Grounding robot motion in natural language and visual perception
The current state of the art in military and first responder ground robots involves heavy physical and cognitive burdens on the human operator while taking little to no advantage of the potential autonomy of robotic technology. The robots currently in use are rugged remote-controlled vehicles. Their interaction modalities, usually utilizing a game controller connected to a computer, require a dedicated operator who has limited capacity for other tasks. I present research which aims to ease these burdens by incorporating multiple modes of robotic sensing into a system which allows humans to interact with robots through a natural-language interface. I conduct this research on a custom-built six-wheeled mobile robot. First I present a unified framework which supports grounding natural-language semantics in robotic driving. This framework supports learning the meanings of nouns and prepositions from sentential descriptions of paths driven by the robot, as well as using such meanings to both generate a sentential description of a path and perform automated driving of a path specified in natural language. One limitation of this framework is that it requires as input the locations of the (initially nameless) objects in the floor plan. Next I present a method to automatically detect, localize, and label objects in the robot’s environment using only the robot’s video feed and corresponding odometry. This method produces a map of the robot’s environment in which objects are differentiated by abstract class labels. Finally, I present work that unifies the previous two approaches. This method detects, localizes, and labels objects, as the previous method does. However, this new method integrates natural-language descriptions to learn actual object names, rather than abstract labels.
Siskind, Purdue University.
Computer Engineering|Robotics|Artificial intelligence
Off-Campus Purdue Users:
To access this dissertation, please log in to our