Lessons Learned from the SubT Challenge

Aaron Auperlee

As the countdown started, a boxy robot with four big wheels carrying a host of cameras, sensors, communication equipment, autonomy software and the computing power to make it all work together, rolled down a ramp into a dark tunnel.

It did not know where it was, what was ahead of it or where it was going. It was there to explore.

Over the next hour, more robots followed. Wheeled robots rolled through passageways, tunnels and caves with rocks, rails and obstacles. They pivoted around tight corners and passed through doorways and openings just centimeters wider than their axles.

The Spot robot, a dog-like quadruped often seen dancing in carefully planned and choreographed demos, walked down the ramp onto unknown footing, adjusting its balance and gait as it scanned the terrain ahead.

Drones flew through narrow corridors, bumping into and bouncing off walls. Collisions like these would typically send a drone crashing to the ground, often in a heap of parts. But these drones rebounded, righted themselves and returned to their flight.

Team Explorer deployed eight robots for the final round of the Defense Advanced Research Projects Agency (DARPA) Subterranean, or SubT, Challenge — a three-year competition during which teams from around the world raced to develop robotic systems that could autonomously operate in underground environments like caves, mines or subway stations for search and rescue missions.

Over the course of the competition, Team Explorer created robots and drones tough enough to bang into walls, roll and walk over ruts, rails and rocks, and find their way through dark, unknown and treacherous terrain. The robots worked in unison, telling each other where they had explored, pointing out areas of the course for others to explore, and passing along information about what they had found to each other and the humans staged at the course entrance.

And for most part, the robots acted autonomously, searching the course and coordinating amongst one another with little “hands-on” control by the team.

“I think back to three years ago and what we were thinking about was essentially science fiction then,” said Sebastian Scherer, an assistant professor in the Robotics Institute and a co-lead of Team Explorer. “Teams of drones zipping around and ground robots just going out exploring, it was just not a thing. And now, we'd feel very comfortable sending a drone into some random environment and thinking it will have a really high chance of coming back.”

Team Explorer, a mix of students, faculty and staff from CMU and Oregon State University, lived up to its name during the final round of the SubT Challenge. The team’s fleet of robots mapped nearly the entire course, earning the team the Most Sectors Explored award for investigating more of the course than any other team.

In total, eight teams competed in the SubT finals. As their robots explored and mapped the course, they searched for artifacts that would be important to first responders during an underground search and rescue mission, objects like dummies staged as survivors, cellphones, tools, backpacks, vents and even the presence of harmful gases.

The team accurately identifying the most artifacts won, and would take home the $2 million top prize. Second place won $1 million and third $500,000. Team Explorer placed fourth, missing out on the money by only a few points.

"I'm proud of this team. This was hard, and we really rose to the challenge," said Matt Travers, a systems scientist at the RI and a co-lead of Team Explorer.

DARPA staged the SubT Challenge to accelerate the development of technologies that would enable robots to assist with search and rescue operators in underground environments that are too dangerous, dark or deep for humans. The challenge forced teams to design robotic systems that could operate when communication, vision and mobility would be compromised. The precision needed to maneuver in these situations, the complexity of the coordination and logistics, and the sheer amount of data needing to be gathered made this challenge too large for humans to manage without the help of autonomy.

The first round of the SubT Challenge, the Tunnel Circuit, took place in August 2019 in a research mine operated by the National Institute for Occupational Safety and Health in South Park Township, outside of Pittsburgh. Team Explorer decisively won that round. Round two, the Urban Circuit, was held in February 2020 inside an abandoned, never-commissioned nuclear power plant at the Satsop Business Park near Olympia, Washington. Team Explorer placed second. DARPA canceled round three, the Cave Circuit, because of the pandemic.

The final stage took place in September 2021 at the Mega Cavern, an old limestone mine in Louisville, Kentucky. Inside the mine, DARPA built a course that incorporated the tunnel, urban and cave circuits of the previous rounds. The finals course included a narrow mine with rails running down the center and obstacles: rocks, barrels and supplies cluttering the sides. There were tight cave sections with stalactites and stalagmites. DARPA even built a subway station with a platform, rails disappearing into a tunnel, hallways, storage areas and support pillars.

Timothy Chung, DARPA‘s program manager for the SubT Challenge, said first responders and members of the military invited to watch the final competition walked away impressed. The competitors demonstrated the effectiveness of teams of robots working with teams of humans. In an hour, the robots were able to map and search a course that would have taken professionals far longer. The teams demonstrated uses of ground robots, drones and quadrupeds far exceeding what many thought possible.

“Without a doubt, we have accelerated the timeline of this technology making it into the hands of people who could save lives,” Chung said. “Some of the SubT technologies are ready for showtime.”

Chung expects further development and commercial-ization of technologies pioneered during the challenge to happen within a matter of years. That‘s fast, considering it took a decade or longer for technologies developed during the autonomous vehicle challenges in the 2000s to arrive on city streets. Industries such as construction, infrastructure inspection, real estate appraisal and others are interested in SubT technologies, said Chung.

