Are robots successfully replacing human drivers?

DARPA's "Urban Challenge" competition has been held every year since 2004. In the competition held in December 2007, the robotic vehicles improved significantly, and even awarded their developers with prizes of millions of dollars

By: Israel Binyamini, "Galileo" magazine

Robotic cars move around the city, alongside cars driven by human drivers. They obey traffic laws and bring their cargo and passengers safely and comfortably to their destinations. This is a vision we have already seen in science fiction movies, but is it very far from the capabilities of the technology that exists today? It turns out not.

On November 3, 2007, the final stage of the "Urban Challenge" competition took place in Victorville, California. This competition (link at the end of the column) was conducted by DARPA, the Agency for Advanced Research Projects of the United States, and is the third in a series of robotic vehicle competitions. In the first competition, in 2004, no vehicle met the requirement to cover a route of about 400 km between Los Angeles and Las Vegas in less than ten hours, completely independently, without any remote control.

In the second competition, held in 2005, four vehicles (out of 23 that started the route) managed to finish the route in the required time. The route included roads, dirt paths, crossings through gates in fences, sharp turns, tunnels and other difficulties. Vehicles that deviated even slightly from the track could have encountered obstacles or sunk into impassable sand. The team that arrived in the shortest time was that of Stanford University, and it won a prize of two million dollars.

Encouraged by the fact that in the second competition all teams (except one) performed better than the best team in the first competition, DARPA decided to design a much more difficult challenge: driving in an urban area. The area was not a real city, but an area specially prepared for the competition inside an abandoned US Air Force base. Within this area, the vehicles had to complete a route of about ninety kilometers within six hours at the most, find parking spaces and park, avoid hitting the sides of the road or other vehicles, and obey California traffic laws.

Three teams managed to meet all the requirements, so the winner was determined by the time it took to complete the trip. The three robots arrived within a few minutes of each other, but since the cars started at different times, and there were other factors that affected the calculated time, the teams had to wait until the next day to receive the judges' summaries. The final result: the robot named Boss, from the Tartan Racing team (link at the end of the column), won first place and a prize of two million dollars.

Collision in slow motion

Not all robots were equally successful, and mistakes and careless driving were also recorded, such as the incident described below, which occurred a few hours after the start of the race: the Skynet robot (Skynet, as the name of the murderous computer network in the "Terminator" movie series) entered a traffic circle by turning a little too sharply , and stopped after almost hitting the guardrail on the side of the road. After a short wait, as if the robot had stopped to think, it moved back, stopped, moved forward again, and stopped again.

At that moment, another vehicle - from MIT (Massachusetts Institute of Technology) - arrived at the same traffic circle, noticed Skynet standing on the side of the road, and decided that it could be overtaken - just as Skynet decided to move again. The collision was at low speed and did not cause any damage.

The human supervisors "froze" the two robots using the remote controls included in each robot, then sent them to continue their tasks - first Skynet and then the MIT robot, which finally finished the competition in fourth place. There were also glitches in the autonomous driving systems: even Boss couldn't get off with the "opening shot" of the competition, when electronic interference from a huge TV screen blocked the reception of his GPS system.

His successes in the earlier stages of the competition gave the boss the right to start the race first, but the malfunction caused him to finish tenth, and he lost further time when he got "stuck" behind slower robots. It will be interesting to see if the next robots will learn to honk in such a situation.

Despite this and similar failures, DARPA director Dr. Anthony Tether noted that most of the robots worked and behaved correctly. This is an important achievement for DARPA, which aims to develop a military capability of carrying out "dangerous, dirty or boring" tasks by autonomous vehicles. An army equipped with such a capability will better protect its soldiers and will be able to use its resources more efficiently. In doing so, DARPA is acting according to an order of the US Congress from 2001, which required the US military to reach by 2015 a situation where a third of the active ground vehicles will be unmanned.

Such vehicles are also widely used in built-up areas, such as for delivery missions. As we know, autonomous vehicles also have great civilian potential. This situation is not foreign to DARPA, whose previous projects have already contributed to the creation of the Internet, modern operating systems and technologies for artificial intelligence such as speech understanding.

Anatomy of an autonomous vehicle

Bose's software is divided into three main areas: perception, planning and behavior. The purpose of the concept is to build a "situation picture" as detailed and accurate as possible describing the location, direction and speed of the robot, as well as similar information about the location of other objects (vehicles, fences, etc.) in relation to it. Some of these objects are "anticipated", that is, known in advance to the vehicle according to the city map loaded into its memory, and some of the objects are not expected: new obstacles and barriers on the road, parked vehicles, vehicles moving on the road, etc.

