The author | Zhou Yongliang
Edited | Jingyu
When it comes to mobile robots, you may often encounter robots, such as in hotels, hospitals, catering and other service areas, you will often see the figure of robots. In addition, Boston Dynamics' influencer robots and various videos have also attracted a lot of attention on the Internet.
However, in some more complex and extreme environments, robots are still not popular, and they urgently need deep breakthroughs in technology.
In such a situation, the former unmanned driving has also encountered the same problem. From 2004 to 2007, the Defense Advanced Research Projects Agency (DARPA) sponsored the Driverless Challenge. In addition to the popularity of the lidar company Velodyne, the competition also directly spawned the hot self-driving track for more than a decade.
Following the same line of thinking, in 2018 DARPA organized the Subterranean Challenge (SubT) and set a $5 million prize to attract teams of engineers and scientists from institutions and companies around the world to participate in the competition.
Darpa SubT Underground Challenge Final Result | IEEE
Now, the series, which has lasted for nearly three years, has come to an end, and the CERBERUS team and the CSIRO Data61 team have won the first and second places in the finals, respectively.
In addition to the results of the competition, the greater value of this challenge lies in the mapping of reality. Looking back at the problems encountered in this competition and the solutions given, we can have a deeper understanding of the development trend of robots and accelerate the popularization of robots.
Robots' "Death Challenge"
Speaking about the direct opportunity to host the underground challenge, Timothy Chung, project manager at DARPA's Tactical Technical Office, said that this mainly takes into account the needs of combatants, as well as the needs of advance teams. They need search and rescue missions in a variety of underground environments, from densely populated accidents to natural disasters to mines.
At the same time, on the official website of this challenge, Clausewitz quoted a sentence in "War Theory", "The relationship between war and terrain requires us to have the ability to quickly and accurately obtain terrain information in any region."
Cerberus' team's ANYmal robot explores the cave autonomously | IEEE
In the selection of the competition venue, this underground challenge selects the three most challenging scenes of underground tunnels, urban underground and caves. In such an environment, robots will face many limits, such as the terrain is very complex, gps can not be used, communication is limited, visibility is low, mobility is also challenged, and so on. Simply put, DARPA's goal is to find robots that can still pass through extreme environments.
According to the organizers, they want the participating teams to push the boundaries of robotics in four areas: mobility (how to move quickly), perception (how to understand the outside world), networking (to transmit data back to the server), and autonomy (autonomous decision-making by machines).
The rules of the challenge are that teams search and test the venue to find and locate the exact location of 10 to 30 items set by the organizers, which may include mannequins, doors, electric pumps, valves, backpacks, fire extinguishers, radios or cell phones, and more abstract things like gas leaks. The team's final score depends on how many items they can find and how long it takes.
At the same time, the contestants were forbidden to enter the tunnel of the mine, and all the work had to be done by the participating robots. The marking accuracy is within 5 meters, which is considered to have completed the positioning and can obtain points.
It is still very difficult to achieve the success of the game. Because it involves tunnels, cities, caves, etc., a holistic solution is needed to achieve a relative balance of capabilities. As a simple example, a robot needs to be small enough to pass through narrow passages, but it needs to carry sensors and computers large enough to map, navigate, and make autonomous decisions for itself; in addition, in terms of power systems, robots need to be more energy efficient to travel several kilometers in extreme environments, but autonomous decision-making and mapping require large power requirements.
The official title for this race is "Underground Challenge", similar to the triathlon race, not to find the strongest swimmers, runners and cyclists, but to find the best overall results of the three. Many people find this statement too euphemistic and should be called the "Robot Death Challenge" or the "Robot Olympics".
Robots & Dungeons
To truly understand this challenge, you need to briefly review the whole process. It is understood that in August 2019, 11 robot teams from 8 countries gathered in a mine in Pittsburgh, USA, which was the venue of the first Tunnel Circuit competition. Along with the team, 20 drones, 64 ground robots and an autonomous airship called Duckiefloat appeared.
Participating teams in the SubT Underground Challenge | IEEE
The strategy of most teams is to first send robots to reconnoiter and probe the environment, and then select suitable robots to complete the positioning task according to the scene and robot performance. At the very beginning of the race, things were going well. But the accident soon happened, and communication became the biggest problem of the competition. Just as the robot turned the first corner of the mine, many teams lost contact with the robot because radio waves could not penetrate the hard rock.
Then, fast forward to six months later, the second race (Urban Circuit) came to the Sazorpe nuclear power plant near Seattle. The plant was built in 1977 and when it was 80 percent complete, the government decided not to invest any more money and the plant became an abandoned project. Today, it is the venue for the challenge.
If you say, in the tunnel challenge, communication is a huge challenge. In this urban underground track, communication is no longer a big problem. The participating teams upgraded and iterated to build a mesh network using deployable network nodes. As the robot moves forward, a small communication robot is placed every other distance, and then it will establish a network node on its own, so that it can work for a long time outside the communication range.
The communication problem was solved, but the low temperature became the problem. At that time, the temperature inside the abandoned nuclear power plant was -4 degrees Celsius, even lower than outside (2 degrees Celsius). Therefore, the participating team members and robots could not stay in the competition venue for too long, and the hands immediately moved to a warm place at the end of the competition.
In the end, CoSTAR relied on better performance and performance, ranking first with a total of 16 points on the City Underground Tour, 5 points higher than the second-ranked Explorer, and the CTU-CRAS-NORLAB team finished third.
DARPA SubT Underground Challenge Final Poster | IEEE
The third is the Cave Tour, which was supposed to take place in the fall of 2020 but was cancelled due to COVID-19. So, the event jumped straight to the final in September 2021 at the Louisville Mega Caverns, a very challenging venue. The entire cave is approximately 4 million square feet (about 370,000 square meters) and features a combination of three DARPA-designed scenes, irregular passages and large cave systems, as well as complex layouts of underground structures.
