The truth about driverless cars

They are on the way, but they are not what we were promised

By Steve A. Schlodover, The article is published with the approval of Scientific American Israel and the Ort Israel network 18.08.2016

• The auto industry and the press created exaggerated expectations for the arrival of automatic cars. Simple road encounters are a huge challenge for computers, so in the coming decades we are not expected to see robotic drivers.
• Automated vehicle transportation systems that depend on human support are particularly problematic. And yet, in the next decade we will still see automatic systems that will drive vehicles under certain conditions and for specific purposes.
• Automated parking lots, shuttles traveling at low speed within campuses, dense convoys of heavy trucks and automated systems for traveling on fast roads and using special lanes - all these are possible and perhaps inevitable.

Soon electronic drivers will drive us where we want to go, when we want, and in complete safety - as long as we don't have to turn left while crossing busy roads. Changes to the surface of the road will also pose problems. So are snow and ice. It will be very important not to harm traffic policemen, security guards and emergency services vehicles. And in an urban environment, where pedestrians are expected to jump in front of the wheels, it might be best to walk or use public transportation.

All these simple encounters that human drivers deal with every day are huge problems for computers. Solving these problems will require time, money and efforts. Even so, a significant part of the public is convinced that fully automatic vehicles are already waiting for us around the corner in the near future.

How did this expectation gap come about? Part of the problem lies in the terminology. The popular media indiscriminately uses phrases like "autonomous", "driverless", and "self-driving" even though these are very different technologies. As a result, important distinctions blur. The auto industry itself does not help to clarify the situation either. Marketers working for vehicle manufacturers, equipment suppliers, and technology companies carefully craft the advertising material they market to allow readers a wide range of interpretations about the degree of automation their products offer drivers. Journalists covering the field have an incentive to adopt the most optimistic forecasts: they are simply the most exciting forecasts. The result is a feedback loop that increasingly reinforces unrealistic forecasts.

This confusion is unfortunate, because automatic driving is a reality that is getting closer, and it can save lives, reduce air pollution and save fuel. But it will not happen as we are told it will happen.

What is automatic driving?

Driving is a much more complex activity than most people think. It involves a long series of skills and operations, some of which are easier to automate than others. Maintaining a constant speed on an open road is a simple task to implement, which is why conventional cruise control systems have been installed in cars for decades. With the advancement of technology, engineers have been able to achieve automation of additional subtasks involved in driving. Today, many cars are equipped with adaptive cruise control systems that maintain both the correct speed and the distance behind other vehicles. Lane keeping systems, such as those installed in new Mercedes-Benz and Infiniti models, use cameras, sensors and steering control mechanisms to keep the vehicle in the center of its lane. Today's cars are very sophisticated, yet moving from such systems to fully automatic driving would be a huge leap.

A scale of concepts with five levels, defined by the World Organization of Automotive Engineers (sae), can clarify the issue of automatic travel. In the three lower levels of the automation scale (level zero, which indicates the absence of automation, is not included in the scale) there are technologies that depend on humans as a backup for emergencies. Adaptive cruise control systems, lane keeping systems and other similar systems belong to the first class. Systems from the second level achieve automation of more complex driving tasks by combining the technological capabilities from the first level (such as lateral and longitudinal control that allows cars to maintain their travel lane and adaptive cruise control systems). This is the maximum level of automation found in vehicles marketed today. Systems belonging to the third level will allow drivers to transfer the vehicles to automatic driving in certain situations, for example traffic jams on fast roads.

The next two steps in the ladder are very different from the lower steps because in these steps the vehicle functions without any human assistance. Systems from the fourth level (high automation) will handle all the secondary tasks of driving, but only in very specific scenarios: in closed parking lots, for example, or in the lanes assigned to such vehicles on fast roads. At the top of the scale is the fifth step - fully automatic travel. One has to assume that such cars are what many envision when they hear things like those uttered by Nissan CEO Carlos Ghosn who confidently announced that by 2020 automated vehicles will be driving on the roads.

The truth is that no one expects that by this time vehicles will be marketed that will be equipped with automatic systems from the fifth level. Most likely, such a degree of automation still belongs to the distant future. Even systems of the third order may be equally distant. But systems from the fourth degree? You can expect them already in the next decade. To understand this confusing state of affairs we must talk about software.

