SOLAR-GP: Sparse Online Locally Adaptive Regression Using Gaussian Processes for Bayesian Robot Model Learning and Control

Abstract

Machine learning methods have been widely used in robot control to learn inverse mappings. These methods are used to capture the entire non-linearities and non-idealities of a system that make geometric or phenomenological modeling difficult. Most methods employ some form of off-line or batch learning where training may be performed prior to a task, or in an intermittent manner, respectively. These strategies are generally unsuitable for teleoperation, where commands and sensor data are received in sequential streams and models must be learned on-the-fly. We combine sparse, local, and streaming methods to form Sparse Online Locally Adaptive Regression using Gaussian Processes (SOLAR–GP), which trains streaming data on localized sparse Gaussian Process models and infers a weighted local function mapping of the robot sensor states to joint states. The resultant prediction of the teleoperation command is used for joint control. The algorithm was adapted to perform on arbitrary link manipulators including the Baxter robot, where modifications to the algorithm are made to run training and prediction in parallel so as to keep consistent, high-frequency control loop rates. This framework allows for a user-defined cap on complexity of generated local models while retaining information on older regions of the explored state-space.