UC San Diego

In this work, a data-efficient method is applied to learn a model of the dynamics of a self-folding robot driven by vibration. These robots can be autonomously fabricated and deployed, but complex dynamics lead to challenges in modeling. Learning from a limited set of observed experiments, a model is developed to control the locomotion of the robot along a desired trajectory. The model is fit assuming a probabilistic Gaussian model and a neural network. The two methods are benchmarked against a differential drive algorithm.