Mobile robots (such as the Pepper robot shown in the above picture, which will serve as practical robotic platform for this thesis) must navigate in sometimes complex indoor environments. This navigation must rely on some kind of map of the environment, allowing the robot to find out where it is at any given time and which paths can be followed to move to other locations. A popular way of building such a map is to use Simultaneous Localization and Mapping (SLAM) techniques. The goal of SLAM is essentially to let the robot roam around its environment and gradually build a map of its surroundings, using the sensors at its disposal (sonars, laser scanners, cameras etc.).
However, running a SLAM system in a completely autonomous mode creates a number of challenges. First, the robot may sometimes get stuck or fail to perceive some important spatial features such as door openings. Second, this process only creates a raw metric map, but does not associate linguistic concepts to areas of this map (such as the fact that a room is called the "coffee room" and that the one next to it is "the corridor"), which is crucial for human-robot interaction.
The goal of this thesis is to extend an existing localisation and mapping procedure (such as one based on landmarks) with relatively simple clarification dialogues with a human being following the robot. The idea is to have the robot ask simple questions to the human ("what is the name of this room?", "where should I go now?") and follow simple instructions ("turn to the left", etc.). The tricky part is to find out when to ask questions (we do not want the robot to overload the human with too many questions) based on its current map, including uncertainties related to aspects of this map.
Prerequisites:
Good programming skills in Python, and ideally some knowledge of Java/Kotlin. Interest for human-robot interaction and dialogue systems.