Intemittent control with reinforcement leaning

In this study, a hierarchical architecture for the intermittent control under the minimum transition hypothesis (MTH) was implemented. A two-stage hierarchy was adopted to perform the high-level and the low-level control respectively. The high-level controller performed the intermittent control by setting a sequence of goals for the low-level controller. Goal planning as the intermittent control policy was learned with hierarchical deep deterministic policy gradient (h-DDPG) proposed in this study, which is a hierarchical version of the conventional DDPG. The model successfully learned to temporally decompose a complex movement into a sequence of basic motor skills with sparse transitions, as shown in results of the two validation experiments: the trajectory following and the obstacle avoidance tasks.