Raphael Thys
Cornelius, a hog-sized robot with fat rubber tank treads, has come to a stop in a small, verdant courtyard on the Spanish revival campus of California State University, Channel Islands.
“It’s either autonomous or broken,” Kevin Knoedler says, squinting into the summer sun, his face obscured by a mask and a hat with ear flaps. Knoedler, who has been building robots for decades, knows that it can be hard to tell the difference between a machine that’s kaput and one that’s cogitating.
“Autonomous,” says Andrew Herdering, a fourth-year mechatronics engineering major.
Suddenly, Cornelius sparks to life. The robot charges toward a backpack lying on the ground about 15 feet away. But then, halfway into its journey, it gets marooned on a large rock.
“Oh, no!” a third-year named Sara Centeno cries.
“It saw the backpack, and the way it’s programmed at the moment, it just mindlessly drives toward it,” Herdering says.
With some difficulty, I gather that Cornelius is in “detection mode,” which obligates it to seek backpacks, obstacles be damned.
What might look like the quotidian work of robotics undergrads anywhere is actually the fevered run-up, by a team known as Coordinated Robotics, to a huge event in the world of autonomy—the final round of the Subterranean Challenge, hosted by the US government’s Defense Advanced Research Projects Agency, or Darpa. In September 2021—a few weeks from now—Cornelius and the 20-odd other robots in the Coordinated fleet will be trucked off to Kentucky’s Louisville Mega Cavern to compete.
Darpa has held public challenges like SubT since 2004. They’re meant to draw talent from beyond the hermetic world of military R&D and jump-start innovation on very hard problems—forecasting the spread of an infectious disease, say, or launching a satellite on short notice. In the first Darpa challenge, a Humvee called Sandstorm autonomously drove 7.4 miles in the Mojave Desert before overshooting a turn and getting stuck. In the follow-up challenge a year later, five teams finished the full 132-mile course. Yesterday’s self-driving Humvee is tomorrow’s driverless taxi.
The SubT Challenge, which kicked off in 2018 and will conclude in the Mega Cavern, forces both robot and roboticist to confront the forbidding set of hurdles that exist underground—poor visibility, bad connectivity, hidden topography. It consists of both a physical competition and a virtual one. In the final physical contest, robots will snake through claustrophobic passages, clamber up stairs, and struggle through mud and fog—maybe even mock avalanches—as they search a course in the Mega Cavern for “thermal mannequins” (i.e., humans) and other “artifacts.” In the virtual competition, simulated robots will do all the same things inside a computerized rendering of the Mega Cavern course. At stake is $5 million in prize money.
The premise of the virtual competitions is that anyone with enough smarts and access to a computer—say, the quiet guy in dad jeans who tells fellow soccer parents, when they ask, that he “does robotics stuff”—can contribute meaningfully to the research. Knoedler (pronounced “nayd-ler”) excels at these contests. Darpa’s program manager for the SubT Challenge, Timothy Chung, calls him “a phenomenal software developer,” “very disciplined and methodical and practical.” But when the code has to interact with the real world, things get complicated. Knoedler quips that “you can solve 90 percent of the problem in the simulation and the other 90 percent on the robots.”
In the run-up to the Subterranean Challenge, Kevin Knoedler practiced piloting drones in his backyard in Southern California.
The SubT Challenge has attracted giants in autonomous research, including well-funded engineers from Caltech, Carnegie Mellon, MIT, and NASA’s Jet Propulsion Laboratory. They have top-of-the-line equipment, ample test facilities, and an army of graduate students they can throw at whatever problems meatspace may present. Next to them, Coordinated Robotics is the scrappy upstart. One robot in the team’s fleet is an old security crawler snagged off Craigslist. Another was assembled from plywood and hoverboard wheels by the robotics club at the middle school Knoedler’s kids attend. Where many competitors use surveying equipment that’s accurate to within a thousandth of a degree to orient their robots on startup, Coordinated often relies on a plumb bob (cost: a few bucks).
“You can’t do robotics seriously without simulation. You will never be able to test everything.”
When I meet the team on campus, the mood is frantic. They have just returned to the lab after many months in Covid lockdown. Their robots employ a vast array of software systems, which have to be painstakingly integrated with the equally vast array of sensors. “All of them have 20 different versions, which work with 20 different other versions,” Knoedler says. This is the annoying other-90-percent stuff.
Herdering is writing code to get the depth-sensing camera to display its data to the remote robot operator (i.e., Knoedler). Centeno is feeding the robots images of backpacks and ropes the way they’ll look in Darpa’s cave—spotlighted in the dark. “For some reason, if a rope is hanging vertically it detects it every time,” Herdering says. “But if it’s horizontal, like lying on the ground, it doesn’t.”
