One afternoon in late September, a yellow four-legged robot named Spot prances and spins on a replica of a dirty subway platform built in a huge limestone cave below the Louisville Zoo. The spots spy around the platform, sucking in data through cameras and sensors arranged on the torso the size of a vacuum cleaner. The robot's small feet kept close to the edge of the platform dangerously, and then returned to a safe place. In the end, Spot was obviously satisfied with what it knew, and flexibly walked down the stairs to investigate further on the roadbed. Back on the now deserted platform, a poster on the wall declared: "The future is now."
What a wonderful future. In this carefully constructed scenario for the end of the DARPA Underground Challenge (a carefully designed three-year, $82 million Pentagon Robotics Competition), something bad has happened to the underground humans, and the robots are leading the way. Come to the rescue. Spot and its robotic teammates and competitors — dozens of walking, driving, and flying robots — are looking for "survivors" (human models that emit body temperature and sounds) and items such as cell phones, backpacks, and helmets. The robot scores points by sending the position of the object back to the human teammate. Finding all items means exploring a maze full of traps. It has a half-mile-long passage and has three environments from scratch: city, subway, storage room, and office; tunnel (simulated mine); and a cave, which is A claustrophobic mashup of cave exploration's hottest songs.
This game is a major test of the proposition that one day the robot team can help emergency personnel assess the disaster area before risking their lives. This also marks a bold step towards robot independence, because robots will have to work beyond human control to a large extent. Eight teams of more than 100 world-leading robotics experts tracked the action remotely (and somewhat helplessly) from the underground assembly area. They are not only attracted by the $3.5 million in prize money-$2 million for the first place, $1 million for the second place, and $500,000 for the third place-but also the potential for professional bragging.
All of this is a science fiction fantasy three years ago. At that time, the Defense Advanced Research Projects Agency-DARPA, the Pentagon department dedicated to advancing breakthrough technologies-invited leading robotics experts to participate in the competition. A few months later, they gathered at a gold mine near Denver to introduce the well-known SubT Challenge. Before that, the role of robots in disaster areas or bomb teams was limited, and manual controllers were usually required to guide individual robots with joysticks or tethers. Although robots are becoming more and more common in retail and manufacturing, they tend to execute pre-programmed programs or operate in primitive, structured environments. Rarely, if any, robotic teams are required to collaborate autonomously and explore rugged, unknown terrain. There will be no GPS in the SubT Challenge, so robots must make and share their own maps with each other. In addition, there will be unstable WiFi, so the robot must establish its own communication network. No single robot has all the required capabilities, so teams of different types of robots must have perception, movement, decision-making, and network capabilities that have never been deployed together in the real world.
"Three years ago, when DARPA started this project, we did not have the technology in this area," said Steven Willitz, a member of the Carnegie Mellon University team. "No one in the world can do such a thing."
The research problems that DARPA solves are so basic that their relevance often goes far beyond war. (In the years leading up to the coronavirus pandemic, DARPA has been funding research to accelerate vaccine and antibody production, including work that led to the Moderna vaccine.) DARPA adopted a competition model for awards that stimulated innovation in the early 2000s, a series of autonomous driving The major challenges of the automobile, these challenges are considered to promote the continuous development of the automotive industry. For the robotics competition, DARPA's three-year timetable can be said to force the emergence of technological solutions faster than the speed left to private companies.
[How a secret Pentagon agency lays the foundation for a rapid treatment of the coronavirus]
Preliminaries were held at a coal mine in Pittsburgh in 2019 and an unfinished nuclear power plant southwest of Seattle in 2020. Now, after more than four days in the 100-acre cave in Louisville, there has been an amazing breakthrough in the terrible malfunction. Self will be harmed, hardware will be destroyed, reputation will be damaged. In the end, the highest prize of $2 million is tied, and the tiebreaker rules must be invoked to determine the winner within a few seconds. Along the way, as the robots found their way and passed their findings back to their creators, they clarified a bigger question than their competitors: how far is the day when humans no longer need it?
The Louisville Giant Cave has thick rock walls, clay floors and approximately 30-foot ceilings. It was once designated to accommodate 50,000 people in the event of a nuclear war, and now features ziplines and holiday light shows. It took DARPA more than a month to transform it into a temporary underground civilization, partly a research laboratory and partly an elaborate robot standoff.
