Robots as Students: Towers of Hanoi

crossposted from willowgarage.com

Nate Koenig of the Interaction Lab at USC is continuing his work here at Willow Garage after a busy summer. Nate carried out an empirical study investigating the use of people as teachers for robots, while also researching learning by demonstration with PR2.

Participants in Nate's study used learning by demonstration to teach PR2 how to solve the Towers of Hanoi puzzle. As the name implies, learning by demonstration relies on a human teacher to provide a robot "student" with demonstrations of a complex task. In this case, the robot uses the state of the puzzle (i.e., location of red disk compared to blue and green), along with the teacher's command, to learn the demonstrated task. Volunteers used a web-based teaching tool to guide PR2 through the three-disk puzzle board. In one condition, teachers were able to directly see PR2, while in the other condition teachers viewed the robot's actions through a small video feed on the web tool. This manipulation allowed Nate to study if robot visibility affects teaching strategies and outcomes. Based on participants' commands and other observations from the environment, the robot learned how to solve the puzzle on its own.

Results from this study indicate that teachers perform better when visually separated from the robot. Performing "better" means that participants made fewer unnecessary or repetitive moves when teaching the robot. While this may seem counter-intuitive, the teachers who could see the robot were easily distracted from the task and seemed to build inaccurate mental models of PR2's capabilities.

In addition to this experiment, Nate worked on a number of smaller projects including developing the first video streaming ROS node and creating a web-based graphical interface for interacting with PR2. You can find many of these contributions in the hanoi package for ROS. Nate is also creator and lead developer for Gazebo, a popular open-source 3D robot simulator. Gazebo is heavily used at Willow Garage to simulate the actions of PR2, and Nate is providing us with numerous improvements.