Inferring Intent for Novice Human-in-the-Loop Iterative Learning Control

This paper studies human-in-the-loop iterative learning control for output-trajectory tracking of linear time-invariant (LTI) systems. Note that knowledge of the intended output is necessary to ascertain the tracking error and find the corrective input needed to reduce the error at each iteration step. In human-in-the-loop systems, the controller does not have direct access to the intended output, and therefore, the tracking error needs to be inferred from human actions. A challenge, however, in inferring the intent and the tracking error is that the human-response dynamics affects the relation between the tracking error (known to the human) and the resulting human action. The main contribution of this paper is to infer the tracking error by inverting a model of the human response and using it for iteratively controlling general LTI systems. This is enabled by modifying the controlled-system dynamics perceived by the human to be of a form for which the human-response model is known. An advantage of the proposed approach is that output tracking is achieved in spite of (potentially) imperfect corrective actions of the novice human user. Results of human-in-the-loop tracking experiments with nine human subjects show that the proposed approach can estimate the human intent (to within 97% accuracy). Moreover, the iterative approach substantially reduces the tracking error (about 93%) when compared with the tracking error achieved by the human alone.