Design of neural network controller for a billiard robot

This paper focuses on the cue ball controlling problem for a billiard robot. A neural network (NN) controller is designed, and the trained robot is able to stroke the cue ball moving to the target point after colliding with objective ball and cushions. Since the problem is non-linear and non-smooth, the solution is divided into several steps. First, the stroking model and the coordinate definition are described. Second, the kinematic model for cue ball motion and the mirror model for cushion rebounds are established under the ideal smooth assumption. Then, the neural network method is used to modify the ideal models, and the pattern recognition method for trajectories is presented. In the verification test, the trained robot is able to master the cue ball controlling with each pattern. The statistic results tally with the model analysis. Compared with simply adopting neural network method, the method combined with theoretical kinematic analysis will effectively improve the network quality and reduce the training error.