Learning Free Terminal Time Optimal Closed-loop Control of Manipulators