Subterranean Showdown: DARPA Pushes Underground Robots to Their Limit

DARPA photo

LOUISVILLE, Ky. — It was 6 a.m. and Lt. Col. Dan Riley was laying in his hotel room, staring at the ceiling.

Though he is an active duty Air Force officer, Riley was not “chair flying” an airplane, a pilot’s way of practicing procedures before a flight.

Instead, his mind was far below ground — running through every possible obstacle facing the menagerie of robots under his control in a network of caves, tunnels and urban underground environments.

As an operator in the final competition of the Defense Advanced Research Projects Agency’s Subterranean Challenge, he was the only one on his assembled 16-person robotics team MARBLE who was allowed to issue commands during the perilous hour of competition.

Only the day before, professional sportscasters hired by DARPA analyzed and speculated about failed unmanned aerial system take-offs, friendly robot collisions and fog-obscured sensors in the challenge’s preliminary runs. He also was the lead for the competition that took place in the virtual realm, helping write the code for the autonomous systems to explore a cave and tunnel simulation. But Riley’s long history in the Air Force equipped him for racing against the clock and operating delicate systems under pressure, he said.

“There’s definitely a lot of stress involved,” he said. “You feel the weight of everyone’s expectations riding on what you’re going to do.”

This challenge was the final round of DARPA’s SubT project aimed at accelerating the development of autonomy and robotics for search-and-rescue operations. Firefighters and first responders could soon command fleets of robots capable of pinpointing unseen hazards and locating survivors, Program Manager Tim Chung said in late September under the glow of the stage lights that filled the event’s watch party at the Louisville Mega Caverns in Kentucky.

With $3.5 million on the line for the final challenge, it was not the time to play it safe.

On the surface, the eight teams competing in the systems challenge had a simple objective: locate as many objects as possible in one hour. But as the three winners who left the caverns with huge, lottery-style checks for $2 million, $1 million and $500,000 would find out, the labyrinth that took weeks to construct would test the limits of mapping, autonomy, robotics and communications capabilities that some of them had been developing for years.

The course was designed to highlight what is possible for the warfighters and first responders who could one day use the mapping, austere navigation, robotics and autonomous hardware and software to conduct subterranean operations, Chung said.

Prize-winning teams Cerberus and CSIRO Data61 scored 23 points, and the lowest scoring teams — Robotika and Coordinated Robotics — each detected two objects.

After an earthquake or a collapsed building, “There’s always a gear turning in the back of my head — and for many of our competitors I’m sure ­— that in the event of some kind of emergency, what one of our robots could go out and help?” he said. “You’ve been able to push the entire frontier of this technology, but there are many people out there in this audience are grateful for even just the one piece of technology that you’ve advanced. That’s so meaningful to me.”

During the competition, each team had one hour to deploy robots onto an obstacle course made up of three subdomains — cave, tunnel and urban underground — which was a combination of the terrain from the previous three competitions in the subterranean series. DARPA hid “artifacts” such as thermal vents, mannequins, cellphones and backpacks for contestants to locate using the robots.

Once the systems detected an object using the laser beam scanning technology lidar or other sensors, they would report the location to their operator. If the location reported was within five meters of the actual location, the team scored one point. If the robot made a mistake, the team lost one of their available 45 report attempts and didn’t score. The fastest time for the final point would decide the winner of the competition if there was a tie.

This configuration emphasizes accuracy, which first responders will need because the stakes are high, Chung explained. The robots could potentially save lives by both locating survivors and identifying hazardous areas first responders should avoid.

“What they really need is the ability to understand where the dangers lie, even before they want to dispatch their units to respond to a natural disaster or a similar scenario,” he said.

Because of that, there is no minimum score needed for teams to demonstrate that they have technology that could be useful for first responders, he noted.

Chung said he hoped the teams would connect with representatives from the Army, Marine Corps and civilian agencies to further their technology.

“[I] encourage you now that you’ve developed this capability to go to … those firefighters and those first responders and find out really what they are needing, if there’s a little deviation that will make an impact on their day-to-day” efforts, he said.

