Research Article

Dynamical Motor Control Learned with Deep Deterministic Policy Gradient

Figure 1

Schematic illustration of the dynamical control. (a) Conventional motor control takes the state feedback as input to generate the control signal , and it behaves like a regulator or spatial filter to the feedback state. (b) The dynamical controller generates the control signal by its internal dynamics. Note that the dynamical controller loops by itself and theoretically the initial state and the goal state are sufficient to generate the control command, with or without the feedback state (dotted arrow). (c) The dynamical controller is trained using DDPG with the reward information from the environment (shown as the Env box). The broken arrow indicates that the controller parameters are tuned with the gradients from DDPG.