Neural Network-Based Calibration of 6D Force-Torque Sensors in Humanoid Robots

Precise force-torque sensing plays a fundamental role in humanoid robot control, impacting tasks from manipulation to movement. A conventional technique for cali- brating 6D force-torque sensors linear regression, which in some cases could not give good results, especially in the field of robotics. The goal of this thesis is to im- prove the accuracy and efficiency of 6D force-torque sensor calibration in humanoid robotics by introducing a novel neural network-based method. We employ machine learning methods, particularly neural networks, to simulate the cross-axis coupling effects and nonlinearities present in the sensors. As part of the process, data are collected during in-situ experiments, meaning that the robot has the sensor installed already. Such data are used to train a neural network to predict the real wrench (i.e. a 6D vector containing the components of force and torque) from the raw sensor readings and also other quantities. The performance of the proposed neural network calibration method is evaluated against traditional calibration techniques. Results indicate a significant improvement in sensor accuracy. This thesis contributes to the advancement of humanoid robotics by providing an efficient and accurate method for force-torque sensor calibration. The integration of neural networks not only simplifies the calibration process but also paves the way for more intelligent and autonomous robotic systems capable of complex interactions with their surroundings.