AUTONOMOUS PERCEPTION AND NAVIGATION IN UNKNOWN INDOOR ENVIRONMENTS

Standard off-the-shelf SLAM algorithms allow robots to build 2D maps of their environments and consequently enable them to navigate to (x, y) coordinates in those maps. However, this is a large step removed from a robot finding and going to a professor’s office or locating an elevator and taking it up one floor. The robot would have to robustly detect and localize doors and elevators in a hallway. Additionally, given directions to this hallway, the robot would have to accurately follow them in a previously unknown environment. In this thesis, we propose solutions to these two key challenges associated with finding a goal in an unknown indoor environment. We present a robust algorithm that relies on image and laser-range data to detect doors. This algorithm is combined with a set of common-sense rules to enable a robot to efficiently find a specific door in a hallway. To follow directions in an unknown environment, we propose a convolutional neural network-based approach that takes a local crop of the 2D SLAM map and a command as input to produce navigational goal points and feedback for the robot as output. All of these methods are deployed on a real robot and evaluated in the form of live trials in previously unseen and unmodified office environments.