Control Proprioception for Robust Autonomous Systems

This paper show how neural networks, configured for regression, can be used to learn the relationships between Inertial Motion Unit (IMU) data collected on a robotic platform and the robot's commanded system state. By learning how the IMU data relates to commanded robot state we can use the neural network to predict what commands have been issued to the robot. By comparing the prediction to the actual commands we can determine if the perceived behavior of our robot matches the commanded behavior. This enables the vehicle to identify issues with control and potentially take corrective actions needed to enable long-duration autonomy.