The teams — and the technology — have come a long way since the series of SubT-related events began in 2018, Chung said. He joked that there was a lot of “colorful language” used to describe the level of difficulty of the first circuit in 2019.

“[Chung] came in with a vision, which all of us inside DARPA thought was crazy,” said Stefanie Tompkins, the agency’s director. “I heard somebody say off to the side and somebody say directly to my face that when they first heard about this, they were absolutely positive [that] it was impossible. So thank you for ignoring your gut instinct and diving into the competition,” she told competitors.

The majority of the eight teams in the final competition scored in the double digits, a marked improvement from the two teams that earned double digit points in the 2019 tunnel circuit.

Additionally, the competition proved that technological innovation doesn’t have to come at the expense of the government, Chung said.

According to budget documents, DARPA’s bill came out to about $82 million over the course of five fiscal years, as innovators put their own dollars on the line as well.

Team CoSTAR’s Joel Burdick, one of the few participants who had been in the running since 2019, said DARPA’s funding wasn’t enough to cover the bills for its “Spot” legged vehicles, wheeled vehicles and aerial drones. The team had to look for other methods such as grants or equipment donations from “day one,” he said. “Any way you can stretch a buck,” said the professor of mechanical engineering and bioengineering at the California Institute of Technology.

Similarly, Coordinated Robotics, another team made up mostly of student researchers and their professor from California State University-Channel Islands, had to get scrappy to deploy their fleet of 10 robots. The team was self-funded for the last round of competition, and it couldn’t afford the legged robots built by Boston Dynamics that many other teams used, said Hugo Quintero, a member of the pit crew.

“It would have been an amazing opportunity to work on a Spot, but now, we have plenty of work here with these, and they did pretty well,” he said.

Different obstacles represented real challenges that could arise in disaster settings, said Viktor Orekhov, designer of the course and a Booz Allen Hamilton contractor. The mobility, perception, autonomy and networking of the robots were tested by different artifacts and their locations, he said.

To measure mobility, the course had a variety of different environments for the robots to traverse. For example, stairs, which were easy for the legged robots, were harder for wheeled systems. One treaded robot jumped its tracks when it tried to go over railroad tracks DARPA put in its path, Orekhov observed.

The large size of the cave and the sheer distance the robots had to travel to communicate tested the limits of the robots’ programmed autonomy. Only the operator had contact with the machines throughout the run. If the algorithms failed, the robots could end up circling endlessly in one area — like a one-legged unmanned ground vehicle that traveled up and over a bridge did during the competition.

“It’s autonomously exploring, so it kind of got caught in the loop and it went up and over that bridge probably seven times in a row,” Orekhov said. While the system was confused, it still managed to complete a great feat, he noted. “It’s autonomously exploring and managing to traverse a terrain — that’s a big deal.”

Building trust in autonomy for the first responders who will eventually use the tech is partly why the challenge is important, Orekhov said. If the robots are going to save lives, first responders have to trust that the machine will be able to return with valuable information.

“In disaster response scenarios, there are lives on the line. If I send a robot out, if I only restrict it to communications range, it’s only useful so far,” he said. “But if I can trust it beyond communications range ... now that system is way more useful and beneficial to humans.”

The length of the course — about 2,900 linear feet — pushed communications to the limit. Distance between the home base forced teams to figure out new ways to transfer information, including dropping communications nodes for navigation and a truck that was designed to carry a very long ethernet cord into the course, extending the comms range.

Dynamic obstacles like fog and smoke could damage the robots’ sensors, impairing their ability to perceive their surroundings. The diversity of artifacts meant that it wasn’t enough for the platforms to have one kind of sensor.

The course was so difficult that there was one artifact — a fire extinguisher — that was not found by any robots on any run, Orekhov noted.

The amount of fog in the area prevented the systems from detecting its existence. By the time the fog cleared out, most robots had moved on to the other sections of the lengthy cave, he said.