The agency installed a TV studio in one area, and Camryn Irwin, a sports host covering ESPN events, and Scott Walker, a golf channel veteran, broadcast daily live broadcasts on YouTube via "DARPAtv." Owen and Walker have completed their homework, working with guest experts to gain insight into the weeds of robot design and strategy, while at the same time using heavy metal soundtracks to amplify the drama as much as possible.
In the other part of the cave, DARPA laid the floor for the team's "garage", and fenced fences were installed on both sides of the spacious and bright limestone corridor. Here, dozens of sleep-deprived engineers tightened the last screw on the robot chassis, frantically typing the last few lines of code into the laptop. (In addition, nine teams, including four teams from on-site challenges, participated in a virtual-only competition.)
Throughout the summer, I called the team to learn about them and their robots. Now, during the week of the competition, I hang out between the garages, meet with robotics experts, and appreciate the various solutions they plan to propose in the challenge. I was shocked by the way competitors work almost side by side in a workspace that is open to everyone. At this point, they know more or less what other people have, just like athletes who study the highlights of their opponents. In the finals, this will boil down to execution, and whose method can best counter the surprise prepared by DARPA in the course.
Matthew Travers, a systems scientist at Carnegie Mellon University’s Robotics Institute, nodded to his competitor: “In the past three years, this community has consolidated.” “People are very much. Excited. This is a big thing we have been working on. Obviously we want to win, but it’s good to be able to participate, learn, and contribute."
Pittsburgh-based robotics giant Carnegie Mellon University is under pressure. CMU is SubT's favorite and won the 2019 preliminaries with a huge advantage-featuring the tunnel course. Its team is called Explorer, and it is allied with researchers at Oregon State University. Unlike most other teams buying existing robots and enhancing them with unique algorithms and unprecedented exploration capabilities, CMU’s engineers have built rugged tractor-type robots and some anti-collision drones from scratch, while other durable drones From the CMU spin-off company established during the challenge. Their plan is to plant a communication node dropped from a tractor in the maze, so that the robot can share information and transmit the location of the treasure hunt game object to score points. The drone will be launched from the rear of the tractor for deeper and higher searches. "When everything goes well," Traverse said, "it's actually very complicated."
Another most popular is CoSTAR. The team is led by NASA's Jet Propulsion Laboratory and works with large engineering companies such as California Institute of Technology, Massachusetts Institute of Technology and other partners. CoSTAR won the 2020 urban track and plans to deploy quadruped robots, wheeled robots and drones. "In order to complete the entire task... you really need to push the most advanced boundaries," Ali Agha, a robotics expert at JPL, told me before the game. He added that one day, the technology pioneered here can be used to explore below the surface of the planet, where it is easier to find evidence of life (if it exists) than on the outside. "There is no [perfect] solution today," he said, "and when we started three years ago, there were fewer solutions."
Just like athletes with disabilities in racetracks, DARPA has supported six of the eight teams, including CoSTAR and Explorer. Their method is very promising and is eligible to receive $1.5 million in phased funding from DARPA within three years, up to a maximum of $4.5 million per team. This money gave these teams a decisive advantage, but it was DARPA's way to reward the best ideas and give them the greatest opportunity to get it out of the laboratory and into the real world.
DARPA also supported an Australian laboratory team called CSIRO Data61-a collaboration between the Commonwealth Scientific and Industrial Research Organization, a Brisbane company called Emesent, and Georgia Institute of Technology-and a team called MARBLE, the team Composed of researchers from the University of Colorado. Partners such as Boulder and Denver. In the first two games, CSIRO and MARBLE have completed the midstream.
I should not be surprised that there are two teams from the Czech Republic. After all, the word "robot" was coined by Czech playwright Karel Capek in his 1920 play "RUR" (in the play, robots are exploited, so they resist. They killed Most humans, until they realize that they have killed people who only know how to make robots.) The CTU-CRAS-NORLAB team is a partner of the Czech Technical University in Prague and Laval University in Quebec City. It performed well in the early circuits. DARPA decided to fund it to enter the finals. Another Czech team with American and Swiss partners is called Robotika, which is self-funded.
