Robot Task Learning With Motor Babbling Using Pseudo Rehearsal

Abstract

The paradigm of deep robot learning from demonstrations allows robots to solve complex manipulation tasks by capturing motor skills from given demonstrations; however, collecting demonstrations can be costly. As an alternative, robots can acquire embodiment and motor skills by randomly moving their bodies, which is referred to as motor babbling. Motor babbling data provide relatively inexpensive demonstrations and can be used to enhance the generalizability of robot motions, but they are often used for pre-training or joint training with target task demonstrations. This study focused on the concept of pseudo-rehearsal and retaining the embodiment information acquired from motor babbling data for effective task learning. Pseudo-rehearsal has beneficial features that allow robot models to be retrained and distributed without access to the motor babbling dataset. In this paper, we propose a pseudo-rehearsal framework that can be jointly trained with task trajectories and rehearsed motor babbling trajectories. Using our proposed method, robots can retain motor skills from motor babbling and exhibit improved performance in task execution.