A bunch of DIY bots with surplus-grade sensors from a school that hardly anyone’s heard of with a team made up of undergraduates, their professor, and a stay-at-home dad—none of this seems like how you’d expect the skunkworks of the world’s most powerful national defense agency to revolutionize autonomy. But we live in a world in which the US military must guard against the threat of “irregular forces” flying weaponized off-the-shelf drones. Just as threats can come from motivated small-time actors, maybe solutions can too.
Sara Centeno, a third-year student at CSU Channel Islands, solders a communications beacon, which the robots use to relay messages underground.
The first robot that Knoedler remembers making an impression on him, as a 7-year-old in Colorado, was Big Trak, a six-wheeled programmable toy tank. A TV ad touted its 16 different commands, which allowed it to “get out of a tricky spot,” “complete the mission,” and “head for home base.” Knoedler became fascinated with the idea of getting a machine to do something as efficiently and reliably as possible.
That fascination stayed with him into adulthood. After attending MIT, where he studied computer engineering, he went to work for Teradyne, a company that develops automatic test equipment. Starting in the mid-1990s, with some coworkers, he began dabbling in TV contests like Robot Wars and BattleBots, sending nasty-looking spike-tipped bots with names like Monster into arenas from Long Beach to Las Vegas. It was more spectacle than serious robotics. “You were allowed to use programming,” he says. “But pretty much nobody did—it was all remote control.”
A few years later, Knoedler heard about the first Darpa challenge, the one in the Mojave Desert. He contacted a number of teams, looking to join as a free agent, and ended up on one called TerraHawk. Knoedler mostly worked on path-planning software—converting 2D lidar signals into a 3D terrain map, then plotting a route with a shortest-path algorithm. The team qualified to compete, but the night before the contest, the air compressor powering the pneumatic steering died. No steering, no race—TerraHawk was out. For Darpa’s next challenge, Knoedler joined a different team, which managed some 16 miles of autonomous driving in the desert before a failed USB hub stopped the run.
In 2007, Knoedler left Teradyne to become a full-time parent. “Kids are challenging, but it was a good choice,” he says, with characteristic brevity. His schedule loosened when the children started school, and he was soon drawn into more contests. In 2017, NASA held the Space Robotics Challenge, a virtual contest that offered a $125,000 prize to the team that could most successfully program a humanoid R5 robot to “resolve the aftermath of a dust storm that damaged a Martian habitat.” Knoedler decided to enter as a solo competitor. He didn’t necessarily want to work alone, but he still had a full slate as a dad—volunteering with the middle school robotics team and coaching various soccer teams and a local chapter of Odyssey of the Mind, a problem-solving competition. “I just didn’t have the time to coordinate with others,” he says.
Knoedler works on a robot called Joe.
The challenge took place in a simulation engine designed by Open Robotics, a nonprofit based in California. The organization is best known for creating the Robot Operating System, ROS, which has become widespread in the world of autonomy, especially for jobs in big facilities. Not counting Amazon’s warehouses, says Brian Gerkey, the CEO and cofounder of Open Robotics, “pretty much any other robot you see roaming around in an environment like that is probably running ROS.” Julia Badger, an autonomous-systems lead at NASA, says that ROS helps “to get people up and going quickly.” Now it’s not as hard to make a robot’s brain talk to its body. “It used to be that we had to write our own middleware all the time,” she says. “Now there’s a package for everything.”
While simulation might seem a poor substitute for real-world robotics, where fog can scatter your lidar beam and mud and rocks can de-track your crawler, Gerkey argues that it’s essential. “You can’t do robotics seriously without simulation,” he says. “You will never be able to test everything thoroughly in the physical environment.” In virtual space, you can unspool myriad hypotheticals at next to no cost: What if I deploy 100 robots instead of 10? What if I make the environment a dozen times bigger? How will my robot react to falling down a hill?
Knoedler’s simulated robot completed all the Martian challenges flawlessly. He won, collecting a total of $175,000. After the competition was over, he traveled to the New England Robotics Validation and Experimentation Center, where he transferred the code for his R5 to an actual R5. “We got it running basically on the first day,” Knoedler recalls. “Normally that transition to get things running on the real hardware can take a month or more.”
Flush with the NASA winnings, Knoedler began readying for the next Darpa challenge, SubT. It seemed like a natural fit, he says: “I like robots. I like caves.” The first phase took place in a research mine in Pittsburgh. Once again, he entered as a solo competitor.
