Скачать или смотреть Inverse Design of Origami for Trajectory Following using Deep Learning

ABSTRACT: This work presents a Variational Autoencoder (VAE) based approach for the inverse design of rigidly foldable and single degree of freedom origami mechanisms for trajectory following. The work first develops an automated graph grammar procedure based on the Principle of Three Units (PTU) to generate a large dataset comprising 8 million crease patterns. The distribution of origami crease patterns in the dataset is projected onto a continuous latent space using the VAE from which new designs can be sampled. The aggregated posterior distribution is approximated using normalizing flows to receive a Gaussian latent space and improve the reconstruction accuracy of crease patterns. By utilizing the continuous lower-dimensional representation of crease patterns in the latent space, the work proposes an inverse design method to design origami for trajectory following. Results demonstrate the capability of the approach to efficiently generate large datasets of rigidly foldable patterns, which in principle, can be used beyond the scope of this work. The results also show that the learned lower dimensional representation of crease patterns enables efficient optimization of crease patterns using a series of target trajectories of varying complexity.

KEYWORDS: Artificial Intelligence, Mechanism Synthesis and Analysis, Application of Machine Learning, Kinematics, Dynamics and Control of Mechanical Systems

VIDEO DESCRIPTION: Optimized origami performing trajectory following with a vertex for 9 different target trajectories. The target trajectories include Sine Wave, Helix, Simple Spline, Spherical Bi-Angle, Complex Spline, Pringle, Ellipse, Parabola and Triangle.