Scaling Reinforcement Learning to Robotics by Exploiting the Subsumption Architecture