Building Conversational Agents for Military Training: Towards a Virtual Wingman

In military training simulations there is an increasing need for synthetic agents as team members that are capable of replacing human role players. Human-agent interaction using spoken natural language is an important capability for such agents, especially when communication skills are part of training. Interaction technologies such as speech-to-text, text-to-speech, natural language understanding and dialogue management are maturing, though their use in human-in-the-loop simulations is scarce. In this paper we address design challenges for building conversational agents in the domain of military training. In this domain, agents often have to adhere to specific communication standards, protocols and use of jargon. We propose a data-driven design method to tailor conversational agents to a particular application domain, while minimizing human authoring. We demonstrate this method within an existing military training simulation capability and show that agents are able to effectively build shared situational awareness and coordinated their activities in a shared task environment.