Semantically Grounded Learning from Unstructured Demonstrations
Speaker: Scott Niekum, University of Massachusetts Amherst
Date: Friday, March 29 2013
Time: 3:00PM to 4:00PM
Location: Star, 32-D463
Host: George Konidaris, Learning & Intelligent Systems
Contact: Teresa Cataldo, cataldo@csail.mit.edu
Robots exhibit flexible behavior largely in proportion to their degree
of semantic knowledge about the world. Such knowledge is often meticulously hand-coded for a narrow class of tasks, limiting the scope of possible robot competencies. For this reason, learning from demonstration (LfD) has become a popular alternative to traditional robot programming methods, aiming to provide a natural and intuitive mechanism for programming robots. Unfortunately, LfD often yields little semantic knowledge about the world, and thus lacks robust generalization capabilities. In this talk, I will present a novel method to generalize complex, multi-step tasks by learning semantically grounded skills from natural, unstructured demonstrations. A Bayesian nonparametric algorithm is used to segment demonstrations into motion categories that can be recognized across demonstrations and tasks, maximally leveraging available data. These motion categories are then further subdivided into semantically grounded states in a finite-state representation of the task, enabling intelligent, adaptive replay. Finally, performance is incrementally improved through interactive corrections that provide additional data where they are most needed. Together, this allows for intelligent discovery and sequencing of grounded skills to create flexible, adaptive behavior that can be improved through natural interactions with the robot.
See other events happening in March 2013
