The perception and exploration problems are two of the fundamental problems of mobile robots. Perception is a preliminary step and solves recognition problems such as object recognition and scene recognition while the exploration problem has more to do with decision making and control. During the past few years, deep learning has made many breakthroughs in a lot of research fields, like natural language processing and computer vision. Deep neural networks take a hierarchical structure that imitate human nervous systems for information processing and avoid the need to calculate hand-crafted features, which bring about many possible solutions to robotic problems.

This work introduces an integrated framework for both perception and exploration of mobile robots, with the use of...[ Read more ]

The perception and exploration problems are two of the fundamental problems of mobile robots. Perception is a preliminary step and solves recognition problems such as object recognition and scene recognition while the exploration problem has more to do with decision making and control. During the past few years, deep learning has made many breakthroughs in a lot of research fields, like natural language processing and computer vision. Deep neural networks take a hierarchical structure that imitate human nervous systems for information processing and avoid the need to calculate hand-crafted features, which bring about many possible solutions to robotic problems.

This work introduces an integrated framework for both perception and exploration of mobile robots, with the use of deep neural networks, and focuses on real-time performance, which is essential for robotic applications. A deep insight of the hidden structure of deep neural networks as well as statistical analysis are made. Then a principal component analysis (PCA)-based algorithm is proposed for efficient execution of deep neural networks. The structure is tested for image classification problems, considering both pixel-wise classification and patch-wise classification. Furthermore, an indoor obstacle avoidance approach based on deep neural networks is proposed and the algorithm is tested in both simulated and real world environment, considering noise observation. We further check the feasibility of indoor exploration by introducing and analysing the information gain when dealing with the obstacle avoidance task.