Inferring human intent in novice human-in-the-loop control tasks

Abstract

Inferring intentions is fundamental to successful interaction among two or more agents. For example, in learning from a teacher the student must be able to first understand the teacher’s intent from the displayed demonstrations which is then committed to memory by imitation and self-practice. Analogously, robots may also be taught new skills using this Teaching by Demonstration (TbD) paradigm, in which case inferring the human teacher’s intent becomes necessary, especially when precision in learning new skills is important, such as in precision assembly tasks or surgery. The primary challenge in learning from a human-in-the-loop operator is that the actions do not always match the intended goal. This stems from the inherent dynamics of human response, which affects the overall closed loop tracking performance when the human is placed in the control loop. This work proposes an inversion approach to compensate for the human-in-the-loop dynamics to infer the underlying intent of the human-in-the-loop operator’s actions. As a result, the task representation using this inferred intent is potentially improved over that based on imperfect human demonstrations. The following are the main contributions of this work: (1) Stable inversion of human dynamics models to iteratively infer human-in-the-loop intent, with convergence guarantees robust to modeling uncertainty and output noise, (2) One-shot human intent prediction with guarantees on improved tracking accuracy compared to demonstrations, (3) Experimental validation of iterative and one-shot human-intent prediction and its application to human-guided robot learning with multiple human subjects.