Abstract

The field of soft robotics is a growing research interest due to the unique benefits of soft robots over traditional robots, such as durability, compliance, and low-cost. These benefits allow soft robots to perform a variety of tasks that would be difficult for traditional robots, such as throwing, hammering, or whole-body manipulation. However, there is a significant variety of materials, actuators, and designs used in soft robotics. While helpful for completing more tasks, this variety also makes efficient modeling and control of soft robots difficult. In this thesis, an open-source repository, Modeling for Learned Dynamics (MoLDy), is presented. This repository is designed to assist other researchers in creating learned dynamic and kinematic models for novel soft robots. The workflow is data-driven, meaning that no analytical models or prior knowledge of the system is required. The repository includes functionality to create synthetic datasets, optimize neural network hyperparameters, train a variety of learned models, and validate the accuracy of the learned models in open-loop prediction and closed-loop control tasks. Users of the repository will also have access to several case studies, including simulation and hardware systems, that walk through the process of generating and validating learned models. As neural networks can require significant amounts of training data to extract patterns in the data, this thesis also explores fine-tuning, which is a method to reduce the need for real-world data. We show that a learned model trained on one system can be fine-tuned on data from a different system to achieve similar performance with as little as 13.8% of the data from the target system. The contributions within this thesis will enable other researchers to model and control a wide variety of soft robots with a simple approach. Additionally, the published code provides examples that facilitate implementation on novel systems and benchmarking with other state-of-the-art methods.

Degree

College and Department

Ira A. Fulton College of Engineering; Mechanical Engineering

Rights

https://lib.byu.edu/about/copyright/