To collect information about the world, the Boss robot, like most of the other robots that participated in the competition, uses more than a dozen different sensors: lasers, radars and cameras, and of course GPS, which is used to locate the vehicle. Each of the sensors has a different "field of expertise". For example, some are used to identify nearby objects, and some are suitable for identifying more distant objects; But a combination of all the sensors makes it possible to cover mistakes and malfunctions, complete information and improve accuracy.

The challenge in the perception system is not limited to collecting information from the sensors: while the cameras, for example, can only report the intensity and color of the light at each of the millions of points in their field of vision, the purpose of this part of the software is to provide information such as "the fence on the right is 160 cm away from it the right side of the vehicle, approaching at a rate of about 10 cm per second; In front of us is a vehicle moving in the same route and direction, and we are approaching it at a relative speed of 10 km/h; At a distance of 130 meters in front of us there is an intersection with a road that connects to the left, and from this road a vehicle approaches the intersection at a speed of 25 km/h; This vehicle is currently 70 meters from the intersection."

To extract such information from the points of light and color reported by the camera, and from their combination with information from lasers, radars, etc., advanced software is needed, which combines artificial intelligence capabilities of image processing, pattern recognition and probabilistic inference to find the interpretation with the highest probability. These abilities have been improving rapidly in recent years, and are among the main reasons for the progress made since the first competition in 2004.

The planning system operates according to the mission definition: the destination the vehicle must reach and the action, such as parking, that must be performed upon reaching the destination. According to the current location and the location of the destination, it is possible to create an "action plan", which at a high level may be similar to the route planning that can be obtained from any navigation software ("go straight for 200 meters and then turn right; at the third intersection turn left and continue another 830 meters until reaching the destination").
Unlike the familiar navigation software, the robot's action plan also dictates additional details, such as speed for each section of travel. This speed is selected according to the driving laws of the state of California and the maximum speed allowed in each section, and according to the information available on the road and traffic conditions. At the lower level, each such operation has sub-operations required for its execution, such as remote identification of the junction where to turn, stopping at the stop line, etc.

The behavior system chooses actions at every moment according to a combination of the information from the perception system with the action plan. If you have to reach the intersection, stop (according to the traffic laws) and then turn right, the behavior system will decide when to slow down, check if there is a vehicle approaching the intersection, and decide if it is permissible and correct to enter the intersection. The software can also deal with unexpected behavior of another vehicle, such as not stopping at a stop sign or making a sudden turn. The group's engineers call this ability "defensive driving".

Robot, who gave you a license?

The Tartan Racing group, which built Robot Boss, was created as a collaboration between Carnegie Mellon University of Pittsburgh and the car manufacturer GMC (short for General Motors Corporation). The group also received support from other companies, including high-tech companies such as Google, Intel and HP, vehicle companies such as Caterpillar and the Israeli company Mobil-Eye, the leader in the field of driver assistance technologies based on computer vision. The other competing groups, at least two of which completed all the tasks almost as quickly, employed similar collaborations of academic organizations and commercial companies.

It is easy to see the great potential of the capabilities demonstrated by the vehicles that participated in the competition, and not only for the future battlefield. Software can react much faster than a human driver, who needs precious tenths of a second to recognize a dangerous situation and another tenth of a second to decide how to react, and to carry out the decision. Software also does not get tired or fall asleep, and can retrieve information about road conditions and road signs (even if they have fallen or been hidden) from the communication network. As you know, even if the computer is not reliable enough to drive itself, it is possible to start with computerized help for the driver. Such devices are already available today in certain models of new cars, or for installation in existing cars.

What else is missing to bring "automatic driver" to mass use? DARPA's competition is just one of a large variety of research and development around the world, whose goal is to improve the solutions that have already been presented, add capabilities and continue the experiments to find out if there are more challenges and problems that the experiments have not encountered so far.

Some are strongly opposed to the DARPA project and the academic collaboration with it, because they see a real danger in the development of driverless military vehicles. Among their claims: if the enemy cannot attack soldiers on the battlefield, it will be an incentive for him to attack civilians in terrorist ways; On the other hand, the ability to go into battle without endangering soldiers will allow leaders to decide too rashly on military measures, which will ultimately result in the loss of many lives on both sides.

Dr. Teter, the director of DARPA, disagrees. He is very satisfied with the results, and in his opinion the development is already close to the point where it will be transferred to other entities: "DARPA is an interesting organization. We never finish the job. What we are doing is showing that it is doable. We eliminate the technological excuse, to the point where it is no longer possible to say, 'This is a very interesting idea, but you know it is impossible to carry it out.'

I think we've come close to this point, and it's time for this technology to be promoted by someone else." It is not possible to conclude from these things that the day is near when we will all be driving cars that drive themselves, but in the competition described here it was proven that the technological limitations are no longer an insurmountable barrier. Therefore this day, even if it is not really close, is closer than we dared to imagine.

Israel Binyamini works at ClickSoftware developing advanced optimization methods. The article was published in Galileo magazine.