We propose an architecture for learning complex controllable behaviors by
having simple Policies Modulate Trajectory Generators (PMTG), a powerful
combination that can provide both memory and prior knowledge to the controller.
The result is a flexible architecture that is applicable to