Identification of Ground Surface for Biped Robot Locomotion Using Foot-Signature Classifier

The generation of efficient joint trajectories for the locomotion of a Biped robot relies on various assumptions, such as whether the terrain is flat or grassy. If the robot encounters a different terrain that violates the terrain assumptions, the robot can fall. Therefore, it is necessary to identify the landscape during the step-touch moment. So, the preventive strategies can make required changes in trajectories to mitigate the chance of falls. This research article focuses on identifying the ground surface for the humanoid robot using the foot-signature matrix, which consists of raw data from various sensors. For this purpose, a universal activation function-enabled convolutional neural network is developed, and the foot signature is used as input. The training dataset comprises the raw data from simulation and real-time for the six different terrains. The presented classifier outperforms the other state-of-art methods for both real and simulation datasets. The result shows that the accuracy of the presented classifier is around 99.77 ± 0.2 % and 98.2725 ± 0.7 % for simulation and real-time respectively.