Error Prediction of a Differential Drive Wheeled Robot with a Swivel Caster Wheel

Abstract

In this study, we propose a method to predict the heading error of a differential drive wheeled robot due to misalignment of a swivel caster wheel. A swivel caster wheel, which is often used for balancing a differential drive robot, produces undesired torque to the robot when it changes direction. An error prediction model will allow a controller to compensate for the behaviors of a caster wheel. Support vector regression (SVR) and artificial neural network (ANN) were selected as our prediction methods due to their capabilities to model nonlinear systems. 966 trials of experiments were conducted to obtain the training and testing data. We were able to predict the heading error with initial caster orientation and speed command as the input variables using SVR with radial basis function kernel and multilayer perceptron with ReLU and tanh as the activation functions. The RMSE of SVR was 2.13°, and the RMSE of ANN with ReLU and tanh were 2.12° and 2.11° respectively. Moreover, we also confirmed the robustness of our model by testing it with testing data at 800 mm/s, which is faster than the maximum speed of 600 mm/s in the training data. The RMSE at 800 mm/s were 8.83° for SVR, and 5.92° and 6.25° for ReLU and tanh, respectively.