Exploring Deep Learning Based Robot Perception Techniques for Navigating Outdoor Terrains

When autonomously navigating to its objectives, a ground robot encounters formidable challenges detecting and recognizing its surroundings and objects. From its sensory input, the robots AI has to semantically segment the scenes such as terrain, vegetation, human-made structures, debris, water streams, etc. The on-board perception then has to assess intelligently and determine what parts of the scene the robot can traverse safely to the objective. The project goal is to develop a novel method of vision-based perception for assessing the navigability of terrains that an autonomous ground vehicle may encounter traversing in natural or structured environments. With the great success brought by the advances in deep learning, computer vision sometimes exceeds human-level performance in object recognition tasks. These algorithms, however, require a large size of class examples to perform accurately. Although visual data are abundant, images relevant to ground navigation, especially those with class labels are scarce. Therefore, there is a need for a computer vision algorithm capable of high performance with small training sets and capable of recognizing novel objects. We propose to investigate the GAN-based data augmentation approach and the efficient scene understanding approach to tackle the data scarcity issue of the perception problem related to autonomous robotic maneuvers in previously unseen environments. Expected relevant robot maneuvering environments and scenes are typically unusual, and the data for the current paradigms of deep learning is either scarce or non-existent. Hence, it is expected that the GAN-based data augmentation approaches provide solutions to developing terrestrial robotic vehicles capable of perceiving and understanding novel environments.