Field Report: Lessons From First Leg of DARPA Subterranean Challenge

Team PLUTO prepares to enter the underground mine in the DARPA Subterranean Challenge. Image: Ray Linsenmayer

August 29, 2019      
Ray Linsenmayer

In a dark and damp abandoned coal mine outside of Pittsburgh, competitors from around the world gathered last week to see how robots, both ground and aerial, can best locate humans and objects underground. The tunnel circuit, the first of three environments planned over the next two years, was part of the DARPA Subterranean Challenge.

Over the course of four days, 11 teams participated in simulated search and rescue missions, with early lessons learned that could have big implications for the military, first responders, and even space exploration.

DARPA Subterranean Challenge Timothy Chung

Dr. Timothy Chung explains the details of the challenge during the opening briefing at the Subterranean Challenge media day. Image: Ray Linsenmayer

DARPA placed 20 artifacts inside the coal mine, and teams were tasked to find as many of these as they could. The challenge was designed to “focus the technology on mobility, autonomy, networking, and perception,” said Dr. Timothy Chung, a program manager at DARPA. “It’s not about having technology that only works in one location at one time, but about having technology that is versatile and agile.”

In six months, the robots will be tested in urban areas, and six months later it will be caves. “This doesn’t give teams enough time to go back to the drawing board” and completely rework their robots, notes Chung. “They have to be thinking about these different terrains from the first competition.”

Of the 11 teams that took part in the tunnel circuit, seven were “Track A” and funded with research grants from DARPA. These included six led by academic institutions, and one comprised only of private companies. The remaining four were self-funded “Track B” teams. “Track B is what makes the challenge at DARPA so awesome,” said Chung. “Anybody from anywhere can qualify and compete. These teams can provide insights that benefit everyone.”

Each team was given 60 minutes in each of their four runs in the mines. To earn a point, teams had to find an artifact, identify it, locate it, and then report it back to the DARPA command post within an accuracy of 5 meters. Teams had had very little interaction with each other before convening in Pittsburgh, and ended up adopted a dizzying array of approaches to tackling this mission.

DARPA Subterranean Challenge Team Explorer

The large wheels on the Team Explorer robot helped it win the underground mine circuit. Image: Ray Linsenmayer

Mobility design differences

There was the most variety between teams in the way they approached mobility in the Subterranean Challenge. Team Explorer, made up members from Carnegie Mellon University and Oregon State University, won the circuit by deploying ground robots with large wheels. While the team had aerial robots as well, it leaned heavily on the wheeled robots because of their robustness.

“In the urban and cave environments in the next couple of circuits, there will be areas where the aerial vehicles will have an advantage,” said Dr. Geffrey Hollinger, assistant professor of mechanical engineering at Oregon State University, and one of the leads on Team Explorer. “But in the tunnel environment, many of the artifacts were accessible to the ground vehicles. I’m personally pretty happy so far, because I think we’ve covered the majority of the course with these robots. You’ll see a wider array of robots from us in future circuits for sure.”

Team PLUTO DARPA Subterranean Challenge

Team PLUTO is focusing on agility to make sure its robots could tackle difficult terrains in all environments. Image: Ray Linsenmayer

Team PLUTO, made up of members from the University of Pennsylvania, Exyn Technologies, and Ghost Robotics, deployed both legged and flying robots. Its main focus was on making the systems agile so the robots could tackle the different terrains in all the circuits.

“The legged robot has a very high stance and can also swing its legs pretty high,”said Dr. C.J. Taylor, professor of computer and information science at the University of Pennsylvania, and lead for Team PLUTO. “This lets it step over a lot of different things. Similarly, the flying robot has the ability to ascend vertically” to find artifacts.

Other teams, like CRETISE, made up of members from FLIR Systems and Neya Systems, deployed robots with treaded “arms” that boasted a 360-degree range that enable them to tackle very rough terrain. A version of this robot is already sold to the military, according to David Weatherwax of FLIR, and a unique feature is that it can shift its weight forward when it climbs to prevent itself from falling backward.

DARPA Subterranean Challenge Team NCTU

Blimps created by Team NCTU had longer battery life than drones, but faced some maneuverability challenges. Image: Ray Linsenmayer

Team NCTU, a self-funded team from the National Chiao Tung University in Taiwan, deployed blimps. “They have a much longer battery life than drones” said Yiwei Huang of NCTU, although the team admitted that maneuverability in the mines was a challenge. He said that for the next circuit, the team is considering adding some motors on the side so the blimp can change direction more easily, and will probably make the blimps smaller so they fit better underground. Several teams mentioned that their drones had crashed while completing the circuit.

Aerial vehicles experienced both battery and payload constraints. They “get about 15 minutes of flight time,” according to Dr. Sean Humbert, professor of robotics and system design at the University of Colorado Boulder, and the lead for Team MARBLE. “That basically gets you down the first hallway.” The wheeled robots his team deployed “don’t need a battery charge, and can run for the full hour.” He said the “urban underground challenge next February is going to be very aerial vehicle focused. There are going to be stairs and large dropoffs. I’m thinking this will be like a DC metro station or something. I think a lot of the platforms that are having success in the mines won’t necessarily be able to translate their success” the same way in the next circuits.

