Shape2Gcode: Direct G-code Generation from 3D Shape Data for Automated Manufacturing

Sangmin Hong, Dohee Cho, Mohsen Yavartanoo, Kyoung Mu Lee

Keywords: CNC manufacturing, manufacturing AI, CAD, 3D Vision, Reinforcement Learning

TL;DR: First AI model that generates G-code for CNC manufacturing from 3D data.

Abstract: Modern manufacturing relies on Computer Numerical Control (CNC) machines, which execute machining operations using G-code, a programming language that defines tool movements, cutting paths, and machining parameters. Despite advancements in automation, generating G-code still requires significant human intervention and reliance on Computer-Aided Manufacturing (CAM) tools. To address these challenges, we propose Shape2Gcode, an end-to-end framework that directly generates optimized G-code from 3D shape data. Our approach leverages reinforcement learning to optimize key machining parameters, including tool radius, milling depth, and toolpath strategies. Additionally, Shape2Gcode incorporates a tool orientation selection module to determine optimal rotation matrices, enhancing the flexibility and precision of the machining. We evaluate Shape2Gcode on CNC manufacturing tasks using the ABC and ShapeNet datasets, comparing its performance against existing CAD reconstitution and CNC automation methods. Experimental results demonstrate that Shape2Gcode outperforms conventional approaches in reconstruction accuracy, significantly reducing the need for manual intervention. By optimizing G-code generation and minimizing manual adjustments, Shape2Gcode improves CNC manufacturing efficiency, lowers costs, and enables more automated machining workflows.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 15585