Learning human-robot interactions to improve human-human collaboration

Most research in human-robot interaction focuses on either the single-human case or the multi-human case where there is direct interaction between the robot and each human. The multi-human scenario in which some of the humans depend on the robot, but do not interact with it directly, is currently less studied. In this paper, we introduce a human-human-robot collaboration task, in which the robot interacts directly with only one of the humans. The goal of the robot is to find the optimal way of helping the two humans achieve their objective. We decided to use meta-reinforcement learning to solve the task, giving the robot the ability to quickly adapt to new human behavior. We trained and tested an agent on a version of the proposed environment that uses simulated human behavior. Initial results show that our task is learnable.