A lot went wrong. Knoedler crashed all of his aerial drones and finished dead last in the real-world portion of the contest, taking home one of Darpa’s “superlative” awards: Most Robots per Person. But the virtual competition was another story. Knoedler dominated it. He finished first, with more than twice as many points as the nearest competitor, notching $250,000. This he would use as seed money.
Knoedler had wanted a team from the start. For the next stage, to be held around six months later, he knew it would be essential. Up the road from his house, at CSU Channel Islands, an associate professor of computer science named Jason Isaacs was looking for a way into the SubT Challenge. The cost of fielding enough robots was prohibitive, Isaacs says, and “as a small school with no PhD program, there was little chance of winning grant dollars.” When Knoedler reached out to him, offering to team up, it was an obvious match.
The new team quickly proved its mettle. At the second SubT event, held in a never-completed nuclear plant near Seattle, Coordinated Robotics placed second among the self-funded teams in the real-world competition and first in the virtual, taking home a total of $500,000. “The last round, we went in with the goal of scoring one point, so we were ecstatic,” Isaacs says.
The next physical competition, scheduled for the fall of 2020, was canceled due to Covid, but the virtual contest went ahead. It was another first-place win for Coordinated Robotics.
Jason Isaacs, an associate professor of computer science at CSU Channel Islands, works on a mannequin-hunting robot.
The Mega Cavern, originally the home of Louisville Crushed Stone, is a sprawling 100 acres of passageways and vaulted spaces running beneath the Louisville Zoo and all 10 lanes of Interstate 264. One senses that its present-day owners are still searching for a business model. There’s office space under construction, and people also stash boats and cars there. There’s a zip line (the only “fully” underground one in the world, apparently) and bike tours, and around Christmas locals drive their cars through the tunnels to see lighted displays.
This week, a micro portion of the Mega Cavern has been set aside for Darpa. The members of Coordinated Robotics are in their team “garage,” really just an enclosure among seven others. They’re preparing for their final trial run before the competition, and the mood is tense. They joke that they haven’t seen actual daylight in days. Their robots got waylaid in Tennessee and barely made it in time. When I ask whether Darpa has provided a team café, someone looks at a jar of peanut butter on a folding table and says, “That’s our team café.” I see a robot that’s new to me, named Karen. “Is this the robot that asks to speak to the manager?” I ask, attempting to lighten the mood. Herdering stares at me blankly.
Knoedler tinkers with a robot called Karen in the robotics lab at California State University, Channel Islands.
With a day to go before the competition, the team garages hum with activity. The high-pitched buzz of flying drones echoes off the walls. Creepy spider bots and quadrupeds with cheetah-print paint jobs scurry across the floor. As safari-park trams shuttle teams to the start gates for their runs, rivals pause briefly to applaud. The air is dank (“pretty much every screw is rusting,” Herdering tells me), and the smell of overloaded chemical toilets seeps through our pandemic face masks.
As I visit the various teams, I find that many of them are talking about the same things. They’re saying that legged robots, rather than wheeled or tracked ones, are obviously the superior choice, because of the terrain. (Flying drones may seem like the answer, but they sometimes have to deal with what Chung calls “the wall-suck problem”—the strange aerodynamics that can happen when a drone flies too close to a wall. One team tried giving its drones whiskers to help them get through tight spots.)
The teams are also saying that none of this would have been possible even 10 years ago. To take one example: At the first Grand Challenge, back in 2004, the lidar sensors were mostly single-beam, Knoedler says—the grainy film to today’s 4K video. Now multi-beam arrays are basically standard in the Mega Cavern garages. And who supplies Coordinated’s multi-beam lidar? A company called Velodyne, which spent the 1980s and ’90s manufacturing audio equipment before a Darpa-inspired expansion into self-driving tech.
Like the robots, I feel I’m moving around in the dark, never seeing everything at once.
For all the progress, though, there are still any number of ways for things to go wrong. Julia Badger, the NASA autonomist, enumerates some of them for me. A typical robot competing in the challenge might have several motors and motor controllers, a communication system to sync them, numerous gears connected to the motors, myriad sensors, and the software packages to power them, plus the AI to decide whether that black spot is solid ground or a precipitous drop-off. Errors abound: A drone flying near some shelving sucks up a sheaf of papers and crashes. A crawler runs aground on a railway track. “I mean, just getting your webcam to work on your computer sometimes is a bear, right?” Badger says.
On the first day, everyone’s talking about a disastrous practice run by Team CoSTAR. At an intersection somewhere inside the course, one of their drones fell and was promptly run over by a ground robot. (When I catch up with Ali Agha, a leading CoSTAR member and JPL scientist, he says that actually “a bunch of them ran over the drone.”)
