Falling Prediction based on Machine Learning for Biped Robots

Biped robots are expected to be able to work in complex environments. However, these robots will inevitably fall at times and such falls may cause injury to the robot itself or to people nearby. Therefore, it is necessary to detect that the robot is falling to be able to warn the robot in sufficient time when it is about to fall and to switch its controller to protect the vulnerable parts of the robot. Modeling and analysis of the biped robot falling problem cannot be fully accurate and many current learning-based methods rely on large quantities of fall data that are difficult to use to train fragile robots. A machine learning-based fall detection method is therefore proposed in this paper. This method requires only a small amount of training data to obtain good fall detection, making the training process on the robot platform much safer. A support vector machine is used to determine the state of the robot and the decision boundary of the stable state is updated during motion to enable the classifier to match the motion capability of the robot.