DEVELOPMENT OF AN AUTONOMOUS WHEELED MOBILE ROBOT THAT INTEGRATES THE YOLOV11 OBJECT DETECTION MODEL

Many mobile robots are unable to successfully navigate complex unstructured environments because they cannot identify and avoid obstacles. This article presents the design and implementation of an autonomous mobile wheeled robot which combines the YOLOv11 object detection system with a Proportional Integral Derivative (PID) controlled differential drive robot. The new approach uses a single monocular RGB camera and infrared sensors to identify obstacles, followed by path re-planning with a probabilistic roadmap method, instead of the traditional approach using distance-only sensors or an object classifier with a limited training set. The YOLOv11 model was trained using the "Object Detection Dataset" from Kaggle that comprises 7,064 images of 119 industrial and domestic items. Following 220 epochs of training, the model's mean Average Precision (mAP@0.5) and recall were 0.89 and 0.87, respectively. During 20 indoor tests with random obstacle placements, the robot achieved a collision free success rate of 95% in comparison to artificial potential field (75%) and ant colony optimisation (80%) methods. It took an average of 0.27 seconds for the robot to react to an obstacle and avoid it. The system completely solved the problem of "Goals Non-Reachable with Obstacles Nearby" (GNRON). This work shows that deep learning for obstacle detection can greatly improve semantic and reliable robotic navigation. Future research will include the use of other sensors and an edge computing platform.