Software nightmares

Despite the popular impression, human drivers have an excellent ability to avoid serious traffic accidents. According to statistics from 2011, fatal traffic accidents occurred in the US once for every 3.3 million driving hours. Accidents resulting in injuries occurred about once every 64,000 driving hours. These numbers set an important safety goal for automated driving systems. Their level of safety must be, at the very least, equal to that of human drivers. Reaching such a level of safety will require much more development than the enthusiastic followers of automation want to admit.

Think how many times your computer crashes. If the same software was responsible for driving a car, the "blue screen of death" would become more than just a phrase. Even a delay of a tenth of a second in the software's response can be dangerous when driving between other vehicles. The software responsible for automatic driving must therefore meet much stricter standards than those currently accepted in the consumer goods market.

Meeting these standards is an extremely difficult task that will require fundamental breakthroughs in software engineering and signal processing. Engineers need new software development methods that will prove to be reliable and secure even in complex and rapidly changing conditions. Although there are formal methods for analyzing every possible failure in some code even before it is written - you can think of them like mathematical proofs in computer programs - but they are only suitable for very simple applications. Today, scientists are just beginning to think about how to develop analytical tools suitable for handling complex code such as that needed to control fully automated vehicles.

After writing the software code, the software engineers will need new methods for detecting errors in the code and testing it. The methods that work today are cumbersome and too expensive. To get an idea of ​​the scale, consider that half the cost of developing a new military or commercial aircraft is the cost of software analysis and quality control. And the truth is that the software in airplanes is much less complicated than would be necessary in vehicles designed for automatic driving on roads. An engineer can design an automated flight system knowing that only rarely, if ever, will the system have to deal with more than one or two aircraft in its vicinity. The software does not need to know exactly the speed and position of these planes because the distances between the planes leave it more than enough reaction time because the decisions have to be made within periods of tens of seconds. A vehicle driving automatically on a road, on the other hand, will have to follow dozens of other vehicles, locate various obstacles on the way, and make decisions within fractions of a second. The level of complexity of the code needed for this will be tens of times higher than that needed in the software for flying an airplane.

After checking the code, the manufacturers will have to find ways to "prove" the safety of a fully automatic travel system in front of various parties: those in charge of risk management in the company, insurance companies, road safety activists and associations, regulators, and of course also potential consumers. The formal safety tests accepted today are not suitable at all for this purpose. Testers will need to conduct test drives of hundreds of millions of miles, if not billions, to ensure a statistically significant exposure rate of the vehicle to the dangerous scenarios it may encounter in normal use by thousands of customers. Some have started thinking about solutions to this problem. The government and industry in Germany have initiated a multi-million dollar project that is exactly its goal, but these are only initial efforts.

The code that will control the vehicle, its "brain", is not the only thing that will have to pass tests. The sensors that will provide this "brain" with the data according to which it will have to make the decisions will have to pass no less stringent tests. The engineers will have to develop new algorithms for processing sensor data and integrating data from different sources to determine whether objects in the vehicle's path are harmless or dangerous objects. These systems will have to ensure an almost zero rate of not identifying dangerous objects and misclassifying them as non-dangerous, and a very low rate of identifying non-dangerous objects as dangerous (such events may lead to unwanted reactions of the vehicles such as sudden deviation from the route or strong braking).

In commercial aircraft systems, the solution is redundancy - many backup systems. This path is not open to automotive engineers because an automatic car is a consumer product whose price must be within the reach of a large audience. Relying on artificial intelligence (AI) is not necessarily the solution. Some argue that machine learning could allow automated driving systems to learn from data that will be collected over millions of hours of driving, and continue to learn throughout the duration of their operation. But machine learning involves its own problems in that it is non-deterministic: two vehicles can come off the same production line but after a year of encounters in different traffic situations, the behavior of the two systems will be very different from each other.

The future of the fourth degree

In the past I said that fully automated travel systems, from the fifth degree, would not be a possibility before 2040. At some point I was quoted as saying that in 2040 automation from the fifth degree would become a reality on the ground. Now I am saying that fully automated vehicles, capable of driving in any conditions, will not hit the road before 2075. Is it possible that it will happen sooner after all? Sure, but not long before.

