As the great setups for the implementation of human-competitive artificial intelligence algorithms, ping pong robots have found an increasing attention in the literature. One of the multiple factors that affects the performance of these robots, is the paddle trajectory, travelled to hit the ball. A good hitting task trajectory should be accurate and fast enough to ensure that the ping pong paddle will hit the ball at the correct position/orientation, in the correct time, and with the correct speed. Moreover, considering the limitations in the speed and torque of the robot’s actuators, the generated trajectory should be implementable by the robot. In this paper, three of most popular path generation curves, namely Spline, Bézier, and Dubins curves, are used to generate the hitting task trajectory for different hitting conditions. The generated trajectories are then compared based on a multi-objective cost function that includes the maximum values of the jerk, acceleration, and torque imposed to the robot’s actuators. Finally, an artificial neural network is used to learn the best trajectory for each hitting condition. The neural network takes the hitting condition (three-dimensional hitting position and velocity) as the input and selects the best trajectory at the output. The results show that the trained neural network selects the true trajectory with an 85% success rate for the test data. The proposed method enables the ping pong robots to instantly select the optimum trajectory for each hitting condition, with no need to timeconsuming optimization procedures.

Keywords