Grounding Natural Language Instructions to Robot Behavior
For general-purpose robots to become widespread, users need to be able to instruct robots to perform a wide variety of tasks. Currently, this requires knowledge of robot programming (and debugging), which is too much to expect from most users. Instead, we envision robots that can interact with users naturally, similar to the ways humans interact with each other. In particular, we focus on using natural language to instruct robots. I will describe a line of work that interprets natural language navigation instructions and converts them into robot goals and actions. On a technical level, we frame this as a machine translation problem, and we take advantage of recent advances in neural machine translation to accurately convert English instructions into a semantic goal representation. The design of these goal representations highlights a trade-off between the expressiveness of achievable tasks and the accuracy/efficiency at which we are able to interpret/accomplish the given tasks; we will explore several points on this spectrum in this talk.
Lawson L.S. Wong is a senior research associate at Brown University, working with Stefanie Tellex and George Konidaris. He will be joining Northeastern University as an assistant professor in Fall 2018. His research focuses on learning, representing, and estimating knowledge about the world that an autonomous robot may find useful. More broadly, Lawson is interested in, and follows many topics within, the fields of robotics, machine learning, and artificial intelligence. He completed his Ph.D. in 2016 at the Massachusetts Institute of Technology, advised by Leslie Pack Kaelbling and Tomás Lozano-Pérez. Previously, he received his B.S. (with Honors) and M.S. in Computer Science at Stanford University. He has received a Siebel Fellowship, AAAI Robotics Student Fellowship, and Croucher Foundation Fellowship for Postdoctoral Research.