Research on the Switch Autonomous Operationof an Inspection Robot Based on Button PoseRecognition

Background: Accurate estimation of button-switch poses, as critical operational components for station inspection robots, plays a significant role in enabling the robot's autonomous operation. Objective: To achieve the autonomous operation of the robot on button-switches, and provide a new patent for the button detection field, an algorithm based on stereo cameras for estimating the poses of buttons was proposed, and experiments were conducted on the UR robot. Methods: Firstly, a dataset of commonly used button-switches in the factory was collected by capturing images with a camera. Subsequently, a neural network was trained to detect the captured RGB images. Secondly, depth image information was fused with RGB data to generate three-dimensional point clouds of the target regions. These point clouds were subsequently processed through clustering and segmentation to isolate the point clouds corresponding to the detected button-switches. Finally, the pose information of the manipulation point of button-switch was calculated based on the segmented point clouds. The effectiveness of the proposed algorithm was validated through operational experiments conducted on a UR robot. Results: The experimental results show that the detection model achieves a mean Average Precision (mAP) of 93.64% and a recall rate of 88.69%. The success rate for pose estimation of manipulation points is 92.31%, with an average computation time of approximately 86.63 ms per point. Discussion: The detection accuracy of the model is highly dependent on the quality and diversity of the dataset. An insufficiently diverse dataset may limit the model's generalizability to a broader range of button-switch types. Conclusions: The results demonstrate that the proposed method is applicable for the autonomous operation of button-switches by station inspection robots.