CommonGrid: Building Common Ground through Belief Maintenance in Situated Communication

Topics on Multiple artificial embodied agents have shown their popularity in recent years. Compared with asking one omnipotent agent to accomplish all the tasks, multiple agents with fundamental functionalities are able to widen the task space through well-correlated collaboration. For example, given a task to lift a heavy television, one agent may find it hard to lift it up and move safely by its own, while two agents that hold each side of the television can easily accomplish it. Such a well-correlated collaboration reveals new possibilities for extending the task space to more complex genres, meanwhile achieving it with relatively low cost compared to training and creating an omnipotent agent.

To achieve a well-correlated collaboration among multiple agents, a wise communication between agents is highly significant. In prior works, a centralized system that holds all the information and control can to some extent get rid of communication between agents. However, if we involve humans into collaboration, a centralized system is no longer available. Therefore, we expect our embodied agents to have independent self minds, wisely communicate, and develop their own believes of the whole world when completing the tasks.