Robot Tutors Offer Helpful Ideas

Just 11 inches high and looking like a cute yellow rubber snowman, Keepon sits on the desk next to the computer. “I have an idea that might help you,” Keepon says.

Keepon is a robot.

But this particular Keepon has been programmed by professor of computer science & mechanical engineering & materials science Brian Scassellati to tutor students as they solve puzzles known as nonograms. Like the similar-looking sudoku puzzles, nonograms can be skillfully solved using logical inferences, but unlike the popular sudokus, most American puzzlers don’t have any experience with nonograms.

Scassellati, working with doctoral student Daniel Leyzberg and undergraduate alum Sam Spaulding (now a graduate student in MIT’s Media Lab), has used this lack of nonogram experience to study how human learning can be improved with a robot tutor. The team’s results, recently presented at the 9th ACM/IEEE Conference on Human-Robot Interaction, show that even relatively simple personalizations in the robot’s approach can significantly improve student comprehension.

A key component of such tutoring was in the connection Keepon maintained with each student. Keepon had perfect knowledge of student’s actions through a software link to the puzzle, but the robot still turned toward the student’s computer screen to “watch” the puzzles get solved. Keepon would turn back toward the student when offering advice or when welcoming the student to the experiment, always excitedly bouncing up and down. If a lesson was repeated, Keepon would even apologize before delivering the strategy reminder the second time.

But while the connection was maintained with every student, the researchers made changes to how Keepon offered advice to some students completing the nonograms. While Keepon’s prerecorded strategy hints were always relevant to solving the puzzles, for a portion of the students the robot tailored which lessons it delivered, suggesting only strategies that those specific players weren’t fully utilizing. For other students, Keepon offered the same strategies, but delivered in random order.

Students receiving the personalized training solved the nonograms faster than those who received the lessons that hadn’t been tailored, a finding that aligns with previous studies on human-human tutoring and classroom instruction. However, the effectiveness of Keepon’s personalizations is notable because of how little customization was involved: the tips were the same, with only the selection of which tips to use being varied. Minimally customized robots can therefore make a significant difference in student learning.

And as Keepon says, that is an idea that might help you.