Modern manufacturing increasingly relies on computer vision systems to detect defects, monitor processes, and ensure product quality. However, training reliable vision models remains challenging due to the scarcity of labeled data for rare defects, process anomalies, and failure modes. Collecting such data in real manufacturing environments is expensive, disruptive, and often infeasible. As a result, vision models trained on limited or idealized datasets often fail to generalize to real-world conditions. This project proposes TwinForge, a digital twin–driven framework that generates physics-aware synthetic visual data for manufacturing vision systems. Unlike conventional synthetic data pipelines that focus primarily on visual realism, TwinForge models the physical manufacturing process itself, capturing how defects originate, evolve, and propagate across multiple stages of production. The project focuses on an end-to-end manufacturing workflow that includes additive manufacturing (3D printing), robotic pick-and-place operations, and conveyor-based transport and inspection. By coupling physics-based simulation with realistic sensor modeling, TwinForge produces synthetic datasets that are causally consistent and representative of real manufacturing variability.

The TwinForge framework models the manufacturing process as a connected digital twin rather than a collection of isolated simulations. The pipeline begins with a digital representation of a manufactured part, continues through robotic handling, and ends with visual inspection on a conveyor system. At each stage, physical properties such as shape, mass, surface texture, and motion are preserved and passed forward, ensuring that downstream behavior reflects upstream manufacturing conditions.

Students will use open-source simulation and modeling tools to implement each stage of the pipeline. Three-dimensional part models will be created or imported using tools such as FreeCAD and Blender, where realistic printing defects—such as warping, surface roughness, or internal voids—can be introduced in a controlled manner. These defected models are then transferred into a physics simulation environment such as PyBullet, where robotic pick-and-place operations are simulated. Because the same digital object is used throughout the pipeline, printing defects naturally affect how the robot grasps and moves the object, leading to realistic handling anomalies such as slipping, misalignment, or unstable placement.