Formalizing Human-Robot Mutual Adaptation via a Bounded Memory Based Model

Abstract

We present a formalism for mutual adaptation between a human and a robotic assistant, which enables guiding a human teammate towards more efficient strategies, while maintaining trust to the robot. We propose BAM, a model of human adaptation based on a bounded memory assumption. We integrate BAM into robot decision making by including it as part of a mixed-observability Markov decision process (MOMDP) formulation, wherein the human adaptability is a partially observable variable. In a human subject experiment (n=69), participants were significantly more likely to adapt to the robot strategy when working with a robot utilizing the proposed formalism (p=0.036), compared to cross-training with the robot. Additionally, the performance as a teammate of the robot that executed the learned MOMDP policy was perceived to be not worse than that of the robot that cross-trained with participants. Finally, the robot was found to be more trustworthy with the learned policy, compared with executing an optimal strategy while ignoring the adaptability of the human teammate (p=0.048). These results indicate that the proposed formalism can significantly improve the effectiveness of human-robot teams, while achieving subjective ratings on robot performance and trust that are comparable to those obtained by state-of-the-art human-robot team training practices.