Open-Ended Continuous Learning of Compound Goals

摘要

Continuous learning of increasingly difficult tasks is envisaged to provide learning autonomy to robots. The tasks, however, are often human-designed or require significant external intervention. This article proposes a domain-independent goal generation mechanism to generate goals at different levels of complexity. Using a mobile robot application, we demonstrate how an agent generates compound goals by combining the state space attributes from the states that it has experienced during exploration and uses task-independent reward functions to learn the solutions to those goals. Finally, the whole process is repeated when its environment changes, thus, forming a continuous learning architecture. Results from the experiments show how the agent can combine complementary and contradictory groups of state attributes to form expressive goals and learn behaviors akin to wall following, avoiding obstacle and lane following without any previous knowledge of its environment.