RoMPLA: An efficient robot motion and planning learning architecture

Robot motor skill learning is currently one of the most active research areas in robotics. Many learning techniques have been developed for relatively simple problems. However, very few of them have direct applicability in complex robotics systems without assuming prior knowledge about the task due to two facts. On one hand, they scale badly to continues and high dimensional problems. On the other hand, they require too many real learning episodes. In this sense, this paper provides a detailed description of an original approach capable of learning from scratch suboptimal solutions and of providing closed-loop motor control policies in the proximity of such solutions. The developed architecture manages the solution in two consecutive phases. The first phase provides an initial open-loop solution state-action trajectory by mixing kinodynamic planning with model learning. In the second phase, the initial state trajectory solution is first smoothed and then, a closed-loop controller with active learning capabilities is learned in its proximity. We will demonstrate the efficiency of this two phases approach in the Cart-Pole Swing-Up Task problem.