Robot Teamwork Could Help Save Lives in Search-and-Rescue Missions
By Hannah Jane DeCiutiis
Reporting Texas
Robots such as the iRobot 510 Packbot were used during the cleanup of the Fukushima nuclear plant. Robots can operate in disaster conditions, including high radiation levels, that would be too dangerous for human rescuers. Photo courtesy of iRobot Corp
Amid the wreckage of disaster sites, robots have been working alongside humans on large-scale search-and-rescue missions since 9/11.
Robots have significant advantages for rescue work: They can operate in environments where humans can’t, such as in the high radiation levels from the Fukushima Daiichi nuclear plant disaster in Japan in March 2011. They can help survey a disaster site to reduce danger to human rescuers. During 9/11, robots located the remains of many victims in the rubble of the World Trade Center.
Rescue robots would be even more useful if they could operate in teams, as human rescuers do. But they were programmed differently when they were built, and robots sent by different companies or universities to help in disasters can’t coordinate with robots they’ve never met before.
A researcher from the University of Texas at Austin has developed a way to overcome that limitation, allowing autonomous robots from different sources to communicate with each other and work in teams.
Sam Barrett has developed an algorithm that allows robots that have never interacted before to complete tasks together. His thesis, “Making Friends on the Fly: Advances in Ad Hoc Teamwork” was published in 2014, when Barrett earned his PhD in computer science. He hopes that his algorithm – a set of instructions that would be programmed into robots as they are built – will allow robots to quickly adapt to new teammates in a variety of areas. Search-and-rescue missions, Barrett said, are the clearest and most critical application.
“Humans are very good at adapting to new teammates, for example in games of pick-up soccer,” said Barrett, who now works at Kiva Systems, a robotics company owned by Amazon. “For robots to interact with robots from a variety of other sources as well as humans, I think that it’s desirable for them to exhibit this type of adaptability.”
Christopher Boyer, chief operations officer of the National Association for Search and Rescue, said the typical search-and-rescue mission protocol is dictated by the acronym LAST: locate, access, stabilize and transport. The first step – locating a missing person – is the most urgent and the one where coordinating robots could be most helpful.
“Using autonomous robots who could talk to each other and share information with each other could be highly valuable, and could help teams with situational awareness,” Boyer said. “They could be useful in gathering those pieces of information which help us decide where to search next or where not to search at all. This set of clues emerges and the probability of finding someone becomes higher.”
He said human rescuers perform a wide range of tasks at disaster sites. For example, some survey the sites by vehicle or helicopter, while others use dogs or horses to search for victims.
“They may not have worked together before, but because they understand what the goal is, they can work together to achieve it,” Boyer said, “I would say it would be similar for a group of robots who all have different capabilities – if they all understand that what they’re trying to do, they could work together.”
In his thesis, Barrett presents what he said is the first algorithm to address all three critical areas of robot teamwork: the ability to deal with a variety of teammates, the ability to accomplish diverse tasks and fast adaptation. The algorithm is called Planning and Learning to Adapt Swiftly to Teammates to Improve Cooperation, or PLASTIC for short.
PLASTIC allows the robot to size up a new teammate, decide which of its old teammates the new one is most like, and then adapt accordingly.
Katie Genter, a PhD student who has worked with Barrett in UT Austin’s Learning Agents Research Group, also studies robot cooperation. Genter said robot teamwork would be invaluable in disaster situations, giving the hypothetical example of a search-and-rescue mission where several universities have sent robots to help.
“These robots are only programmed to work with robots that they’ve worked with before,” Genter said. “Say one of your robots breaks – you may have to fly in a robot to replace it. Wouldn’t it be great, though, if UT and MIT’s robots could work together, instead of having to wait two or three days for a replacement?”
Christopher Amato, assistant computer science professor at the University of New Hampshire, has a very different approach from Barrett’s. As a post-doctoral researcher at MIT, Amato worked on solutions to coordinate teams of robots to make the best possible decisions when they are uncertain about the outcome of a situation. Amato’s team has worked on simulated search-and-rescue missions so far, and will begin working this summer on hardware for deployable robots.
“They are complementary approaches, in a way,” Amato said. “I’m making the assumptions that I know who my teammates are and what their capabilities are. Sam is making the assumption that you don’t know who your teammates are, but you know what your environment is.”
Coordinating a team of robots is a complex and somewhat new area of research. Genter said robots working with unknown teammates are almost certainly part of the future for artificial intelligence but that the approach may take off slowly in mainstream robotics. Though algorithms like Barrett’s could be useful, Genter said large robotics companies don’t have an incentive to investigate how to use them because they are financially able to build coordinating teams of robots.
“We see this as the way robots will work together in the future,” Genter said. “At this point, it’s about how many people we get to reach out to so that hopefully someone will pick it up and say ‘this is amazing’ and then start using it.”
So far, Barrett’s work has been applied to computer simulations where robots play soccer together. Applying it to other real-world scenarios such as disaster rescue would take longer, Barrett said.
The robots need a representative set of teammates for PLASTIC to work, and that data can take a long time to acquire. “Overall, I think that this work could be applied given some effort, but it’s going to depend a lot on the scenario,” Barrett said.
As artificial intelligence continues to develop, so will robot coordination, Amato said. His and Barrett’s work are just two pieces of a much larger puzzle that involves humans and robots working together to solve problems.
“Robots are relatively cheap to make now and relatively capable in terms of what they can do,” Amato said. “There’s going to be many more robots around, and they’re going to need to be able to coordinate. This is something I think is very much a big part of the future of AI.”