Later that day, I rejoin Coordinated Robotics. They’re putting robots through their paces in a test area, netted to keep errant drones from flying away. The team seems a little frazzled. “Don’t step on that cable!” Knoedler snaps as I nearly tread upon a robot’s fragile communication tether. A last-minute change to try to improve the robots’ imprecise sense of where they are was not particularly successful. The lockdown-induced lack of testing is catching up with them. I ask Knoedler how optimistic he’s feeling about tomorrow. “Given today,” he says wearily, “not so much.”
The next morning I am ushered, with a few others, into the media viewing area, a small, curtained-off section in a cavernous hall, where we watch grainy feeds from the robots already on course. I’m able to parse some of what’s going on. The quadrupeds from Carnegie Mellon that look like horses pooping on the trail? They’re dropping communications nodes, creating a kind of mesh network through the labyrinth. The rest of it is a mystery. That robot paused at an intersection—is it stuck in an endless decision loop, or is it a comms relay for other robots? Like the robots, I feel I’m moving around in the dark, never seeing everything at once.
When Coordinated’s run is over, the team is awarded two points. Both Knoedler and I know that’s nowhere close to the podium. “The team actually got all the robots up and working, so that was really impressive,” he says. They were plagued by navigation issues. “Our robots couldn’t localize themselves very well very deep into the course,” Isaacs tells me. “So even though we found things, we weren’t able to report them accurately enough to score points.”
I ask Knoedler why the robots that did so well in Seattle came up short today. “We’re not totally sure,” he says. One theory: The bumpier floor in today’s course threw off the localization. Isaacs is upbeat about the whole thing. “This is the kind of learning that we cannot possibly duplicate in the classroom,” he says. “Kevin’s been a great mentor.”
On the final morning, after all the teams have completed their runs, everyone is ushered into a large, vaulted space with a big stage at one end, flanked by video monitors, under a giant Darpa banner, to watch a highlight reel. We sit in darkness on socially distanced folding chairs, watching robots move through caves, and I feel like I’m at some conceptual-art Kraftwerk concert. Camryn Irwin, who normally calls things like beach volleyball for networks like ESPN, hosts the proceedings, and Julia Badger plays John Madden with the expert commentary. I let the familiar cadences and clichés of televised sports wash over me. The emcees, to their credit, manage to bring substantial drama. “I was holding my breath watching this,” Irwin says, as a legged bot prances right to the edge of a subway platform.
And then, suddenly, we’re alerted that we can visit the actual course. We’re given hard hats, flashlights, and 30 minutes to explore. As we pass through the mine entrance, we enter Darpa’s faked cave, constructed out of a snaking maze of prefabricated metal pods. The composite walls of the narrow passage are fashioned after image scans of actual cave walls. Stalactites obscure the way. There’s a thunk, and Herdering says, “I just discovered why they gave us helmets.”
Knoedler in the Louisville Mega Cavern at the SubTerranean Challenge final.
As we walk, the team members call out the artifacts the robots saw, or didn’t see. Wedged into one alcove is a thermal mannequin that says, repeatedly, “Welcome to the SubT Challenge Final Event.” The underground network is impressively detailed. In the mine office, a faded “Employee of the Month” calendar is affixed to the wall. The subway station has graffiti, faded posters, even signs announcing planned closures—someone at Darpa was clearly having fun. I feel a new sense of respect for the robots sent into this foreboding, perplexing, obstacle-filled fun house.
Coordinated Robotics finishes next to last in the real-world competition but third in the virtual—good for another $250,000. The overall winner of the virtual competition is a solo competitor, Hilario Tomé, a Barcelona-based roboticist. Tomé, who wasn’t at Louisville, tells me later that his success came in part from sheer effort—120 hours a week for nearly a year and a half—and in part from going beyond the test worlds that Darpa provided to help people prepare for the competition. His simulated robots represented a “truly generalizable solution,” he says.
“There’s always a temptation to ‘study to the test,’” Chung says. But overfitting solutions to known problems leaves robots unable to handle the unknown, which was “the crux of the problem that Darpa was interested in.” Tomé plans for his SubT work, and the funding from it, to power his new company’s forthcoming physical robot. He has already announced a pilot project with Barcelona’s fire department.
The fact that Knoedler was usurped by this relative upstart is, for Chung, less a reflection on his ability than a sign of how quickly innovation is happening in the sphere. Being an incumbent “is not as long-lasting” as it used to be, he says. What’s simulated today is real tomorrow.
This story was updated to clarify that Coordinated Robotics placed second among the self-funded teams in the Urban Circuit’s real-world competition.
Let us know what you think about this article. Submit a letter to the editor at mail@wired.com.