Machine Learning Approach for Tensegrity Form Finding: Feature Extraction Problem

Abstract

In this paper, the problem of form finding for a tensegrity structure with the use of a machine learning pipeline, including feature extraction and regression, is studied. Tensegrity robots present a range of new opportunities in a number of areas in robotics, however the lack of efficient and scalable tools for controlling their motion slows down their deployment. This study aims to add a new tool for solving one of the basic problems encountered in tensegrity robot control: finding a stable equilibrium state (and its inverse task). Using machine learning approach makes the solution scalable and potentially faster than iterative optimization-based methods. The paper provides a view into the issue of feature extraction, and provides implementation of a number of feature extraction methods: Principle Component Analysis (PCA), Kernel PCA, undercomplete, denoising and sparse autoencoders.