Feature-Based Geometry Sorting Using Autoencoders: A View Based Approach To Found-Object Fitting

At the intersection of robotic automation and sustainable design, this research tackles issues of bio-digital processes by developing a machine-learning and multi-agent framework for finding natural source objects and matching their geometric features to target geometry patches. AI-driven multi-agents enable the collection of source artifacts, while feature matching using stacked convolutional autoencoders addresses the reusability of found parts to refit geometric surface patches into target geometries. By prioritizing available materials, the research aims to reduce environmental impact and promote efficient resource management. Given the high energy and material consumption of conventional metal forming processes, this work presents a two-stage methodology for sourcing reusable natural dies. This approach employs advanced machine learning techniques to enhance the applicability of natural dies in sheet metal formation, minimizing deviation of the target geometry’s surface in the machining process with limited waste.