MIT EECS Draper Laboratory Undergraduate Research and Innovation Scholar
Leslie P. Kaelbling
Electrical Engineering and Computer Science
Teaching Robots To Understand Natural Language Tasks Programming robots to perform everyday tasks requires interacting with humans and their rich language structures for conveying information. Robots need to operate in environments which require a lot of noisy sensor input to build a world model from which the robot needs to draw conclusions. One important conclusion in human-robot interaction is determining the object a human is referring to. Evaluating lambda expressions with functions that define an object lets the robot determine the probability that it has found the object referred to. We need to find a way to translate natural language into lambda expressions so the robot can include expression evaluation in its plan including making observations that increase certainty of the world.
My project for SuperUROP integrates a lot of small questions I have spent last year answering. For instance how would a robot determine which object is heaviest? How does a robot create a model for what color an object is? I hope to get more mathematical background to find out interesting ways of thinking about we structure language think about objects make plans and how to apply that to robotics.