Sparse Contextual Task Learning and Classification to Assist Mobile Robot Teleoperation with Introspective Estimation

Abstract

This report proposes a novel approach to learn from demonstrations and classify contextual tasks the human operator executes by remotely controlling a mobile robot with joystick, aiming to assist mobile robot teleoperation within a shared autonomy system in a task-appropriate manner. The proposed classifier is implemented with the Gaussian Process (GP). GP is superior in uncertainty estimation when predicting class labels (i.e. the introspective capability) over other state-of-art classification methods, such as Support Vector Machine (SVM), which is probably the most widely used approach on this topic to date. Moreover, to keep the learned model sparse to limit the amount of storage and computation required, full GP is approximated with a state-of-art Sparse Online Gaussian Process (SOGP) algorithm, to maintain scalability to large datasets without compromising classification performance. The proposed approach is extensively evaluated on real data and verified to outperform the baseline classifiers both in classification accuracy and uncertainty estimation in predicting class labels, while maintaining sparsity and real-time property to scale with large datasets. This demonstrates the feasibility of the proposed approach for online use in real applications.