EAPSI:Evaluating Anticipatory Communication Strategies for Human-Robot Teaming

Butchibabu Abhizna, Edison NJ

Investigators

Abstract

Robotic systems are being integrated into complex and safety critical domains, where these systems work with humans as teammates. For example, in domains such as emergency response systems, robotic systems are deployed to conduct tasks that may not be feasible by humans. One of the challenges for fluent human-robot teaming is effective communication, where the human and the robot exchange of the right information at the right time is critical. Prior work shows that high-performing human teams tend to share information by anticipating the needs of their teammates (referred as implicit coordination) and that to work effectively with a human teammate, the robot must be interpredictable and communicate in a way expected by the human teammate. The objective of the project is to develop a computational model for the robot where the robot will proactively communicate with human teammates using implicit coordination strategies. This is one of the first studies to address how anticipated communication strategies could be used by robots to improve team performance when working with humans and how these strategies could be learned by the robots using machine learning artificial intelligence (AI) techniques. This award will provide a U.S. graduate student with the opportunity to work in collaboration with Professor Liz Sonenberg, an expert in human-robot teaming, in the Department of Computing and Information Systems at the University of Melbourne, Australia. To computationally model the robot?s AI, a Mixed Observability Markov Decision Process (MOMDP) framework will be used. This framework is commonly used as part of machine learning to learn from prior data and optimize on a reward function. The MOMDP framework for this study will use previously collected human-human team communication data. In this framework, a reward function will be implemented where using implicit coordination will provide higher reward for the robot. This will enable the robot to compute the optimal policy and solution space for communicating with the human teammate. This model will be then evaluated through human-subject experiments at MIT starting in September after the EAPSI program. This NSF EAPSI award is funded in collaboration with the Australian Academy of Science.