Data-based Iterative Human-in-the-loop Robot-Learning for Output Tracking

Abstract

This article develops a data-based approach for improved iterative robot-learning for output-tracking from novice human-in-the-loop demonstrations. While nominal human-response models can be used to improve iterative learning, the convergence can be slow due to variations in each human operator. The major contribution of this article is to use data acquired during iterative learning to learn the unknown human intent as well as the human-response model, and thereby, improve convergence when learning future trajectories. The proposed method is applied to a robot arm, and results indicate both an increase in the range of frequencies where tracking is achieved (from 0.2 Hz to 0.5 Hz) and an increase of 103% in the tracking error reduction for the same number of iterations.