“This technology will pay significant dividends in the future," Chung said. "It will extend well beyond the subterranean environments. I think we'll see that above ground, in the air and in buildings.”

To push this technology forward, Team Explorer made advancements in both hardware and software. The team's simultaneous localization and mapping (SLAM) software, which allowed the robots to construct a map of the course while they explored it, performed so well in early competitions that other teams used versions of it as they competed at the finals. Work out of Oregon State University, headed by Geoff Hollinger (SCS 2007, 2010), an associate professor of mechanical engineering in Corvallis and Team Explorer co-lead, developed software that helped robots make decisions, see objects and explore better.

A student at Oregon State University developed an algorithm to maximize the use of marsupial robots — robots that carry other robots. The algorithm determines where and when to best deploy the second robot to maximize exploration capabilities. During the finals, Team Explorer successfully used a marsupial robot, launching a drone off the back of one of its ground robots once it encountered stairs in the urban part of the course.

Other research examined the behavior trees of robots, or how a team of robots determines the tasks for individual robots. This was useful to Team Explorer as its robots decided where to go, but it will also be useful to robots in other environments. Emily Scheide, a master's student at OSU when she worked with Team Explorer, is now a Ph.D. candidate at the university studying how behavior trees can help robots assist children with disabilities.

“Anywhere you have a multi-robot system making tough decisions, this will be used,” Hollinger said.

Team Explorer built their own robots and drones for the competition. The wheeled ground robots, named R1, R2 and R3, were built smaller and more agile as the competition went along. R1, the team's first robot — a 400-pound beast — was still effective in the finals, exploring the urban circuit and launching a drone. R3, the newest and smallest of the ground robots, used its size and articulated body to make it deep into the finals course, navigating tight cave sections where other robots feared to tread.

However, the plucky robot's ambition may also have been its downfall. As R3 worked its way to the back of the course, it encountered a steep grade that led to a cliff. The robot triumphantly summitted and began exploring the edges of the cliff, until it got a little too close. R3 tumbled about 20 feet down the cliff, coming to rest on its side. Had R3 been able to report back what it saw during its run through the course, the number of artifacts identified might have pushed Team Explorer into first place.

Throughout the competition, the development of the drones had its ups and downs. The team's early drones had six propellors and no prop guards. They crashed often. To lighten the payload, the team tried to design a drone that used only cameras to see, not LIDAR, but darkness and dust ruled out that approach.

The purpose of searching for water on the Moon is not to find ancient signs of life. Rather, any ice found on the Moon could be invaluable for future space missions. Moon ice could not only provide water for astronauts to drink, but they could also split the water into hydrogen and oxygen to provide breathable air and rocket fuel for missions to such places as Mars, Pluto or the moons of Jupiter.

“One of the challenges to space exploration,” Wettergreen said, “is getting enough fuel off the surface of the Earth to get to Mars or Europa or Pluto. If water on the Moon can be extracted, it’s actually more efficient to get fuel from the Moon than to launch it into orbit from the Earth.”

Finally, Astrobotic and NASA plan to launch a much larger lander known as Griffin in 2023. This lander, weighing more than 1,000 pounds, will have seven rocket engines, as well as dedicated power and communication capabilities. NASA expects the mission to carry the VIPER — Volatiles Investigating Polar Exploration Rover — that will include instruments and a drill designed to extract water on the Moon.

Additionally, Team Explorer built an all-metal drone. It was indestructible, but too heavy. Then the team moved the propellers from above the drone's payload to below it. They also added a robust propeller guard that allowed the drone to bump into walls and obstacles.

These modifications, coupled with flight control software that adjusted roll and pitch when a drone did bump into something to keep it flying, produced a design that caught the attention of other teams at the Urban Circuit competition.

It also caught the eye of a few companies, inquiring about the design. That led Vai Viswanathan, a member of Team Explorer who left after earning his master of science in robotics from CMU, and Jay Maier, a mechanical engineering student at the University of Pittsburgh, to spin off the technology and found Canary Aero. The company makes drones that can carry a heavy payload and get close to and even bump into objects without crashing.

“You don't have to be too afraid of hitting the walls,” Viswanathan said. “It gives you a little more freedom.”

Canary Aero already has a few commercial customers lined up. Team Explorer used a Canary 15, named for its 15-inch propellers, in the finals. The drone‘s collision tolerance was certainly put to the test. While exploring the urban portion of the course, the Canary drone bounced off the walls in a narrow hallway. Seemingly spinning out of control and headed for the ground, the drone righted itself and kept on going.

Robots, whether on the ground or in the air, are not typically designed to bump into things. Team Explorer's were. The team incorporated robust designs and software that allowed their air and ground robots to hit walls and survive, and to free themselves from getting stuck. Scherer said robots typically stay away from those situations.

“But that’s not always the right answer for something that explores,” Scherer said. “Sometimes you need to bump.”