Event Location: Newell Simon Hall 3002

Abstract: The use of robots in industry, domestic applications, and even space exploration is no longer the realm of science fiction. As robots become commonplace tools in diverse applications, though, human operators must be able to instruct the robots in how they are to complete the tasks assigned to them. Rather than rely on experts to program robots for each new task, we expect any person capable of demonstrating the skill to be able to instruct the robot to do so. In this thesis, we examine the use of demonstration to program or, more aptly, to teach a robot to perform a precise motion.

Programming by demonstration is an approach that has already seen some success in research applications. We propose a new method of learning from examples that is capable of producing provably collision-free motions without requiring special training for the teacher. This approach is capable of learning precise motions, even when the precision required is on the same order of magnitude as the noise in the demonstrations. In this work, the robot is an active participant in the learning process, requesting a limited amount of demonstration while maximizing the area in which the robot can confidently plan, thereby ensuring efficient use of the human’s valuable time.
We have developed a proof-of-concept system that is capable of learning motion policies in two-dimensional spaces. We have also conducted experiments in dimensionality reduction for higher-dimensional configuration spaces, showing that this approach can be applied to more complex systems as well. We propose to continue the development of this system and complete an active learning component to ensure the efficiency of its operation. We also propose to investigate the applicability of this approach in various domains and improve the effectiveness and efficiency of the component algorithms. We will evaluate this approach as a replacement for standard trajectory programming techniques currently used by experts in the field, and as a tool for inexperienced robot programmers to teach skills to robots.

A copy of the thesis proposal document is available at:
http://www.cs.cmu.edu/~nmelchio/files/melchior_proposal.pdf

Committee:Reid Simmons, Chair
Manuela Veloso
Jeff Schneider
O.C. Jenkins, Brown University