The automation of vehicles is today one of the most discussed issues of modern mobility. Tesla, who seems to be the global market leader in the automation of cars is a brand mark of which everybody seems to be thrilled or at least many people who are ready to wait long months for their own Tesla car.
Still, watching demonstration films on Youtube, we yet may have some doubts about whether fully automated cars could manage on public roads while participating in the same traffic flows that vehicles led by a human hand. (Photo by Roberto Nickson on Pexels.com).
Traveling the world, we already see successfully implemented automated vehicles on railway tracks. A prominent example is the Paris metro line 1. But automatic trains on dedicated tracks often fenced and hence not accessible for pedestrians, and other traffic participants are in any way nothing beyond fully programmable machines riding according to a defined schedule on a single track at a time slot dedicated to them. Their movement is just programmable. No artificial intelligence needed. According to present technical standards – a piece of cake.
But automated cars on public roads are another story.
THE LEVELS OF VEHICLE AUTOMATION
Intelligent to fully automated vehicles are at least theoretically classified into five classes (levels). Level 1 is cars driven by a driver but containing some driver assistance features. Level 2 is partial automation with some steering automation, but still, it is the driver who drives the car and monitors the environment.
Level 3 is conditional automation. The driver may let the car drive and does not monitor the surrounding environment, but on the vehicle’s notice, he or she must be ready to take the wheel. Level 4 is a highly automated vehicle that can autonomically ride under specified circumstances. Still, there are circumstances under which the vehicle is not able to drive automatically.
Only level 5 is full automation under all circumstances. The vehicle can perform all driving functions under all conditions. At levels 4 and 5, the driver has an option to take over the wheel.
Source: US Department of Transportation, National Highway Safety Administration
By scrutinizing the above features, we can quickly get to the conclusion that what differentiates the vehicle automation levels is automated driving under only specified or all conditions. Under some conditions, it is better when the driver takes over the wheel.
THE PROBLEM OF COMPUTING POWER
The basic explanation is that the driving conditions may be so complex that the data flow needed by a machine to be processed in a split of a second is still beyond the presently available computing power. The complexity of traffic conditions concerns in particular road transport vehicles. As I write this text, Microsoft specialists assess the data flow of an intelligent car (level 1) as 25 GB per hour. An automated vehicle needs at least 30 GB processing capacity per second.
While watching films of automated driving uploaded by Tesla, we can clearly observe the vehicle computer tracing (in bounding boxes) all fixed and moving objects in the vehicle’s environment. But, the more complex the situation, the car gets slower in decision making. Is the complexity of traffic conditions only a problem of the computing power, or are there some other reasons that the on-board computer cannot still manage the trip on its own?
THE THEORETICAL BACKGROUND
The issue of too complicated driving conditions is already well explained in the traffic engineering and transport economic theory, at least from the fifties of the XX century. The so-called multiple interactions of traffic participants make the fundamental diagram of traffic (functional dependency between traffic density and the traffic flow) developed by engineers only a theoretical function. Traffic density is the number of vehicles on a given road section. Traffic flow is the number of vehicles that can trespass a cross-section of a road in a given time. Theoretically, the maximum traffic flow should correspond to the designed road capacity. But in practice, it is never so. The diagram is, as said above, only theoretical.
Further explanation was given at the beginning of the XXI century by physicists who, upon the molecular physics experiments, explained the so-called breakdown of traffic conditions. The purpose of the experiments was to explain the nature of spontaneous traffic jamming. But this theory may also be useful when talking automated vehicles and the failure to be autonomic under all traffic conditions. They concluded that due to traffic instabilities, the maximum capacity is never reached. The theoretical traffic function gets scattered by higher densities.
THE PUBLIC GOODS
The economics theory divides all the goods we consume into the so-called private goods and public goods. (There are some more detailed divisions like club goods, but for the purpose of this text I will omit them as non-relevant). Private goods and public goods are not about the ownership of goods. They are about the nature of usage. A good that is consumed at one time by one user is a private good. A good that may be simultaneously consumed by many consumers is a public good with the growing interference of each additional user imposed onto all other users.
An all-access road is a public good. It can be used by anybody who holds a driver’s license and by any vehicle that is allowed to the road system. So, on public roads, we deal with a variety of vehicles of different lengths and acceleration parameters and speeds. And the drivers are different people, with differing driving skills, habits, and tempers. Not all of them obey the rules. And I do not talk here only of the speed limits. Drinking and driving with all consequences for the other drivers and pedestrians might be just an example. And, besides direct traffic participants, a pedestrian can any time emerge on a road or a street. Pedestrians also not always obey the rules or simply forget them while holding smartphones in their hands. If we add e-scooters and bikes on roads and streets, the unpredictability of traffic conditions gets higher and higher. And all of the conventional users may use the road at the very same moment as an automated vehicle is on automation modus on the same road …
JUST PROGRAMMING IS NOT ENOUGH FOR UNPREDICTABLE TRAFFIC
This simultaneous usage of a road or a street by all possible users makes the road a very unpredictable environment. The more interactions among users are possible, the more decisions in a split of the second have to be done by a driver. In the unpredictable traffic environment, the choices and the driving itself do not correspond to any algorithms. They are getting intuitive. It is more about the psychology of traffic than the well-defined driving rules. As the traffic flow is based on the individual and often intuitive decisions of single drivers, it often gets unstable. No traffic engineering functions apply. The density and flow relation gives back scattered results. The traffic is no longer explainable and hence programmable by using mathematical functions.
