Real-time deep learning-based position control of a mobile robot

This study uses PID (Proportional-Integral-Derivative), fuzzy logic, and deep learning algorithm to experimentally achieve real-time position control of a four-wheel-drive symmetric autonomous mobile robot whose design and prototype are realized. At the same time, the convolutional neural network-based YOLO (You only look once) algorithm is used to detect and get around obstacles by classifying the items in front of it during the robot's movement, to get the robot to reach the position specified as a reference a fuzzy logic controller is created. As the robot can be used outside, It is necessary to recognize more than one type of obstacle. For this reason, YOLO training is conducted to classify eighty objects that the robot may encounter in the external environment, such as cats, dogs, people, and chairs. In this research, a single obstacle is studied and it is determined as a human class obstacle, which is one of the obstacles that the robot is most likely to encounter in the laboratory environment. While the target location points are sent to the robot with the designed Android application, all the controls needed by the robot are carried out by the microcontrollers on it, independent of any computer. As a result of the experimental studies, it is seen that people are detected in real time with YOLO within the range of 90 cm-130 cm specified in the algorithm, without any problems. In addition, the mobile robot reached the target points sent to it from the Android application without any errors by overcoming obstacles within the specified approach distance (0.05 m or 0.1 m). While the maximum mean absolute error in the speed controls made on the motors throughout all experimental studies is 0.351 rpm, the Robot moved with a maximum absolute linear speed error of 2.8 mm/s. This shows that the robot is successfully controlled with fuzzy and PID controllers in line with the information obtained with YOLO.