la complessa sintassi e la logica manifatturiera implicita. L'introduzione dell'algoritmo di post-processing "Snapper" ha permesso di allineare gli output continui del modello alle entità geometriche deterministiche del CAD originale, riducendo gli errori di coordinata di oltre il 70% e garantendo una precisione sub-millimetrica indispensabile per gli standard industriali. Automatizzando l'intero workflow CAD-to-robot, questa ricerca propone un approccio data-driven per accrescere precisione, velocità e autonomia nella fabbricazione robotica avanzata.
Automated pipeline for robot code generation from CAD models using deep learning
LATIFI, AIDIN
2025/2026
Abstract
The transition from Computer-Aided Design (CAD) to robotic manufacturing is a bottleneck in modern industry, often relying on manual intervention and expert knowledge, which hinders efficiency and introduces variability. This thesis presents an automated two-stage deep learning pipeline to generate robot code for laser-cutting operations directly from B-rep CAD models, preserving full geometric fidelity. The first stage employs an adapted BRepNet, a graph neural network architecture, to perform edge classification directly on the topological structure of the CAD model. By learning from geometric and topological features, this model effectively identifies manufacturing-relevant cut paths without resorting to precision-losing discretization methods. The second stage frames robot code generation as a sequence-to-sequence translation task. An encoder-decoder Transformer architecture is developed to translate the identified geometric primitives (lines and arcs) into a sequence of machine-executable instructions. This model utilizes a multi-head prediction system to simultaneously generate the heterogeneous components of each command. Experimental results on a dataset of structural I-beams demonstrate the pipeline's effectiveness. The BRepNet-based feature recognizer achieved 100% classification accuracy on the test set. The Transformer model successfully learned the complex syntax and implicit manufacturing logic, generating valid and logically coherent robot code. A final "Snapper" post-processing algorithm was introduced to align the model's continuous outputs with the deterministic geometric entities from the source CAD file, reducing coordinate errors by over 70% and achieving sub-millimeter precision. By successfully automating the CAD-to-robot code workflow, this research demonstrates a data-driven approach to enhance precision, speed, and autonomy in robotic fabrication.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_LATIFI_Thesis.pdf
non accessibile
Descrizione: Thesis text
Dimensione
4.8 MB
Formato
Adobe PDF
|
4.8 MB | Adobe PDF | Visualizza/Apri |
|
2026_03_LATIFI_Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary text
Dimensione
2.26 MB
Formato
Adobe PDF
|
2.26 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/252855