Near the end of the garage, four red four-legged fire trucks curled up to rest. I can't help thinking of the robot dogs in "Fahrenheit 451", even though these creatures are defenseless. They are called ANYmals and are the main competitors of the yellow Spot walking robot. In the SubT Challenge, every team capable of buying Spots (Boston Dynamics' commercial version, each priced at about $75,000) did so and customized them to explore independently—except for one. The CERBERUS team developed its strategy around ANYmals manufactured by ANYbotics in Zurich. CERBERUS, also funded by DARPA, is a partner of the Research University of ETH Zurich, the Norwegian University of Science and Technology, Berkeley, Oxford, and the University of Nevada Reno. In Greek mythology, hellhound is the terrifying three-headed watchdog of the underworld. Kostas Alexis, professor of robotics at the University of Norway, said that the three leaders of the team are legged robots, flying robots and autonomous algorithms.
[Spot is a robot dog worth US$74,500 in our dystopian dream]
Every game requires brave losers, and I found them at the other end of the garage: the self-funded Coordinated Robotics, led by independent robotics expert Kevin Knoedler, and professors and undergraduates from California State University Channel Islands. Nordler is so skilled in simulated robotics that the team won hundreds of thousands of dollars on the virtual track of the SubT competition; he used prize money from past competitions to fund Coordinated Robotics. He searched for safety robots on Craigslist, and the team rebuilt the list for the game. Unable to afford the fancy communication nodes that fell from the robot, the team reused the Raspberry Pi computer-a credit card-sized device for beginners to learn programming-and designed a deployment mechanism that includes paper clips, kinks, and Shotgun pellets.
The team won $250,000 in the City Tour last year, making it the second-highest self-financing group. "We are a small and aggressive team, but this is also one of the joys of doing this," said Jason Isaacs, associate professor of computer science in the Channel Islands. "This shows my undergraduates... how big boys play in the world's top graduate school of robotics. And you can do it too."
Timothy Chung, DARPA Challenge Project Manager, wanders among the flickering and buzzing robots that avid humans are studying. He is a game master and carefully designed a huge laboratory experiment disguised as entertainment and sports. Before joining DARPA in 2016, Chung was a drone expert with faculty members of the Naval Graduate School. Under the system where DARPA only grants project managers a temporary term to avoid becoming obsolete, Zhong has a few years to make a technological sensation-he intends to make the SubT challenge one of his.
"When you know that there is innovation to be carried out, the huge challenge is really great. You know that there is a breakthrough waiting to happen, but it is still unclear which direction the breakthrough will take place," Zhong told me on the phone from Zoom when we arrived in Louis. Before Will. "We will build an arena and a sandbox, and you can come and propose your solution."
When we met in the cave, he pointed out the calibration method of the game to produce actual results and theoretical breakthroughs. Chung consulted with military commanders, the New York Fire Department and other first responders, as well as mine safety engineers and cave rescuers, and came up with the various obstacles and booby trap hazards that rescuers hope robots can investigate before human rescuers enter.
Although the Pentagon will certainly benefit from any discovery, there is a well-known element of idealism for this huge challenge. I want to know whether the $82 million that DARPA budgeted for this project is worth it, especially because the results will be open to the world and teams from other countries have already benefited from it.
"To really emphasize this desire to create a new community, perhaps a brand-new industry, of course a breakthrough technology, it really requires and guarantees the injection of a lot of resources," Zhong said. He also mentioned the history of DARPA, which was established after the Russians launched satellites in the late 1950s to ensure that the United States would not fall behind in technology again. "On the other hand, DARPA is not only interested in creating technological surprises, but also in preventing it," he explained. "This is a great opportunity to understand what we did well and what we did not do well."
Curriculum designer Viktor Orekhov led the press team to visit the labyrinth. I am grateful to him for letting us wear helmets, because when I stumbled through the cave part, my head kept hitting the synthetic stalactite. "This is a science, it's a big experiment," Orekhov said of his booby trap. "It also has an art... how do you design a course that is difficult enough for ground robots and aerial robots so that you can see what is possible and really push us to see the direction of technological development, And not so hard that no one exceeds the first 100 feet?"
