Development of the control architecture of the mobile robotic platform ARGO for undercarriage train inspection

With a forecasted enlargement of the global railway infrastructure, ensuring more efficient and safer maintenance procedures are some of the main challenges that railway companies have to overcome. In this context, one of the most relevant and time-consuming maintenance procedures is the visual inspection of the underbody of the rolling stocks. It requires special maintenance facilities and it lies in the framework of reactive maintenance. A shift towards a predictive maintenance approach would allow for an increase in both efficiency and safety. This newer approach relies on the collection of sensory data, which is absent in the traditional visual inspection procedure. These are some of the reasons why Next Generation Robotics s.r.l. has developed the ARGO robotic platform for performing visual inspection of the trains’ undercarriage. In this context, the work carried out in this master’s thesis has focused on the development of the control architecture of the teleoperated ARGO robot. The high-level control structure has been developed in the ROS Noetic framework. It handles the tasks of both converting the inputs provided by the human operator into viable setpoints for the actuation of the DC motors powering the robot’s joints and providing an estimate of the robot’s odometry. During the development of this thesis, particular attention has been paid to the task of accurately estimating the robot’s odometry. Computing a precise estimate of the robot’s position and speed is necessary to provide the human operator with feedback on the robot’s state in scenarios where the robot may not be visible. It also allows attachment of metadata to the collected sensory data for later analysis. For the purpose of estimating the odometry, a sensor fusion algorithm based on a constant acceleration Kalman filter has been developed and tested, obtaining a satisfactory estimate of the robot’s state. The filter fuses together data coming from the optical encoders mounted on the transmission wheels together with the estimate of the robot’s speed returned by a motion tracking algorithm running on the point clouds returned by two 2D LiDAR sensors mounted on the robot. The algorithm estimates the robot speed by tracking the motion of reference objects identified in the point clouds. For this purpose, two motion tracking algorithms have been developed and tested. First, a simpler one based only on geometric considerations. Then, a second one based on Kalman filtering in order to improve the results obtained with the first one. Experimental tests have shown that the algorithm based on Kalman filtering returns a better estimate of the robot’s speed and acceleration in terms of root mean square error because of its ability to better reject the measurement noise. Satisfactory odometry estimation allows closing the high-level control loop for teleoperation by a human operator. Finally, the low-level control structure has been developed with the goal of actuating the DC motors powering the robot’s joints to follow the position and speed reference signals computed by the high-level logic. The controller has been designed to implement a digital cascaded PI controller with a position, a speed and a current control loop. It has been developed and tested using a STM32 Nucleo board based on an Arm® Cortex M7 micro-controller unit and the drv-8245 IC h-bridge driver by Texas Instruments®. The low-level controller has been tested experimentally, achieving satisfactory results for the robot’s actuation.