Candidate
Sheran Christopher Lalith Wiratunga
Title
Training Gaussian Process Regression models using optimized trajectories
Supervisor
Dana Kulic
Abstract
Quadrotor helicopters and robot manipulators are used widely for research and industrial applications, but are often difficult to accurately model. Supervised learning methods such as Gaussian Process Regression (GPR) can estimate a dynamics model from experimental data without requiring knowledge of the structure of the model. GPR has been used to learn the inverse dynamics of a manipulator, but has not been applied to quadrotors. Additionally, collecting training data for supervised learning can be difficult and time consuming, and poor or inadequate training data may result in an inaccurate model.
We propose a method for systematically collecting training data for a GPR model by parameterizing the trajectory and optimizing the trajectory parameters to maximized the GPR variance over the trajectory. This approach is tested in simulation on a quadrotor, and in experiments on a quadrotor and a 4-DOF manipulator.
GPR has a large computational cost, and we compare three sparse approximations of GPR. Finally, we evaluate different methods of training the GPR hyperparameters. These hyperparameters have a significant impact on the prediction made by the GPR model, and hyperparameter training is important for making accurate predictions.