The stadium stretches for a little more than half a mile and consists of three interconnected underground worlds-cities, tunnels and caves-built under the larger caverns. For most of the tour, we had to double it; once, we had to climb a steep hillside covered with artificial rocks by hand. Some parts include features designed to challenge the agility of the robot, such as water obstacles, muddy, loose gravel, stalactites, three stairs and two sets of railings. There is at least one fog machine to confuse the robot's sensors, and three heavy doors swing behind the robot, so they must find different ways to go back.
[How do you teach robots to navigate new places? Learn young children. ]
Forty survivor mannequins and other handicrafts, such as mobile phones, are scattered everywhere. In addition to dissipating heat, the survivors also said: "Welcome to the SubT Challenge Finals!" When the mobile phone plays DARPA video and sends out wireless signals, all of these can be detected by robots equipped with cameras and sensors to obtain heat. , Sound and signal. In the urban area, a simulated office is filled with carbon dioxide, which can be detected by gas sensors.
The team knows the types of artifacts, but does not know the layout of the route or the details of obstacles, so they cannot over-prepare their robots. The whole thing is already very difficult for humans to negotiate. What will the robot do?
The first two days of the competition included a short preliminaries to solve problems in the robot system. DARPA changes the details of the course every night, so the team never knows what will happen. Highlights of everything that went wrong are the main entertainment of the long night of fixing the last error.
I heard laughter from the Australian’s garage and went over to see what happened. In the storage room near the subway track in the city, they were looking at a laptop and watching footage of what happened during their second initial trial. A CSIRO drone took off from the back of the tank robot and quickly identified a backpack. So far so good. It then sucks a piece of plastic packaging material into the propeller. The drone flipped and fell like a rock. The robot expert laughed—it turns out that the drone only needs a new prop. "Unfortunately, I have never taught drones to avoid flying pieces of paper," said Glen Wagner, who taught drones almost everything else.
(Wrong plastic debris is not a trap set by Orekhov-it is real debris left over from the construction of the stadium. But because it interferes with the running of a team, to be fair, Orekhov Carefully replace it with future runs.)
In the CoSTAR garage, JPL's Agha was even more frustrated with what happened in the team's first round of preliminaries. The team's three robots-drones, walking Spot, and wheeled robots-met at an intersection. There may be an algorithmic argument, but the walking robot stepped on the drone. Then the tractor drove twice on the drone. This accident and other issues in this round of competition resulted in one of the biggest upsets in the history of robotics: Mighty JPL-Caltech-MIT tied for last-behind Coordinated Robotics and its Craigslist hardware and paper clips.
But after the second round of preliminaries, Aga and his teammates are beaming and fist each other. Their score is better than any other team, which is the result of the strategy changes and software fixes they encountered around 5 o'clock that morning.
[Tesla said it is building a "friendly" robot that will perform trivial tasks and will not fight back]
After the preliminary round, CSIRO topped the list, followed by Explorer, CERBERUS, CoSTAR and MARBLE. Preliminary scores have nothing to do with the bonus rounds on the third and final day of the competition. Each team made it to the finals in a clean and tidy manner, and the money on the table was there. They carried out the final adjustment with excitement and tension, and tried to sleep.
"Three years-they were all shortened to 60 minutes," CSIRO teammate and Emesent co-founder Farid Kendall told me.
"DARPA will throw everything they have to us," Explorer's Willits said. "We have done our best to prepare for this, and now we just need to do our best to implement our plan.... And the crazy thing is that (in the first few rounds) the last team can bring Go home 2 million dollars."
"The results will be great," CERBERUS's Alexis predicted. "I don't know if it means one, two or three.... Our goal is to be in the top three."
The members of CoSTAR gathered in a circle, taking turns recalling the final stage of the intense journey. "We just went to enjoy it and see what the robot did," Agha told his teammates. "It's not us anymore."
On the day of the game, the teams packed their robots and entered the maze one by one. They boarded the tram and transported the machines and personnel to the staging area at the entrance of the stadium. As each team passed by the hallway of the garage, their competitors stood up and applauded.