DARPA Subterranean Challenge Team MARBLE

Team MARBLE focused on autonomy and a long battery life for its ground-based robot. Image: Ray Linsenmayer

Autonomy

Many teams focused on multi-robot coordination. This will only become more difficult in future competitions, said Hollinger of Team Explorer, “because we’ll be thinking about larger environments, more 3D structures, and potentially more robots in these environments.”

Penn’s Taylor said Team PLUTO decided to make its legged and flying systems completely autonomous. “Once they are launched, they go into the mine and make all the decisions themselves. Like where to go, how to avoid obstacles, and when to communicate. They detect the artifacts on their own, and relay that information back to the base station. Once the robots are launched, the base station operator’s only job is to look at what comes back and check it off before it goes back to DARPA.”

Humbert of Team MARBLE agreed: “Once you escape line of sight, you really don’t have any way to communicate with the base station. What we’ve done to compensate for that is to build really strong autonomy into our platform.”

Networking and communications challenges

Closely related to autonomy, networking and communications were cited by several teams as the biggest challenges of the competition, because radio waves don’t go through rock. Again, each team seemed to have a different way of handling this. “We have a hybrid approach,” said Hollinger of Team Explorer, “where we allow the vehicles to move out of communication with the operator and come back in. That’s where the autonomy kicks in. The operator also has the ability to control and guide the vehicles if they are within communications range.” Team Explorer’s robots dropped communications nodes in the mine to develop an ad hoc communications network.

Humbert of Team MARBLE said their robots also established networks by dropping beacons with tiny routers. “These beacons just fall of the vehicles,” so the robots don’t have to come all the way back to communicate with the base station. “This might save 10-15 minutes over the run,” he said.

Team PLUTO, for its part, didn’t drop nodes because, according to Taylor, “we are focused on full autonomy.” Each of its robots was “outfitted with a mesh networking system that allowed them to connect with each other when they were in communication range. We also built a distributed database system that allows the robots to share information both with each other and the base station.”

DARPA Subterranean Challenge Team Cretise

Team CRETISE’s Cobra robot includes miniature treaded vehicles called “First Looks” that are deployed to maintain communication links. Image: Ray Linsenmayer

Because these systems are expected to go in and out of communication, Penn focused on developing a fault-tolerant system. “Just in case one of these robots goes down, the information is replicated so it has the best chance of coming back towards the base station,” Taylor said. “This could be robot to robot communication, robot to base station, daisy chain, or even a bucket brigade.” Taylor emphasized that his team wasn’t relying on this, and their default assumption is that they wouldn’t be able to communicate with the robots.

One of the more unique approaches came from Team CRETISE. Its treaded robot, Cobra, carried miniature treaded vehicles called “First Looks” into the mines. “We drive the Cobra in as far as we can,” said Weatherwax, “then it drops the First Looks to maintain a communications link. These nodes are dropped off the Cobra, and have the ability to autonomously get themselves right-side up and mobile.

Perception

There was broad general agreement about the sensors the teams deployed. Because weight is more of a concern on the drones, however, harder choices had to be made. In every case, the sensors deployed included some combination of “time of flight” cameras to look underneath the robot, infrared to help detect objects, inertial odometry to help gauge the motion of the platform, and LiDAR systems used for range sensing.

“Using all of this camera data,” said David Weatherwax of Team CRETISE, “we can detect the different artifact types that have been specified. We know there are only a handful of artifact types we’re looking for, so we’re training a neural network to detect those artifacts autonomously. By using the different sensors on board the robots, we can figure out where those artifacts exist in space and transmit that information back to DARPA during the run itself.” The AI does the initial detection, and the human operator matches what he is seeing.

“While there’s wonderful technology here at the challenge,” said Humbert of Team MARBLE, “I think the biggest impact is going to be on the students who are coming out of this program. We’re training a new generation of autonomous and field roboticists. That just doesn’t happen. Sure, we can go out in the lab and drive our little robots down the hallway of the engineering building, and that’s good enough for paper. But to have this field experience and to have the students – first, second, and third-year Ph.Ds doing this stuff, is just amazing. In my opinion, this is going to be the largest effect of the challenge.”

With the tunnel circuit completed, attention in the Subterranean Challenge moves to the urban circuit, scheduled for February 2020. The third circuit, exploring caves, is scheduled for August 2020, with a final event in August 2021.

Results from Tunnel Circuit:

Team Explorer finished with 25 points in the tunnel circuit. Image: Ray Linsenmayer

Teams in the Systems track will compete for up to $2 million in the Systems Final event, with additional prizes available for self-funded teams in each of the Systems Circuit events. Teams in the Virtual track will compete for up to $1.5 million in the Virtual Final event, with additional prizes for self-funded teams in each of the Virtual Circuit events.

  • Team Explorer, 25 points (DARPA-funded)
  • Team CoSTAR, 11 points (DARPA-funded)
  • Team CTU-CRAS, 10 points (Self-funded, winner of the $200,000 Tunnel Circuit prize)
  • Team MARBLE, 9 points (DARPA-funded)
  • Team CSIRO Data61, 7 points (DARPA-funded)
  • Team CERBERUS, 5 points (DARPA-funded)
  • Team NCTU, 2 points (self-funded)
  • Team Robotika, 2 points (self-funded)
  • Team CRETISE, 1 point (DARPA-funded)
  • Team PLUTO, 1 point (DARPA-funded)
  • Team Coordinated Robotics, 0 points (self-funded)