A Method for Neighborhood Gesture Learning Based on Resistance Distance

• Published in: Advances in Human Factors in Robots and Unmanned Systems pp. 427 - 436
• Main Authors: Yanik, Paul M., Threatt, Anthony L., Merino, Jessica, Manganelli, Joe, Brooks, Johnell O., Green, Keith E., Walker, Ian D.
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing 2017
• Series: Advances in Intelligent Systems and Computing
• Subjects: Gesture recognition; Human-Robot interaction; Machine learning; Resistance distance
• ISBN: 3319419587; 9783319419589
• ISSN: 2194-5357; 2194-5365
• DOI: 10.1007/978-3-319-41959-6_35

Multimodal forms of human-robot interaction (HRI) including non-verbal forms promise easily adopted and intuitive use models for assistive devices. The research described in this paper targets an assistive robotic appliance which learns a user’s gestures for activities performed in a healthcare or aging in place setting. The proposed approach uses the Growing Neural Gas (GNG) algorithm in combination with the Q-Learning paradigm of reinforcement learning to shape robotic motions over time. Neighborhoods of nodes in the GNG network are combined to collectively leverage past learning by the group. Connections between nodes are assigned weights based on frequency of use which can be viewed as measures of electrical resistance. In this way, the GNG network may be traversed based on distances computed in the same manner as resistance in an electrical circuit. It is shown that this distance metric provides faster convergence of the algorithm when compared to shortest path neighborhood learning.