Simulation for Ball Landing Control of a Robotic Ping-Pong System

Abstract

For a robotic ping-pong system, it is difficult to return the ball with different spins to a desired landing point. Previous work on this problem mainly focused on the physical models. To simplify the derivation of the model, this paper proposes a racket control method using recurrent neural networks (RNN) for a robotic ping-pong system to return the spinning ball to a desired position. The proposed method consists of following three parts: 1) the framework of using a RNN to predict the ping-pong ball trajectory after it is hit by a racket; 2) a feedforward neural network to estimate ball outgoing velocity based on the prediction trajectory; 3) a feedforward neural network to estimate the racket pose based on the ball velocity change after hitting by a racket. Compared to other regression algorithms, neural networks (NNs) do not require assumptions about the relationship between inputs and outputs of a ping-pong robotic system and can learn its complicated dynamic model. Simulation results show that the proposed method improved the estimation error of the ball outgoing velocity with topspin and backspin, and reduced the error of the landing position. Compared with the modified locally weighted regression (LWR), the percentage of x- and y-direction error was improved by 10% and 25%, respectively.