Towards generating simulated walking motion using position based deep reinforcement learning

Abstract

Much of robotics research aims to develop control solutions that exploit the machine's dynamics in order to achieve an extraordinarily agile behaviour [1]. This, however, is limited by the use of traditional model-based control techniques such as model predictive control and quadratic programming. These solutions are often based on simplified mechanical models which result in mechanically constrained and inefficient behaviour, thereby limiting the agility of the robotic system in development [2]. Treating the control of robotic systems as a reinforcement learning (RL) problem enables the use of model-free algorithms that attempt to learn a policy which maximizes the expected future (discounted) reward without inferring the effects of an executed action on the environment.