The COVID-19 pandemic produced its own set of challenges. CSIRO Data61, the team that won $1 million for a second-place finish, had members who were prevented by Australia’s travel ban from leaving their country to work on the robots, said Navinda Kottege, the group’s leader. The Commonwealth Scientific and Industrial Research Organisation had to ship its platforms to its counterparts at the Georgia Institute of Technology and train technologists there to operate and adjust the robots.

“Our biggest concern was damage to hardware, because we couldn’t send down our engineers and electronic materials,” he said. “In the competition that we’ve had in the past, there had been some damage after each run and we were nervous about that.”

At one point, the team considered not deploying any robots for the preliminary rounds to protect them from the treacherous falling ceilings, sharp drop-offs and railroad tracks that lay in store.

“Given that, for us to come on top of the leaderboards in preliminary rounds, that was completely unexpected,” he said.
Kottege himself maintained a presence in the team garage by video conferencing into the cave using an iPad on a Segway scooter.

Another international team from the Czech Republic, CTU-CRAS-NORLAB, was not able to start practicing with its robots until August because of delays in funding, explained Tomas Svoboda, a program lead.

“We really had no time to do a real integration test,” he said. “We were preparing, we were designing [the] payload, integrating the sensors and preparing for that. … [But] this was actually the first time the complete system ran together.”

While all of the teams in the systems competition — and the virtual competition that took place simultaneously — developed techniques to push the boundaries of robotics capabilities, some of the competitors stood out.

Cerberus, one of the only groups to utilize legged robots throughout the entire competition, developed its own variation of Boston Dynamics’ “Spot” robot, which they called the ANYmal.

Kostas Alexis, program lead for the team, said his was one of the first groups to invest in legged robots from the beginning because it had a vision for employing their advanced mobility. For example, legged robots can walk down stairs and can right themselves if they are knocked over.

Developing the robots from scratch gave the engineers the advantage of having access to the low-level capabilities of the systems, he said.

“That means we can make them more adjusted to the competition because we have access to the software up to the last detail,” he noted.

Additionally, simultaneous localization and mapping, or SLAM, was used by robots to create a map and calibrate their position using data collected from their sensors. The CSIRO Data61 computer engineering team has been refining its proprietary technology package called Wildcat SLAM to maximize its accuracy in the competition, Kottege said.

In addition to its second-place finish, DARPA recognized the team for reporting the location of a drill within 5 cm, the most accurate in the competition. According to Kottege, other teams also used CSIRO’s mapping technology.

“For us, that’s a great outcome to have it used by multiple people,” he said.

But even lower performing teams brought unique advancements. CTU-CRAS-NORLAB finished with just seven points in the final systems competition. However, its electric six-legged robot that was developed for the competition still could have an impact on the industry, said Svoboda. The team decided not to use it in the competition because it was too big for some of the constrained tunnels, but he said it could have implications for robotic mobility. The robot measures the level of force needed to push against the terrain as it moves instead of relying on sensors to determine how much force is needed, he explained.

“It’s also interesting to control it in a reactive way,” he said. “For instance, if you enter some slippery terrain and the robot starts drifting, you can somehow do the countermeasures against it.”

While the systems competition didn’t pan out for CTU-CRAS-NORLAD, the team took $500,000 for the virtual contest.

At the end of the day, the determination and grace under fire for Riley’s team MARBLE paid off. MARBLE won $500,000 for its third-place finish with 18 points and achieved the balance of autonomy and human interaction that DARPA was looking for, he said. He recounted one moment in the final run when he was able to tell a robot to defy its programming and keep going through the fog. Because he was able to turn the platform around, the team scored five more points.

“Being flexible was key to our success in the end,” he said.

Chung added that because DARPA was ahead of industry in solving subterranean problems, the mission needs — such as improved mapping and reliability — have stayed consistent. Now, there is a foundation from which everyone can build, he said. For example, the testbed used for the virtual competition is available online for any interested technologists to put their own autonomous software through its paces.

“There’s no limit to where these technologies can go from here,” Chung said.

Topics: Robotics, Robotics and Autonomous Systems