Deep Learning-Based Calibration Method of a 6-Axis Force-Torque Sensor for a Wheeled Humanoid Robot

This work proposes a calibration methodology for a force-torque sensor of 6 degrees of freedom by using deep learning. This force-torque sensor will be used to estimate external forces applied to the mobile platform and/or torso of a wheeled humanoid robot, as well as the contact location. The results proved that the neural network is capable to deal with the non-linearities produced by the mechanical cross-talk of the force-torque sensor, and therefore, it will allow a better estimation of interaction forces and the contact location in the wheeled humanoid robot. The experimental results also show that the model avoids the over-fitting problem because the gap between the training and validation curves is practically zero. Moreover, the inference time of the neural network model was optimized in order to achieve a real time operation, and this will allow the implementation of a force-feedback control algorithm.