MITRE | The MITRE Digest | Designing Robot Teams for Military Use

Lieutenant Rachel Goodwin glances at her display in the command and control vehicle and sees that the action is going well. Three robotic vehicles just finished mapping the terrain a half-mile ahead. Robot vehicle Alpha now verifies with its computer vision that an abandoned truck, which didn't show up in reconnaissance photos taken an hour earlier, is sitting on a key route to Goodwin's objective.

Alpha had help locating the truck. A batch of tiny ground sensors dropped by aircraft earlier in the week picked up and recorded the truck's movement on the road up to the moment it stopped. Then the sensors picked up the soft vibration of footsteps moving quickly away from the truck. The truck and footstep data were transmitted to Alpha.

On seeing the truck, Alpha now uses a very sensitive chemical-laser sensor that pinpoints the powerful explosives hidden in the truck. Alpha makes a quick decision and calls its more powerful brother, a rocket-firing robot positioned a quarter of a mile away, to take out the truck. Ten seconds later the truck and its load of explosives is blown to bits. Goodwin smiles and advances her squad to their next objective.

The robots depicted in this scenario are not yet real. They're part of the long-term goals of the Future Combat Systems program, a joint research effort of the Defense Advanced Research Projects Agency (DARPA) and the Army.

Dr. Alan Christiansen, a principal engineer in MITRE's Cognitive Science and Artificial Intelligence Department, leads a team that is exploring approaches to the command and control of robot platoons that could be used in the battlefield. This MITRE sponsored research is targeted to provide technology that will be useful to the Future Combat Systems Program.

MITRE's robotics activities go back to the late 1980s when researchers worked on robot arm tele-operation for the NASA Space Shuttle. MITRE's robotic specialists also collaborated with the Woods Hole Oceanographic Institute to develop an underwater robot that performed autonomous search and retrieval. Today we are working on a variety of robotics projects.

"Our work in the command and control of multiple robots involves a relatively new science. The robotics field hasn't emphasized coordinating teams of robots on the battlefield," says Christiansen. "However, for teams of military robots, reliability demands that the human-robot interaction and collaboration must be well-defined. We must also fully understand how autonomous robots will react in complex, hostile environments."

Christiansen notes that it's not feasible now to build robots that are completely autonomous. However, it is feasible to build robots with some autonomous capabilities, augmented with an interface by which a supervisor can interact with the robots. This supervisory control is appropriate for a wide variety of tasks, such as search and rescue operations. Search and rescue tasks are similar to other military applications such as force protection with robot sentry teams, searching for mines on land and under water, and battlefield reconnaissance. "For example, a robot performing a search and rescue task could go to into an area and find the victims within a known space, map the space, and then report the location of the victims," said Christiansen. "Humans would then conduct the actual rescue or extraction."

Putting together a team of search and rescue robots is a complex process that involves integrating a software architecture, obstacle avoidance and mapping, computer vision and other sensors, navigation, and a command interface.

Currently, MITRE researchers are using three Pioneer 2-AT robots made by ActiveMedia Robots. The 60-pound robots move on four rubber-tired wheels and "see" their environment by using computer vision, ultrasonic sensors, and heat sensors.

In one experiment, Christiansen and his team used computer vision to control a team of robots moving through a twisting hallway. One robot leads the way while the other two robots use "color blob" tracking to follow the leader. A blob is just a continuous region of a single color, such as the robot ahead. Software in the following robots judges the distance and direction of the leading robot by comparing changes in color and the orientation of its color blob.

Because this is a new field in robotics, it's been important to share work with other scientists in the field. The MITRE team tests its command and control theories and learns about advances by other researchers by attending competitions specifically designed for search and rescue robots. One such competition was the 2002 Robot Competition and Exhibition sponsored by the American Association for Artificial Intelligence (AAAI), which was held in July 2002 in Edmonton, Canada. (MITRE attended to practice and learn in preparation for a similar competition, RoboCup-Rescue, in the summer of 2003.)

