System Similarity, Performance Guarantees, and Asymmetry in Transfer Learning for Robotics

Abstract

In robotics literature, transfer learning has been proposed in learning-based control frameworks to leverage existing experience from a source robot or task to accelerate or improve the learning process on a target robot or task. It is often assumed without analysis that incorporating prior experience will be beneficial. For robotics applications, inappropriately transferring experience can be unsafe or inefficient. This thesis presents two approaches to this problem. First, we propose an experience selection algorithm based on a dynamics similarity characterization to select source experience that best improves target robot performance. Second, we derive an upper bound on the tracking error of a target robot using an inverse dynamics module from a source robot, and demonstrate how the bound can guarantee a performance improvement on the target robot prior to conducting transfer. We further illustrate that inverse module transfer is asymmetric. We demonstrate both approaches in quadrotor trajectory tracking experiments.