An automated vehicle is a machine, a computer. A machine can either be programmed by applying functions and algorithms, or it may learn to correctly respond only after the so-called deep learning process. To cope with the problem of traffic instabilities, not only more processing power to deal with algorithms would be needed, but also some significant work on the artificial intelligence of automated vehicles. Deep learning of the unpredictable traffic conditions would, in turn, require photographs or short films of thousands, but thousands of unusual traffic situations intelligently described and further matched with the algorithms on how to react to each of those traffic situations.
The exercise is similar to the process of teaching the machines to recognize a cat and a dog in a picture. But the automatic recognition of dogs and cats by machines was possible only after thousands but thousands of photographs at the disposal of AI laboratories had been shown to them. All described either #dog or #cat. The photos with hashtags had been gathered through crowdsourcing (photos with hashtags downloaded by social media users). The deep learning process required, in turn, applying artificial neural networks. The significant difference between deep learning about traffic conditions and learning to recognize dogs and cats is that crowdsourcing of specimens is rather not possible for traffic conditions. The data must be gathered by vehicles at the disposal of those who teach the machines. This requires mileage and time. Much more time than was needed to teach the machines to recognize dogs and cats.
THE AUTOMATION IS EASIER WITH PREDICTABLE TRAFFIC
But not all traffic conditions are not predictable. Under very light traffic conditions, we are in the so-called free-flow traffic. Our interactions with other road users are only a few. We have free choice of speed within the legal speed limit. The overtaking is easy. Tesla simulations show free flow is quite well manageable by the present-day automated cars.
The denser the traffic gets, however, we have more and more to adjust to other users that share the public road with us. Gradually, with more users on the public road, the interactions of road users are multiplying. The traffic gets synchronized. The speed must be adjusted to others. Overtaking requires greater and greater attention. With increasing traffic densities, we are stuck in traffic that moves at the same pace for a lane. The denser it gets, the pace on all lanes gest synchronized. If traffic flow is close to the road designed capacity, the speed of all vehicles stabilizes. Overtaking is possible only by forcing out. Somebody once even said that traffic jamming is the most democratic institution in the world. No matter what car you drive, how expensive it was, you are equal to all other road users. You must behave like all the others, and you have no choice. If traffic flow is higher than the road capacity, the traffic is of a stop-and-go nature, and the traffic jam gets longer and longer.
So, in the free-flow traffic and from highly synchronized traffic onwards, the traffic conditions are very predictable. And this is when potentially automatic driving, even with current automation levels, would be useful … And it already is.
The driving under highly synchronized traffic and stop-and-go traffic is about keeping the proper distance and consuming as least energy as it is possible. Tesla states quite clearly today, ‘The Autopilot enables your car to steer, accelerate, and brake automatically within its lane.’ And I bet, with the current state of machine learning, a machine would optimize the ride within a lane better than a human driver, even if the surrounding traffic is with humans behind the wheel.
Personally, I would give much today if a computer could take over for me in a traffic jam. It would be a relief. But still, in the free flow and lightly synchronized traffic, I would be happy to drive on my own. And the reason is not only that the automation is too risky. It is because I simply like to drive. And, according to various studies I know about, many drivers would at least potentially buy an automated car, but only provided that they would be able to take the wheel to have the fun of driving.
THE PREDICTABLE HIGHWAY TRAFFIC
If we match carefully all we know about the reliability of automatic cars of today with the engineering and physics theory of traffic, we can quickly understand why automatic vehicles of today are best in highway traffic. Automated driving on highways today is about the automation of the driving process from highway on-ramp to off-ramp, including navigating on interchanges and overtaking slower cars if the traffic between lanes allows speed and overtaking like in the free-flow conditions. That what happens on a single lane is usually synchronized flow and, therefore, highly predictable. And there are no pedestrians, no e-scooters and no bicycles on highways, which lowers quite much the traffic risk.
THE PREDICTIONS FOR CARGO LOGISTICS
If we look into the automation of vehicles predictions for the near future, we see hopes for quick automation for the highway heavy vehicle traffic. Already the leading truck manufacturers like Mercedes Benz let their new truck models collect traffic data. Billions of miles of traffic experience collected by heavy vehicles throughout hours of daily trips should enable artificial intelligence to recognize and react to various and complex traffic situations on highways.
Mc Kinsey experts say that already from 2022 on, it will be possible to drive trucks in platoons with only a driver in the leading truck on the whole highway section of an origin-destination single trip, and later from 2025 with no driver in any of trucks in the platoon. On the last mile from the trip origin to the dedicated truck stop before highway ramp and behind the highway ramp to the tip destinations, drivers will have, however, to appear in the cabin. The highway traffic on the outside lane, usually taken by the heavy trucks, takes place under predictable traffic conditions. But the trip to the highway and from the highway section onwards in on public roads with general access and hence changing traffic conditions in a daily pattern from safe free flow off-peak, through unstable traffic conditions to predictable synchronized or stop-and-go traffic in peak hours. The driver’s hand will be needed for sure.
The trucks’ origins and destinations are usually located near highways. These are often big logistics terminals that consolidate cargo flows. The pickups and final deliveries are often made by smaller vehicles that enter conventional roads in denser inhabited areas. If the latter condition is met, the driverless’ highway logistics from- and-to big logistics terminals located near highways does not seem impossible by 2030 or even earlier.