Communicative behavior generation for navigational robots
Shih-Yun Lo, 0000-0001-9926-6236 · 2021
Communicative behaviors emerge as humans navigate a populated space and interact with one another. For robots to smoothly deploy in human environments, they too should be able to communicate. Inspired by the theory of mind in psychology, this thesis creates computational models that enable a robot to act based on inferring not only what humans will do next, but also how humans will interpret the robot’s own actions, using a nested-prediction formulation. Through a performance optimization process, the models automate the decision of whether, when, and how the robot should communicate. The generated communicative behaviors can shorten time delay by a third compared to prior work. Moreover, when the robot only engages in navigational, task-oriented actions, it generates motions that implicitly communicate its intent. Compared to the legible motions in the literature, ours have higher ratings by humans on intent clarity and other subjective measures. This thesis demonstrates performance optimization as a means to improve robot communicative behaviors in the navigation domain, in terms of both efficiency and human interpretability.