Minimum time motion of manipulators along specified path amounts to design the motion in a way that end-effector moves along the specified trajectory as fast as possible. It is shown that such a motion is bang-bang in terms of the tangential acceleration of end-effector along the path. Recently a learning algorithm was proposed to find switching points and near-minimum time solution online. Basic idea behind the method is to move the manipulator on the specified path on consequent segments of maximum acceleration, constant velocity, and maximum deceleration and to teach the manipulator to reduce and adjust the constant velocity period in each step of learning process. As the constant velocity period gets smaller and smaller, the solution converges the time optimal and two switches on the start and final time of constant velocity period converge the real switch. Adjustment of second switch also pushes the final error to zero. In this algorithm, it is assumed that the phase plane trajectory of solution is symmetric. Moreover, it was assumed that the maximum acceleration does not change drastically with the speed of manipulator. Despite novelty and ability to be implemented on line, this algorithm results inaccurate solutions for medial stages of learning for non symmetric and non smooth trajectories along which manipulator shows different ability in acceleration and deceleration. This thesis presents further investigations on non symmetric and non smooth trajectories, and introduces some modifications in leaning rules to account for this defect. Two new learning rules are introduced and their performances in different situations are compared through numerical simulations. Moreover, all previous studies only present minimum time path planners instead of controller, and don’t give guarantee to keep the end-effector on the desired path in the case of unwanted deviation from the path. Since there is no desired time history for joint variables, desired path is not available in terms of time. Therefore the first step to design a closed loop controller is to define suitable concepts of position and velocity errors. Defining these concepts, this thesis, proposes a closed loop controller based on the aforementioned learning algorithm. This controller is able to keep the end-effector on the predefined trajectory in the presence of external disturbances or any unwanted deviation from the path. With this controller we can be sure that after a limited number of learning stages, the motion converges to the theoretical minimum time solution in a safe and secure way. To cut the learning process, this paper presents an index for plannar manipulator which shows how close it is to leaving the desired trajectory. Key Words: Minimum Time Control, Trajectory Planning, Serial Manipulators, Specified Path