By this time, the number of participating teams had been reduced from the initial 11 to 8. For all the participating teams, autonomy was the biggest challenge in the final. Because team members are not allowed to enter the playing field, and only one person per team can remotely control the robot, letting the robot decide for itself where to go and how to go will be the key to winning the game.
At the end of the final, both CERBERUS and CSIRO Data61 teams found 23 items. With less time, the CERBERUS team won the final and took $2 million in prize money, while CSIRO Data61 finished second.
Revelation and future
Looking back at the three years of the Robot Underground Challenge, it should be three years when dreams shine into reality. Navinda Kottege, team leader at CSIRO Data61, said in an interview that the most valuable part of the challenge was that DARPA initially knew that robots did not have this ability, but asked for a competitive robot team after three years.
The original idea did become a reality. Before the challenge, their team did have some cool technology, but no robotic system could reliably work for an hour or more; but after three years, they could deploy robots, let them do some tasks autonomously, and then team members do the work in their own hands, which is a very big step forward.
At the same time, this challenge also brought a lot of technical surprises, which may lead the development of robots. In this series, the participating teams brought many forms of robots, including six-legged spider robots, four-wheeled models, crawler tank robots, flying robots, etc. Among them, wheeled robots offer the most reliable maneuverability, and drones can explore some of the larger caves.
But the most impressive thing is the CERBERUS team's four-legged robot ANYmal C, which is also reliable in the game, can maintain stability even in the event of a collision, and has done most of the team's work.
ANYmal C Quadruped Robot | IEEE
Built by the Swiss company ANYbotics, the robot is a bit like a large dog, weighing about 100 pounds, equipped with a camera, 3D sensors, including lidar for 3D mapping and synchronous positioning and mapping (SLAM), which can perform search and rescue, inspection and other tasks. In terms of movement speed, it can reach 1 m/s, and can easily handle 20 degree slopes and 45 degree stairs, as well as traverse 25 cm of voids through passages of only 60 cm.
What's more, ANYmal C doesn't require a painstaking modeling process and dangerous and costly field testing to cope with the complex topography of the real world. So how did it get there?
According to the team leader, in the case of complex terrain and the inability to accurately detect the environment, the robot must rely on proprioceptive - sensing its own body form at high temporal resolution.
To this end, the Swiss company ANYbotics has proposed a robust controller. This controller uses only proprioception measurements from a combined encoder and inertial measurement unit, which is the most durable and reliable sensor on a leg robot. With the new controller, these robots can easily climb over challenging scenes such as streams, meadows, snow, gravel slopes, and more.
At the same time, the controller is driven by a neural network strategy and trained in a simulated environment. Although there is no real-world data or accurate terrain model, the controller is still able to overcome a variety of irregular terrain in the field. The researchers also emphasized, "Our system can traverse almost all terrain without falling once."
In addition, the developers have introduced some new methods: first, in terms of models, ANYbotics does not use the current mainstream multilayer perceptron (MLP), but uses sequence models, especially the time convolution network (TCN) of the feeling state; second, about privileged learning. The new model is divided into two phases in training, first training the teacher strategy, which has access to privileged information, that is, the real knowledge of the ground and the robot's contact with it; then the teacher instructs the student controller to learn with pure proprioception, which uses only the sensor information available to the robot itself.
In addition, the methods mentioned in this study do not use camera, lidar, or contact sensor information, relying only on proprioceptive sensor signals to improve the adaptability and robustness of control strategies in different terrains.
The CSIRO team has mapped maps with extremely high accuracy| IEEE
In addition to the four-legged robot, the high-precision mapping map is also the surprise of this challenge. In this challenge, although the CSIRO data61 team ranked second, the map they mapped by the robot was less than 1% different from the real map of DARPA.
It is understood that darpa's official map is to invite professional technicians, use very expensive equipment, and spend 100 hours to conduct professional mapping. But the CSIRO data61 team, in less than an hour, got results that were close to the official map. The technology behind this really surprised and curious everyone.
It is understood that behind the CSIRO data61 team, there is a set of independent robot research and development platform called Wildcat and commercialization solutions for industrial applications. According to the team leader, Wildcat uses a nonlinear optimization algorithm to convert 3D LiDAR (light detection and ranging) data from robot sensors into a highly detailed and accurate environmental map in real time, while providing attitude information for robot control at an updated rate.
Robots can share map fragments in real time to help them stay aware of changes in their environment and explore new areas. This allows them to easily integrate different robotic platforms into their fleets.
Wildcat leverages CSIRO's more than a decade of SLAM R&D experience, which is used in conjunction with IMU (Inertial Motion Unit) and LiDAR sensors, with the option to add additional sensors such as chemical, radiological, gas, GPS, and Wi-Fi.
Wildcat solutions have already begun to be commercialized, with drone company Emesent being its first customer, while it also collaborates with several other companies.
A few years ago, Boston Dynamics uploaded a video of the robot on YouTube, which touted its bipedal robots and four-legged robot dogs. However, even if it becomes an "Internet celebrity", boston dynamics' fate in Google, SoftBank and Hyundai Motors reflects the difficulties of the robot industry.
Although the DARPA Robot Challenge tests the search and rescue ability of robots in extreme environments, just like the driverless competition 15 years ago, which gave birth to the autonomous driving track, the terrain scanning, mapping and movement capabilities exercised in the robot competition may soon be applied to the robot products visible on the market, greatly expanding the application scenarios of robots and ultimately accelerating the transition of robots to consumer electronics categories.