Efficient Machine Learning of Mobile Robotic Systems based on Convolutional Neural Networks

Abstract

During the last decade, Convolutional Neural Networks (CNNs) have been recognized as one of the most promising machine learning methods that are being utilized for deep learning of autonomous robotic systems. Faced with everlasting uncertainties while working in unstructured and dynamical real-world environments, robotic systems need to be able to recognize different environmental scenarios and make adequate decisions based on machine learning of the current environment’s state representation. One of the main challenges in the development of machine learning models based on CNN is in the selection of appropriate model structure and parameters that can achieve adequate accuracy of environment representation. In order to address this challenge, the book chapter provides a comprehensive analysis of the accuracy and efficiency of CNN models for autonomous robotic applications. Particularly, different CNN models (i.e., structures and parameters) are trained, validated, and tested on real-world image data gathered by a mobile robot’s stereo vision system. The best performing CNN models based on two criteria – the number of frames per second and mean intersection over union are implemented on the real-world wheeled mobile robot RAICO (Robot with Artificial Intelligence based COgnition), which is developed in the Laboratory for robotics and artificial intelligence (ROBOTICS & AI) and tested for obstacle avoidance tasks. The achieved experimental results show that the proposed machine learning strategy based on CNN provides high accuracy of mobile robot’s current environment state estimation. This book addresses many applications of artificial intelligence in robotics, namely AI using visual and motional input. Robotic technology has made significant contributions to daily living, industrial uses, and medicinal applications. Machine learning, in particular, is critical for intelligent robots or unmanned/autonomous systems such as UAVs, UGVs, UUVs, cooperative robots, and so on. Humans are distinguished from animals by capacities such as receiving visual information, adjusting to uncertain circumstances, and making decisions to take action in a complex system. Significant progress has been made in robotics toward human-like intelligence; yet, there are still numerous unresolved issues. Deep learning, reinforcement learning, real-time learning, swarm intelligence, and other developing approaches such as tiny-ML have been developed in recent decades and used in robotics. Artificial intelligence is being integrated into robots in order to develop advanced robotics capable of performing multiple tasks and learning new things with a better perception of the environment, allowing robots to perform critical tasks with human-like vision to detect or recognize various objects. Intelligent robots have been successfully constructed using machine learning and deep learning AI technology. Robotics performance is improving as higher quality, and more precise machine learning processes are used to train computer vision models to recognize different things and carry out operations correctly with the desired outcome.