The AAAI competition involves finding victims in a simulated collapsed building represented by an arena composed of three successively more difficult zones. The arena was designed by engineers at the National Institute of Standards and Technology to provide an objective evaluation of autonomous mobile robots for search-and-rescue operations.

In each zone, a variety of overt and hidden "targets" are placed for robots to locate and identify. In this competition, the targets are victims played by mannequins dressed in regular clothes. Added to the mannequins are heating pads to simulate body heat. Mechanical devices make the arms move and small tape players play "Help me!" To find the victims, rescue robots must contend with overturned furniture, stairways, collapsed floors, and broken pipes.

Knowing the location of victims is crucial to the rescuers, so it's critical that the robots provide a map of the search area and pinpoint where the victims lie. Originally, the team expected to use computer vision with color blob tracking to detect mannequin victims by their skin tone. But the competition, like real life, had victims with different skin color. And like a real damaged building, gray dust from concrete and plaster coated the victims, changing their skin tones.

Tips for Designing Search and Rescue Robots

1.Use headlights for operating in the dark.

2. Use multiple sensors such as sonar, lasers, and bumpers to avoid obstacles and navigate among acoustically and visually transparent panels.

3. Design localization techniques that avoid or compensate for excess wheel slippage due to dust and loose debris.

4. Keep the robot's physical size small to navigate narrow passages and low overhead obstacles.

Another thing the team learned is the importance of integrating a variety of sensors on the robots. Multiple factors can confuse a particular kind of sensor. For example, in dark areas computer vision can't be counted on. Lasers can be used for determining position, but they don't work if there are mirrors or transparent panels in the way. Sonar can be used in some instances, but if fallen acoustic ceiling tiles lay against walls, the sonar doesn't work well because the unit's sound waves are absorbed instead of reflected back. In addition, sound waves don't bounce back predictably off objects that have sharp angles, and can give misleading images.

The team also learned the importance of providing a good interface to control a robot in unknown terrain, says Laurel Riek, a MITRE senior artificial intelligence engineer. "The video window on the laptop must be updated frequently to give us the best control of the robots," says Riek. "One of our challenges is to figure out how to design the interface so the user can control a robot easily and accurately. Another challenge is how to control multiple robots at once."

Designs of the Future

In future designs the team plans to use optic flow, which detects changes in the robot's field of view image. Optic flow works by comparing these changes from one video frame to another. For example, if an object is seen to be getting larger from frame to frame, the robot can deduce that it is getting closer to the object.

"Another thing we can do with vision is motion detection," says Dave Smith, a lead artificial intelligence engineer. "We keep the robot still for a few seconds and have it look at one area of a room," says Smith. "If it sees anything moving, there's a good chance that it's a victim, but it's not guaranteed. That's why we work with a probability map. If that data gets corroborated with other evidence, either from other sensors or from other robots looking at the same area from a different angle, the probability for that point on the map reaches a threshold. Then the robot can bring it to a rescuer's attention," says Smith.

The MITRE team is also developing special software to show the location of victims on a laptop display. The display would show different color layers to represent what each robot sees. A single composite map combines the input of all three robots to increase the probability of finding victims. The input from the different sensors can also be isolated in layers to determine, for example, what the heat sensors detect. If the human controller sees a promising area with potential victims, the robots can be directed to that area for further exploration.

For future research and competitions, the team will add a tracked PackBot robot that can right itself if it's flipped over on its back. As the team continues its research, it will help bring the goal of completely autonomous reconnaissance robots closer to reality. And future commanders such as Lt. Goodwin will have the confidence that their robot platoons will complete their missions.

—by David Van Cleave

	Solutions That Make a Difference.® Copyright © 1997-2013, The MITRE Corporation. All rights reserved. MITRE is a registered trademark of The MITRE Corporation. Material on this site may be copied and distributed with permission only.
	Privacy Policy \| Contact Us