Sam Spaulding

Smarter Robots for Engaging Interactions


Learning Social Behavior
From Demonstration

Using Learning from Demonstration (LfD) to create autonomous behavior policies for social interaction tasks. Collaboration with Brad Knox.

Learning from Demonstration (LfD) is a popular and effective method for developing robot behaviors. Thus far, however, LfD has been most effective in producing physical, task-oriented behaviors (e.g. manipulation). We are adapting the LfD paradigm for the development of robot social behaviors.

Humans interact with a robot in two phases: during the demonstration phase, the robot is tele-operated in a Wizard-of-Oz style. After sufficient demonstrations are collected, they are provided as input to algorithms that learn an autonomous policy based on the demonstrations. In the autonomous phase, humans interact with the robot that behaves according to the learned policy.

We are currently investigating whether the learned policy can accurately capture the interaction dynamics of a social interactions and how humans interact with a robot in a collaborative, social task in both Wizard-of-Oz and autonomous conditions

Publications

Learning Social Interaction from the Wizard: A Proposal
W. Bradley Knox, Sam Spaulding, and Cynthia Breazeal. 3rd Workshop on Machine Learning for Interactive Systems (MLIS '14) at AAAI 2014



Effects of Personalized Human-Robot Interaction in Educational Tutoring

How can Social Robots most effectively tutor students to help learn problem-solving skills? Collaboration with Dan Leyzberg.


We investigated the efficacy of three different methods of tutoring instruction: verbal instruction, video agent, and embodied robot and found that, even when delivering identical content, students who received instruction from an embodied robot learned a puzzle task significantly more quickly than students in either of the other two cases.

In addition, we developed a method for tracking a subject's proficiency at a complicated cognitive task composed of many unique skills via indirect observation. Rather than prompting a student to demonstrate his/her proficiency at each skill for evaluation, the tutor generalizes probabilistic estimates of proficiency from observed gameplay, without interrupting the learning task.

Finally, we isolated the effect of personalized tutoring in an HRI study and found that even simple personalizations could produce significant learning gains during a 1hr. experiment

Publications

The Physical Presence of a Robot Tutor Increases Cognitive Learning Gains
Leyzberg, D., Spaulding, S., Toneva, M. and Scassellati, B. 34th Annual Meeting of the Cognitive Science Society (Cog Sci ’12), July 2012.

Personalizing Robot Tutors to Individuals' Learning Differences
Leyzberg, D., Spaulding, S., and Scassellati, B. 9th ACM/IEEE Conference on Human-Robot Interaction (HRI ’14), March 2014.