The staging area is a platform that provides space for team robots and a small human repair station staff. Only a few feet down the dusty ground can be seen through the entrance arch. Above is a scoreboard, used to record the time and the number of correctly identified objects. Each team has 60 minutes to find as many survivors and artifacts as possible, and accurately report their locations within five meters. Every correct report will get one point, and 40 points is a perfect score.
A teammate playing the role of "human supervisor" would sit at a table with a computer monitor. Under the command of the supervisor, the robot will walk, fly or roll through the arch and disappear. Gradually, the map of the unknown maze drawn by the robot will spread like a tree branch on the supervisor's computer screen. Workpiece reports are returned from the robot in the form of pictures or other data, and the supervisor’s job is to verify them and send them to DARPA for points. Too many wrong guesses will be punished.
The teams that fell behind in the preliminaries played in the morning but did not find enough objects to compete: CTU, Coordinated Robotics and Robotika scored 7, 2, and 2 points respectively.
The team MARBLE from Colorado has arrived in Louisville and plans to use the fewest robots in one of the purest autonomous integrations-only two wheeled vehicles and two Spots. When everything goes well during the run, Dan Riley, a PhD student as a human mentor, sometimes feels that he does not have enough to do. But at the last minute, the team made a decisive strategic decision to put more control over humans. They added two short lines of computer code. When the robot is within the communication range, Riley can receive the view image from the robot's perspective. This helped Riley recognize objects faster, and at one point let an unwilling robot pass through the fog. "Then I was able to let the autonomy take over again, it started and more artifacts were discovered," he told me. The team scored 18 points.
CoSTAR's robot ran into problems from the beginning. A Spot strode onto the court-then stopped and tried to return. The team is aware of the culprit: Orekhov’s fog machine is launching billowing clouds at the entrance, and some robot sensors interpret the fog as an insurmountable obstacle. The human supervisor adjusted his strategy, but it was too late. The robot explored less than half of the courses and scored 13 points.
I sat on the side of the docking area and watched the four red ANYmals of the Swiss-Norwegian alliance CERBERUS come to life. The spotlight comes on. The sensor rotates. The robot stood up from the crouch, walked to the entrance, and stood on the starting line like a sprinter. Under the order of the human supervisor Marco Tranzatto, they began to act. Two of the ANYmals carry black communication nodes behind them. They will stop regularly, squat down and put down a node, thereby expanding the communication network. The other two are equipped with additional sensors and can roam more widely. Every few minutes, the two roamers will return to the communication range and send back the artifact report. Tranzatto has the ability to tell the robot the general direction, but the robot will do the rest of the work autonomously. The team's drone failed to enter the stadium, and ANYmals explored less than three-quarters of the maze; despite this, they found a large number of artifacts in the area they did reach. Final score: 23.
Explorer's three heavy wheeled robots almost galloped down the city, tunnels and cave passages, and were hanged to death in several narrow places. The team's drone dived through the subway tunnel and explored the platform. One of the wheeled robots is the only machine in the entire competition that climbs steep slopes in the furthest part of the cave. Then it slid down a 15-foot cliff, lying still on its back, the wheels spinning. The crash occurred outside the communication range, which meant that another robot had to retrieve its artifact report to forward it back to the supervisor-but this was not the biggest problem. Explorer's energetic robot queue sent hundreds of subject reports to the supervisor, so that he could not scroll through them all before the time ran out. At least six unscored scores are left on the computer. In the end, the team lives up to its name, exploring 93% of the maze—more than any other team—but only scored 17 points.
The last game of the day was the Australian CSIRO. Soon, when the railway track in the tunnel part fell off one of the pedals, one of the two tank crawler robots got stuck. The drone it carries cannot be launched because the space is too narrow. "So we lost two agents," Queensland University of Technology PhD student Brendan Tidd told me later. Fortunately, the two Spots of the team seem to be exploring independently. But time passed, and one of the Spots failed to return to the communication range to upload its artifact report. Tidd thought it must have fallen into the depths of the maze. He was left with only the second tank trying to retrieve data from Spot. The two shortest roads were blocked by the crippled tank and Orekhov's trap door that closed the passage, so Ted had to make the tank go the long way. "I was flying along that cave path, and the communication nodes were flying on both sides of the robot," Tidd recalled. "I did it in the last five minutes....This is a very tight game." At the same time, another Spot automatically retrieved data from the fallen Spot-but then also fell. It is still able to transmit information to the tank. The exercise paid off. CSIRO created the most accurate maze map of any team-and scored 23 points.
