Planning appropriate trajectories can significantly increase the productivity of robot systems. To plan realistic time-optimal trajectories, the robot dynamics have to be described precisely. In this paper, a neural network based algorithm for time-optimal trajectory planning is introduced. This method utilises neural networks for representing the inverse dynamics of the robot. As the proposed neural networks can be trained with data obtained from exciting the robot with given torque inputs, they will capture the complete dynamics of the robot system. Threfore, the trajectories generated will be more realistic than those obtained by using nominal dynamic equations based on nominal parameters. Time-optimal trajectories are generated for a PUMA robot to demonstrate the proposed method.