The word ''
Articles • Names • Photos • Contact

Army teaching robots to understand language

By C. Todd Lopez

WASHINGTON (May 30, 2017) -- Earlier this month at the Pentagon, a robot called "RoMan," which stands for "Robotic Manipulation," demonstrated how it could pick up a tool box and then put it back down. Nearby, a scientist from the U.S. Army Research Laboratory (ARL) manned the robot's controls.

The future of Army warfare will likely involve robotics deployed to perform an array of functions that Soldiers today perform for themselves, ARL researchers say. But what will likely not be part of that future, they say, are Soldiers who operate those robots with a controller or stick. Instead, those robots will operate independently, after having been issued verbal instructions by a Soldier.

A pentagon icon.

Dr. Stuart Young, who serves as the chief of the asset control and behavior branch at ARL, was at the Pentagon in May as part of the annual Lab Day exhibition. Along with other DOD researchers, he manned a booth in the Pentagon courtyard where he met with military personnel to discuss his team's current projects and innovations.

Young said his team's project involves making robots more capable of understanding natural language, so that they can function more like members of a team rather than just tool that has to be controlled by a Soldier.

Robots, he explained, should be more like teammates to the Soldiers they work with, and should, like the Soldier, understand the intent of the mission they are executing.

"We are focusing on natural language dialogue, which is a bi-directional conversation we can have with the robot," Young said. "I want to be able to naturally engage with the robots. So if I for example say go to that building over there, you might say 'which one?' And I could say 'the tall one.' It's a way for us to have a dialogue and disambiguate what you intend for the robot to do."

Right now, Army robots are tele-operated and require Soldiers to control them directly. With the current technology, Soldiers who are operating the robot must stop their other duties to focus on directing the robot. To operate a robot, for instance, the Soldier might have to put down his own gun and pick up a controller. This might mean that another Soldier has to provide security for that Soldier while he controls the robot -- so operating a robot might take two Soldiers out of the fight. That's got to change, said Dilip Patel, a researcher from General Dynamics who works with ARL.

"Now you have a Soldier managing and operating a robot, instead of having their fingers on the trigger," he said. "We don't want that. We want these robots to be autonomous so you just tell the robot, 'Robot, go in that building. And if you have to break down the door, break down the door. Go inside and communicate with the occupants and tell me what is going on in there.' So the robot is looking at the Soldier and listening, through natural language communication, and knows exactly what to do."

Patel is part of the Robotics Collaborative Technology Alliance Program, managed by the ARL. He explained the goal of their program -- which focuses on robotic perception, intelligence, human-robot interaction, and dexterous manipulation and mobility -- is to provide autonomous, robotic teammates to dismounted Soldiers.

"When dismounted Soldiers are approaching an area that requires somebody to go in and either break down a door or interact with an improvised explosive device, we don't want our Soldiers doing that," Patel said. "We want the robots doing that. That's the purpose of this program, to lay the foundation, the research, that's going to make this happen."

There's a lot more work to be done to get robots to respond to Soldiers the way other human Soldiers do, Young said. And there's a lot of work to be done getting robots to perform like Soldiers as well.

"The robots don't have common sense," Young said. "So we use tools like knowledge bases to be able to imbue the robots with that common sense understanding of the world, so they can understand what the human wants them to do, in context. And once we understand that, then the robots have the difficult challenge of being able to understand the environment. It's perception of the world."

Young said that robots must understand their environment beyond just knowing what things are -- like a building, or stairs or asphalt, for instance. They must understand what those things mean to them as well.

"The grass, the concrete: Those are not just things. They are things you can walk on, or you can drive on," Young said. "If I detect mud, well mud is something you can drive on, but it might not be as optimal, or it may have a higher risk -- I might get stuck. So that's the type of thing we have to look at. And then that goes into some sort of cognitive architecture so we can do intelligent behaviors."

Young also said those robots need to be able to adapt to environments without having prior experience in that particular setting, and they must do so at a speed that allows them to keep pace with Soldiers performing the mission.

"We need platforms that can manipulate the environment and maneuver with the Soldiers at operational tempo, at the speed they need to operate at, so the robots are not slow and deliberate, but are able to keep up and have resiliency," he said.

Dr. John Fossaceca, a research scientist at ARL, predicts that the robot as a teammate concept is about 20 years away for the Army.

"In the future, I envision that Soldiers and robots will be working together as teammates, fully dismounted, and robots will be able to understands the intent of the commander, and will have a feedback loop where the robot can get clarification of instructions and understand even the tempo of the operation, understand the context of the environment we're working in, be able to do automatic exploration, come back with reports, and basically save Soldier's lives," he said. "In 20 years, this will be working."

A tiny four-by-four grid of dots. A tiny representation of the Mandelbrot Set. An oscillator from the Game of Life. A twisty thing. A snowflake.