Hierarchical reinforcement learning for humanoids

Abstract

The control of humanoid robots has always been difficult as humanoids are multi-body systems with many degrees of freedom. With the advent of deep reinforcement learning techniques, such complex continuous control tasks can now be learned directly without the need for explicit hand tuning of controllers. But most of these approaches only focus on achieving a stable walking gait as teaching a higher order task to a humanoid is extremely hard. But there have been recent advances in Hierarchical Reinforcement learning, in which a complex task is broken down into a hierarchy of sub-tasks and then learned. In this paper, we demonstrate how a hierarchical learning inspired approach can be used to teach a higher order complex task, such as solving a maze, to a humanoid robot.