Personal Greetings: Personalizing Robot Utterances Based on Novelty of Observed Behavior

Abstract

One challenge in creating conversational service robots is how to reproduce the kind of individual recognition and attention that a human can provide. We believe that interactions can be made to seem more warm and humanlike by using sensors to observe a person’s behavior or appearance over time, and programming the robot to comment when it observes a novel feature, such as a new hairstyle, or a consistent behavior, such as visiting every afternoon. To create a system capable of recognizing such novelty and typicality, we collected one month of training data from customers in a shopping mall and recorded features of people’s visits, such as time of day and group size. We then trained SVM classifiers to identify each feature as novel, typical, or neither, based on the inputs of a human coder, and we trained an additional classifier to choose an appropriate topic for a personalized greeting. An utterance generator was developed to generate text for the robot to speak, based on the selected topic and sensor data. A cross-validation analysis showed that the trained classifiers could accurately reproduce human novelty judgments with 88% accuracy and topic selection with 95% accuracy. We then deployed a teleoperated robot using this system to greet customers in a shopping mall for three weeks, and we present example interactions and results from interviews showing that customers appreciated the robot’s personalized greetings and felt a sense of familiarity with the robot.