A Practical Vision-Aided Multi-Robot Autonomous Navigation using Convolutional Neural Network

In this paper, a low-cost practical approach for the collision avoidance of a multi-robot system by using a single camera. A convolutional neural network (CNN) is applied to obtain an estimation of the depth of the image at the output of a monocular camera, assisting the team of mobile robots to detect obstacles in an unknown environment, determining navigation strategies, overcoming the limitation of the onboard LiDAR sensor. An avoidance controller was designed over a modified artificial potential field (APF) method, leading robots to avoid obstacles to reach the goal point. This paper provides an alternative solution for range measuring and environment sensing, replacing common distance sensors such as LiDAR sensors and ultrasonic sensors. The camera captures more data about the environment while being relatively cheaper than most sensors. An open-source CNN machine learning model called MiDaS is applied to help estimate the depth of detected obstacles from the input image. Simulations and experiments with three TurtleBot3 mobile robots were conducted to validate the proposed algorithms. Experimental studies have been carried out to test the effectiveness of the proposed approach in the paper.