The future of automation from the third degree is also shrouded in fog due to the very real difficulty of capturing in an emergency situation the attention of drivers engrossed in observing the passing scenery, or worse - drivers who have fallen asleep. I heard from representatives of several car manufacturers that this is such a difficult problem that they do not intend to even try to develop automation from the third degree. Automation from the third degree may never materialize, except perhaps for technology that can control driving in traffic jams, where stopping and moving are slow and the maximum damage may be a fender bender.

Still, we will probably see cars with a high level of automation already in the next decade. Almost every major car manufacturer and many companies in the field of information technologies are allocating serious resources to automation from the fourth degree: fully automatic driving that is limited to defined environments and that does not depend on the backup of humans who may make mistakes. When the operation of automatic driving systems is limited to defined situations, their feasibility increases wonders. (Automated human transportation vehicles have been operating at major airports for years - but only on completely separate tracks.)

There is a great chance that in the next 10 years we will see complete automation of operations such as automatic parking in parking lots. Drivers will be able to leave their car in a special parking lot that will not allow entry to pedestrians and non-automatic vehicles. An automatic system in the vehicles will communicate with sensors that will be deployed in the parking lot to find available parking spaces and navigate to them. Since there will be no need to open the doors, parking spaces will be able to be narrower than today and thus parking lots in areas where parking is expensive will be able to accommodate more vehicles.

In pedestrian-only urban areas, business parks, university campuses and other places where speeding is prohibited, low-speed driverless passenger shuttles will be able to operate. In such environments, sensors with lesser capabilities will suffice to detect pedestrians and cyclists, and if a sensor makes a mistake in detection and causes unnecessary braking, no one will be hurt (at most the incident will anger the vehicle passengers). project CityMobil2 of the European Commission has already been operating such technologies for several years, and the final demonstration of its capabilities is planned to be held in the summer of 2016.

Segregated bus lanes and truck-only lanes will soon allow commercial vehicles to operate with higher degrees of automation. Physically separating these vehicles from other road users would greatly simplify the systems responsible for detecting and responding to hazards. Eventually, driverless trucks and buses will be able to follow human-operated vehicles in fuel-efficient convoys. Experimental models of systems for transporting convoys of buses and trucks are already being tested by researchers from all over the world, among other things in the program PATH at the University of California at Berkeley, in the Energy Project ITS of Japan, and in projects CONVOY and-Sartre in Europe.

But probably the most common application of fourth degree automation in the next decade will be automatic systems for personal passenger cars on high-speed roads. These systems will allow cars to drive themselves in certain road conditions and on defined sections of expressways. These vehicles will have redundancy of components and subsystems to allow them to "limp" to safety without human guidance in the event of a malfunction. It is likely that their use will be limited to good weather conditions and road sections that will undergo detailed mapping down to the level of the locations of the signs and lane markings. These road sections may even include safe stopping sites for vehicles to reach in the event of a breakdown. Most major car manufacturers are already working on the development of such systems, and next year the Volvo company is planning an open field test of such capabilities in Gothenburg, Sweden, a test that will include 100 experimental vehicles.

These scenarios may not sound as futuristic and exciting as a personal electronic driver, but they do have one advantage: they are possible, perhaps even inevitable, and they will arrive in the near future.

about the writer

Steve A. Schlodover - A mechanical engineer by training and holds bachelor's, master's and master's degrees from the Massachusetts Institute of Technology (MIT). In the 80s he helped establish the PATH program at the Institute for Transportation Studies at the University of California, Berkeley.

for further reading

• Technical Challenges for Fully Automated Driving Systems. Steven Shladover. Presented at the 21st World Congress on Intelligent Transport Systems, Detroit, Mich., September 7–11, 2014
• Towards Road Transport Automation: Opportunities in Public. Private Collaboration. Summary of the Third EU-US Transportation Research Symposium, Washington, DC, April 14–15, 2015. National Academies of Sciences, Engineering, and Medicine, 2015

• Summary of definitions for SAE International's 2014 report Taxonomy and Definitions for Terms Related to On-Road Motor Vehicle Automated Driving Systems
• On the way to cars that don't collide, Steven Ashley, Scientific American Israel, March 2009