Date of Publication :30th June 2023

Abstract: Construction robots are receiving more and more attention as a promising solution to the emerging shortcomings of the conventional construction industry. The development of intelligent control techniques for obstacle avoidance is crucial for guaranteeing the adaptability and flexibility of mobile construction robots in complex construction environments. Most of the existing obstacle avoidance algorithms are based on processing high-precision point cloud data collected by laser sensors to ensure operation fluency. However, because of the limitations of the laser sensors, those algorithms are invalid when detecting transparent obstacles that frequently appear in building environments. Therefore, this study aims to introduce a vision-based process for mobile construction robots to avoid transparent obstacles. To do so, a monocular camera is mounted on the testing robot platform, Turtlebot3 Burger, to collect visual inputs. A convolutional neural network is trained to compute the received videos and recognize transparent obstacles. The vision programs are coded in the Robot Operating System (ROS) to control the robot’s motions. On-site validations are conducted to prove the efficiency of the vision-based obstacle avoidance process. Different from avoiding obstacles using lidar inputs, the vision-based strategy successfully controlled the robot to avoid transparent obstacles. The findings contribute to paving a novel method for robotic obstacle avoidance by combining visual signals and deep learning, which is more efficient for avoiding collisions with transparent obstacles.

Reference :