By Keith Hautala. Photos courtesy of Carnegie Mellon University.
The rescuers search for survivors in the darkness of a vast labyrinth, deep below the surface. They squeeze through tight spaces, navigate blind turns, scramble over obstacles, and struggle to avoid innumerable traps laid for them. One wrong turn could spell disaster. Communication is limited. And time is running out.
The “survivors” in this case are only mannequins, scattered throughout a section of the Louisville Mega Cavern, a former limestone mine that stretches downward and outward beneath 100 acres of Louisville, Kentucky. The rescuers are a band of semiautonomous robots developed by Team Explorer, a group
of graduate students, postdoctoral scholars, and faculty researchers from Oregon State University and Carnegie Mellon University.
They’ve come here, for three days in late September, to compete against other robot-human teams in the third and final round of the DARPA Subterranean Challenge, sponsored by the Defense Advanced Research Projects Agency. At stake is a prize pool totaling $5 million, with the largest share going to the team that locates the greatest number of mannequins and other hidden artifacts — 40 in all, ranging in size from a cellphone to a backpack — in a 60-minute run.
The SubT Challenge seeks “to better equip warfighters and first responders to explore uncharted underground environments that are too dangerous, dark, or deep to risk human lives.” Over a stretch of nearly four years, the competition has engaged multidisciplinary teams from around the world to devise creative solutions to map subsurface networks, on the fly and in unpredictable conditions. It’s the kind of challenge that, by design, strains at the limits of existing hardware and software capabilities — in terms of autonomy, networking, perception, and mobility.
Members of Team Explorer rally behind their banner at the final round of competition, held in September in Louisville.
Geoff Hollinger, associate professor of mechanical engineering and robotics at Oregon State, earned his doctorate from Carnegie Mellon in 2010 and collaborated with the CMU Robotics Institute on an autonomous tunnel-mapping project from 2015 to 2017. So, when the call for proposals came out for the SubT Challenge in early 2018, CMU tapped Hollinger to join them in putting together a dream team. Explorer’s proposal was one of only seven selected by DARPA to receive phased funding of $1.5 million per year, for up to three years of competition.
Rather than use off-the-shelf hardware, Explorer assembled a tight combo of rugged, tractorlike robots and collision-proof drones built from scratch by CMU engineers. The Oregon State contingent led development of the multirobot coordination algorithms that determine where the team should look for artifacts, while also making major contributions to the object-recognition system and user interface.
“There have been many technical challenges involved in getting the robots to successfully coordinate without redundancy and achieving reasonable object recognition with the perception systems, including camera, laser, and gas sensors,” Hollinger said. “The biggest challenge has been integrating all of the systems on our unique team of robots.”
Explorer comes into this final round strong, having proven
a worthy contender in two preliminary trials. The team came in first place in the Tunnel Circuit in the fall of 2019 and second in the Urban Circuit in early 2020. A planned third preliminary round, the Cave Circuit, was scrapped because of travel limitations imposed by the COVID-19 pandemic. The Mega Cavern event includes challenge elements from all three subdomains.
“It was especially challenging to develop computationally efficient approaches that could run in real time onboard the robots as they traversed these diverse underground environments,” said Robert Debortoli, a doctoral student in Hollinger’s group.
Bereft of GPS and cellular connectivity, the robots must depend solely on their own systems to coordinate movements and relay information. Only one human teammate is allowed to provide direct supervision during the competition. Adding to the difficulty, the robots are going into the Mega Cavern blind, with no trial runs and no maps to guide them.
“The SubT Challenge requires the robots to perform well at the first attempt in an unseen environment,” said Graeme Best, a postdoctoral robotics scholar at Oregon State. “This has pushed the teams to develop solutions that work robustly across a wide range of environments.”
Explorer’s strategy involves having the wheeled robots scatter throughout the mazelike course, launching drones from their tails to extend the search farther, and distributing communication nodes along the way to create an ad hoc network.
As the timer ticks down to zero, that approach appears to be working. Explorer has lived up to its name, surveying more of the course — an impressive 93% — than any other team, earning it the “Most Sectors Explored” award. However, the team fell a bit short in points scored, coming in fourth place overall. In the end, Explorer’s robots proved a bit too adept at identifying possible artifacts and overwhelmed their human teammate with reports, including some false positives.
“We did a great job exploring the course, and ultimately we ended up sending back more artifacts than the operator could handle,” Hollinger said. “The operator ran out of time and wasn’t able to get to all of the ones that we saw.” Also participating on Team Explorer from Oregon State were master’s students Yu Hsuan (Chris) Lee and Emily Scheide, and recent graduate Manish Saroya, M.S. robotics ’21. Lee won a best paper award in May at the IEEE International Conference on Robotics and Automation for work performed during the competition.