This method actually does have fixed-purpose computational units if you will (that's why it's called guided policy search). During learning, it basically loops

This method actually does have fixed-purpose computational units if you will (that's why it's called guided policy search). During learning, it basically loops back and forth between a heuristic trajectory optimization (to keep a tight rein on the motor torques) and reinforcement learning with neural networks (to achieve high-level abstraction).