[Researchers imagine that robotic lizards may play a role in future disaster monitoring]
That night, it was clear that MARBLE won the third prize of $500,000. But CERBERUS and CSIRO tied for 23 points each. Chung called the tiebreaker and clearly stated in the 36-page rule book of the game that the team that determined its last artifact first would win. Since the team only has a vague idea about their own points time-and does not know the time of their opponents-Zhong can keep the winner secret until a big reveal made for DARPAtv the next morning.
Robot experts gathered in a conference hall built by DARPA in the cave. The highlights of the Grand Prix were played on a huge screen, but the time of the last key artifact detection was jokingly concealed. Then Chung announced the result: Tidd's crazy sprint for CSIRO scored the last point in less than 30 seconds; Tranzatto reported the last artifact of Cerberus, with just over a minute left. The members of CERBERUS jumped up from their folding chairs and condensed into a tight bounce team like cheering athletes. They jumped onto the stage and accepted an oversized check for $2 million.
As a person, I found the results of the competition to be gratifying: the top three teams have found a way to use the cooperation between human intuition genius and the ruthless drive of robots. Teams that invest too much or too little in robots will fail.
The game master Zhong has foreseen this epiphany. In the past three years, by forcing the team to transcend the known limitations of robot autonomy, he has led them to re-recognize the role of humans. "When you can make a few robots do really meaningful and influential things, you can make human supervisors do other meaningful and influential things... It's really cool," he was talking to a few people. Said in the conversation. Contestants after winning. "This will be the key to providing powerful new features in the future."
Of course, the term "supervisor" has a very different meaning from "operator". The operator uses the joystick or keyboard to command the machine to dismantle bombs or assemble small parts. The supervisor is too busy to focus on the movement of a single robot, but to supervise the whole. Perhaps, one day, we will look back and realize that in the giant cave of Louisville in September 2021, our relationship with the robot has begun to transform from an operator to a supervisor.
Some teams told me that they have made some discoveries that have been commercialized or opened up new research fields in the past three years. Nevertheless, if it is not a robot, the highest score is only 23 points out of the possible 40 points, which is a bit shame for robot experts. This means that almost half of the survivors or artifacts were not found. "It is unclear whether we will remove humans from that picture," Zhong said earlier. "You will never have a'robot rescue team' because I think you still want humans to participate in many things."
[Why it takes years for a robot butler to take over your housework]
The winners told me that the money will be reinvested in the research of their respective institutions. The morning after the award ceremony, I returned to the cave and found that the members of CERBERUS had returned to work. Shehryar Khattak instilled map-drawing skills in ANYmals, and he followed the course with an open laptop to make another map.
He frankly admitted that CSIRO's mapping solution has proven to be excellent. The Explorer's ability to quickly disperse robots in the maze is "amazing." "However," he said, "this competition is not only about technology, but also about strategy. You have to balance all the components. Your human supervisor is as important as the onboard localization system, or maybe more important. This is the system, the algorithm, and the people. Harmony."
Khattak told me about the next problem he wanted to solve: robots navigating in nature. The distinction between shrubs and rocks is complicated. You can swipe one, while the other is difficult and to avoid. The dappled sunlight in the breeze through the tree-lined glades is completely chaotic for the robot.
When I saw the machine paralyzed by feeling overloaded, I remembered all my time in robotics research. Despite the greatest efforts of the world’s greatest roboticists, too much information—too many decisions, too much consciousness—can still cause drones to crash, and the mechanical quadruped is in endless confusion. Turn in a circle. I left the damp cave and breathed fresh air. It's time to go home. A journey of hundreds of miles requires my brain to initiate countless precise movements and make experienced judgments to travel through various complex environments. I would taste the barbecue that night, and the next morning I would go for a run in the sun-dappled clearing in the woods. I was suddenly and strongly aware of and grateful for all the little superpowers that made me human.
David Montgomery is a contributing writer for the magazine.