Automatic task manipulation with constrained image understanding for a tele-operated robot
If you have ever tried to tell someone who is blindfolded how to traverse an obstacle strewn path, you will understand the difficulty associated with giving instructions and having them followed appropriately. Similar difficulties are associated with controlling a tele-operated robot over the WWW. The feedback provided by the robot in the form of audio and video information is usually insufficient to carry out fine control tasks in a timely manner. Even something as simple as traversing through a doorway is often fraught with difficulties, including aligning with the door, opening it, navigating the threshold and finally passing through the doorway. The work presented in this thesis describes an automatic approach to doorway alignment, opening and passage for a low-bandwidth tele-operated wireless robot on the WWW. A systems approach is taken in utilizing doorways serviced by handicapped access mechanisms. Our work presents two major contributions for the traversal task. An automatic positioning subsystem has been devised to allow near optimal alignment of the robot in order to allow a video image of the doorway to be taken. A constrained image analysis technique based, in part, on a neural network subsystem is than employed to locate the handicapped access button within the image to determine the robots position relative to it. When the two subsystems interact, it becomes possible for the robot to align itself with the button and press it to gain access to the doorway. The interaction of these two systems represents an extendible architecture that supports human tele-operation by off-loading mundane yet complex tasks to automatic stable processes ensuring timely and effective task completion.