Autonomous Skill Acquisition for Robots Using Graduated Learning

Gautham Vasan

Abstract: Skill acquisition is among the most remarkable aspects of human intelligence. It involves discovering purposeful behavioural modules, retaining them as skills, honing them through practice, and applying them in unforeseen circumstances [11]. Skill acquisition underlies our ability to choose to spend time and energy on the mastery of particular tasks and draw upon previous experience to solve more complex problems over time with less cognitive effort[10]. If endowed with continual skill acquisition, robots can autonomously improve their skills over time, where learning at one stage of development is a foundation for future learning [23]. It could unlock new possibilities for physical automation with general-purpose robots, just as general-purpose computer processors ushered in the information age [24, 33]. In this work, we propose a novel approach called Graduated Learning, where we ask a robot to acquire new manipulation and locomotion skills repeatedly, using time-delineated experiences of attempts at those skills (i.e., episodes) and some store of previously acquired knowledge (e.g., weights of a neural network). Our proposed approach chooses the order in which an agent learns these skills since the progressive manner in which they are developed plays a vital role in developing a final skill set.