ALBUQUERQUE — Before artificial intelligence wields weapons in real-world battle, the Pentagon needs to test its AIs in virtual environments. That’s why DARPA, the Pentagon’s blue-sky projects agency, is developing an AI that can communicate with humans in games, as well as manage and understand the peculiarities of a human team in action.

See, central to the US military’s enthusiastic embrace of artificial intelligence is the idea that AI and robotics can never replace human troops, but they can work with them in a human-machine team that is much more than the sum of its parts. The military envisions armed drones that fly alongside manned fighters as expendable “loyal wingmen” or scout ahead of human soldiers, AI targeting systems that line up shots for human gunners, and a host of virtual staff officers that do the organizational grunt work on everything from maintenance schedules to strike planning. And AI will be essential to coordinate the staggering complexity of future Joint All-Domain Operations across land, air, sea, space, and cyberpspace.

But if soldiers are going to rely on AI in battle, they will need to trust that the AI is capable of performing the mission, and that they will be able to communicate with the AI to change plans as circumstances change. This is the heart of the Adaptive Distributed Allocation of Probabilistic Tasks, or ADAPT, a $1 million contract DARPA awarded to Aptima, a Woburn, Mass.-based engineering firm.

To explore this kind of human-bot collaboration, Aptima is working on a testing environment that is entirely virtual and familiar to many: Minecraft.

“Teams of humans that are going to be doing some work in Minecraft and getting advised by this AI model,” said Adam Fouse, Aptima’s ADAPT Program Manager. This isn’t Aptima’s first strike at this sort of work with DARPA. It builds on another DARPA contract awarded to Aptima, the similarly acronymic ASIST, or Artificial Social Intelligence for Successful Teams.

The goal is to find a way for humans to communicate information to an AI agent that the AI itself might not have, perhaps because it lacks the sensors to perceive it, perhaps simply because it’s in a different place. Likewise, said Fouse, the AI needs to be able to “communicate back to humans in ways that are efficient and useful, so they can understand at a glance kind of the insights of the AI.”

Once in the virtual environment, the AI will have an opportunity to observe the humans it works with as they actually perform tasks. Building from this knowledge, and from repeated exposure to humans executing these tasks, the AI will work to assign humans and AI agents to perform the tasks for which they are best suited.

Minecraft is well suited for multiple players, part of what keeps it one of the most played games in the world. It allows players to collaborate on tasks ranging from exploration, to demolition, to building, and it includes the option of in-game enemies and combat. Many of the code libraries used to build Minecraft are open source, so it’s very easy for people to modify —  in this case, for Aptima to insert a custom in-game AI agent.

The Minecraft environment is useful for another reason: it allows researchers to record and learn from how people actually work together on tasks, and then observe how they will work alongside an AI agent. Then it allows the AI to see how the people it works alongside actually fight, and adapt.

How might ADAPT manage human-AI teams? For example, Fouse said, “we know that Joe over there always tries to go as far into buildings as possible, to start in, no matter what he’s told. There’s part of the team tasks that says, ‘Well, we need someone to do that. Let’s give that task to Joe.”

If the AI can anticipate human action, and plan accordingly, in the virtual environment, then that AI could become a valuable battlefield aid for a commander. With technologies built upon ADAPT in hand, it would allow human commanders to better understand how a team of humans and, eventually, robots are best equipped for a task, and plan accordingly.

Like most DARPA projects, ADAPT is an early bid in an attempt to seed a new kind of tool. That tool, derived from ADAPT, could be a program that runs on a squad leaders tablet, or it could be scaled up to higher command levels. With this AI-informed tool in hand, commanders could know their own forces’ inclinations and strengths, using recorded evidence, and then use people in battle to their fullest potential.

Using AI to both understand and manage humans in this way is ultimately part of a greater project of reducing the cognitive load of battlefield management. The more an AI can function as a sort of personal assistant for combat, the less time commanders have to spend on minutia. In this way, the AI takes on some of the mental burden of orchestrating a fight, and it lets the people fighting do what they do best, even if that is a headlong charge against the odds. And then, with AI derived from work done on ADAPT and other DARPA AI projects, the machines managing the background detail of war can adjust to human behavior and salvage victory out of human fallibility.

Before any of this ends up in the battlefield, the Minecraft environment should allow the AI to figure out humans, and work alongside them, in a setting where the gravest risk is a misplaced pick-axe, a dropped block of stone, or a pixelated zombie.