A Bayesian Neural Network-based Obstacle Avoidance Algorithm for an Educational Autonomous Mobile Robot Platform

Abstract

Autonomous mobile robots belong to automatically controlled objects that are designed and produced for various demands. This study aimed to develop an inexpensive platform of autonomous mobile robots that can be used for educational and research purposes in technical universities. The robot was built based on popular ultrasonic sensors to detect obstacles and a Raspberry Pi 4, which is a Linux-embedded computer. An effective obstacle avoidance algorithm for the robot was developed using a Bayesian neural network for classification. Training a Bayesian neural network does not require a validation dataset separate from the available data. In addition, the Bayesian approach can effectively handle the uncertainty of the system and result in the best generation for the network when inferring the unseen data. Training data are generated using robot-to-obstacle distances and the corresponding navigation modes. The commands to control the left and right motors of the robot are generated by the pretrained Bayesian neural network for three modes of navigation, forward, left, and right. Finally, this system can be useful as it can be conveniently integrated with advanced robot control algorithms.

Keywords

• autonomous mobile robots
• obstacle avoidance
• Bayesian neural network