An experimental study on a learning-based approach for the push recovery of NAO humanoid robot

Push recovery and keeping balance in humanoid robots are two important issues which ensure that robot can perform the imitation procedure and can be readily integrated in any environment. This paper represents an approach for push detection in NAO-H25 which is based on robot's FSRs (force sensitive resistor) and learning algorithms. Two challenges are involved in this paper, namely, the low quality of NAO FSR and different levels of force which robot should detect. In this study, NAO's sensors data are gathered by Choregraphe. In the stand position (with no push) and by applying the maximum forces, FSR's output specified as base vectors in such a way that the differences between various vectors and base vectors indicate the type of the push (back, front ...). By getting different data set, the learning algorithms can be used in order to detect the type of the push based on the previous data. However, in this paper different approaches which require less computation than classical learning algorithms is used. Some specific data should be updated every time so that the robot can detect a new push better. After detection, robot should be in equilibrium, which is performed by the controlling part. Two actuators which are located in the robot's ankles are used to apply the required control signals. The experimental results indicate the correct detection in less computation volume.