Developing Robot Autonomy via Invariant Representations

点击翻译按钮获取中文摘要

People pay revisits to places encountered in the past seemingly oblivious of the time of the day, the weather conditions or even the season when the initial visits were made. They perform this task effortlessly mostly through the use of their sense of vision to recognize routes and landmarks and make navigational decisions. Can robots do the same? How does a robot process its visual sensory data and build a representation of the environment in such a way that enables it to recognize places that have been encountered in the past independently of the environmental conditions? These questions still challenge the robotics community today, and obstacles to robot autonomy. They are at the heart of this research proposal. This proposal requests funding to support research in visual navigation of autonomous robots, considered a fundamental requirement for a mobile robot to be truly useful, in order to work competently in homes and hospitals, at work sites as well as in areas hard hit by natural disasters such as earthquakes. Much of the research in autonomous navigation has adopted vision as the primary sensor for the robot to build a map of the environment and localize itself within that environment. This proposal tackles a key challenge in visual robot navigation, namely, how to describe an environment from the visual data in such a way that allows a robot to overcome the appearance changes of the environment that still hamper vision. These changes include changes in lighting due to time of the day, weather, season, and changes in the viewpoint of the robot visual sensor during revisits. The changes can be especially significant when robots are deployed on missions over a long period of time. The long-term goal of the proposed research is to achieve autonomous robot visual navigation. The short-term goals of our research focus on solutions to build an environment representation that is invariant to various changes in the environment. Our effort in achieving this invariant representation unfolds on two fronts. First, we recognize that the geometry of an environment does not change with lighting. Therefore, if we characterize an environment in terms of its geometric features, we can effectively achieve an invariant environment representation. Second, we are encouraged by the recent impressive development of deep neural networks in visual object detection and recognition tasks, and we will proceed to use state-of-the-art techniques in deep learning and create a semantic description of the environment that is invariant to its changes. Specifically, we depend on deep learning to detect and recognize constituent objects of a scene, and we match scenes using such a semantic description while taking into account the spatial arrangement of the objects in the scene. Our research will be conducted on both existing benchmark datasets for autonomous navigation research as well as on real physical robots, on land and on water.