Designing Natural Wood Log Structures with Stochastic Assembly and Deep Learning

Abstract

Advances in contemporary 3D scanning and bespoke robotic technologies have enabled architectural structures to be directly constructed from naturally grown wood. However, design precedents using natural wood logs are still dominated by design approaches using predefined geometric models. The limit of this approach lies in the necessity to model the form of every structural member based on the captured geometries of all the materials before design begins. Moreover, human designed rules for joining irregular components are limited and solutions are prone to be limited by empirical knowledge. In this paper, we introduce a method for assembling natural wood log structures with higher goals autonomously using robotic stochastic assembly and deep learning. The novelty of this method is that the design of structures does not rely on prior-knowledge of the to-be-assembled materials but is generated by assembling materials iteratively. A vision system with a position suggestion network based on convolutional neural networks (CNNs) was implemented and trained to drive an industrial robotic arm for negotiating between the topological changes from potential connections and the local assembly constraints of the log. A robotic hand-eye coordination database recording the assembly of birch logs has been established and small-scale wood structures were built by the trained robot. Results show that our robot can find desired structural configurations autonomously and can assemble unfamiliar batches of wood logs. The cost and gain of using stochastic assembly and deep learning as a design strategy are discussed and future research on using different learning strategies and large-scale implementations are laid out.