Juiced and Ready to Predict Private Information in Deep Cooperative Reinforcement Learning

In human-robot collaboration settings, each agent often has access to private information (PI) that is unavailable to others. Examples include task preferences, objectives, and beliefs. Here, we focus on the human-robot dyadic scenarios where the human has private information, but is unable to directly convey it to the robot. We present Q-Network with Private Information and Cooperation (Q-PICo), a method for training robots that can interactively assist humans with PI. In contrast to existing approaches, we explicitly model PI prediction, leading to a more interpretable network architecture. We also contribute Juiced, an environment inspired by the popular video gameOvercooked, to test Q-PICo and other related methods for human-robot collaboration. Our initial experiments in Juiced show that the agents trained with Q-PICo can accurately predict PI and exhibit collaborative behavior.