AbstractThis paper presents an optimal feedback controller capable of driving a non‐linear control system from an arbitrary initial state to a fixed final state in minimum time. The controller is based on a feedforward multilayer neural network trained repeatedly using open‐loop optimal control data which densely span the field of extremals of the non‐linear system. The effectiveness of the controller is clearly demonstrated by a simulation on a two‐link robot manipulator. The effect of sensor/actuator noise and parameter variation is also included to confirm the robustness of the con