SVM-Based State Estimation of Biped Robot

Currently, the research of biped robots is becoming more and more popular, the kinematics-based state estimation, as an important precondition for its dynamic walking, has also become the focus of many researchers. Owing to the hybrid dynamics of biped robot, ground contact of each foot has to be previously evaluated to complete state estimation when using kinematics method. The contribution of this paper is to introduce the soft SVM to detect the contact phase. This SVM-based method free biped from the use of force sensors, which are highly vulnerable under the impact between feet and contact surfaces. Moreover, accurate dynamic modelling is not requisited. These two advantages indicate the strong robustness of SVM -based contact detection. Based on this contact detection method, the paper studies the state fusion method by applying extended Kalman filter to combine kinematics estimation with IMU data. Finally, the SVM-based contact detection algorithm and the complete state fusion method are both verified on our biped robot, with several experiments. The accuracy of the SVM -based method is validated by the comparison with force sensor based method. In addition, in contrast to torque-based method, its accuracy does not highly rely on the selection of algorithm parameters, like torque threshold.