One night in mid-March last year, a woman named Elaine Herzberg was walking across a street in a dimly lit area of Tempe, an Arizona city. She was hit by an Uber SUV with an autonomous driving system (i.e. with a person inside to supervise the car's autonomous choices): the car overwhelmed her, along with the bicycle she was carrying by hand, killing her. The news hit newspapers around the world and Herzberg became the first person to be hit and killed by a self-driving car, leading to numerous doubts about the future envisioned by Silicon Valley companies in which vehicles will drive themselves. and we won't even have to care about them.
The Uber crash
Police explained at the time that the accident could hardly have been avoided even in the case of a human being driving: Herzberg would have started crossing suddenly, suddenly appearing in the middle of the dark road, in a point with no pedestrian crossing. More than a year later, however, the investigation into the Tempe accident has reached a first set of conclusions that are not very encouraging for Uber and other companies in the sector: Herzberg was run over due to software problems in the Volvo XC90 equipped with systems. for driving developed by Uber. According to the US National Transportation Safety Board (NTSB), the US government investigation agency that deals with transportation-related accidents, the system was unable to identify Herzberg as a pedestrian. For now, the NTSB has not indicated a certain cause for the incident, but will provide further details by the end of November.
The experts who conducted the investigation still believe that the Uber hadn't identified the bike as being at high risk of collision until moments before making contact with Herzberg, who was carrying it by hand. At that point the braking distance was not sufficient and consequently the accident occurred. The system was apparently unable to accurately detect an unexpected crossing of the road, in an area with no pedestrian crossing and where pedestrians were not allowed to pass.
Following a decision made at the beginning year, Uber cannot be indicted for Herzberg's death, but the driver who was in the SUV and who was supposed to intervene could still be indicted. From some footage taken by the SUV's internal cameras it seems that she was distracted, a few moments before the accident, and that she was not observing the road. Other clues indicate that he was probably watching a streaming show on his smartphone.
Self-driving
Uber is alone one of the Silicon Valley companies developing self-driving cars. Waymo, a subsidiary of Alphabet (the holding that has control of Google), has been working on self-driving cars for years and has started projects in some cities of the United States to test its systems. Ford and other auto companies are doing the same, with their own research initiatives or collaborations with Waymo.
Along with the shift to electric motors, self-driving cars may be the most important innovation in the history of the automotive industry and for this reason investments of several billion dollars have multiplied in recent years. Uber is particularly interested because it considers the transition to autonomous driving essential to make its economic model sustainable: the current one, based on drivers, is not sustainable and is causing huge losses for the company, which has recently gone public.
Problems and opportunities
Despite the announcements on the progress achieved in the last period, cars that drive themselves they continue to have some problems in managing all the situations, and unexpected events, that arise when moving on the street. They are quite reliable on the motorway, where there are fewer variables and the most complicated maneuvers are exits and lane crossings, while they struggle in city traffic where traffic lights, pedestrian crossings, intersections and the presence of pedestrians add several complications to the software that decide how to do it. moving the car.
Waymo, Uber and the others try to highlight the improvements in autonomous driving systems, while they prefer to put aside the doubts and problems still unresolved. Before the Arizona accident and other less serious ones, these companies tried to promote the safety of their cars as much as possible over traditional ones.
In 2013, for example, Anthony Levandowski – the the engineer behind the project that would become Waymo and later accused by Alphabet of having stolen some documents to gain advantage in setting up his own company for self-driving trucks – he told the New Yorker: “Every year we delay more people die ”from road accidents. Elon Musk, the CEO of Tesla and SpaceX, has often used the same rhetoric, going so far as to argue that those who write articles questioning the safety of autonomous driving “kill people”.
Net of positivism that typically pervades Silicon Valley, it is however true that in various circumstances the sensors mounted on self-driving cars are more precise than our senses, and above all they allow the vehicle to react to a danger with much lower reaction times than ours. In extreme cases, such as the Arizona accident, they fail reminding us that a future without investments or deaths on the road is still distant. According to experts, in the case of Tempe probably a human could not have done better, given the poor visibility and the fact that Herzberg was crossing where it was not allowed.
The safety systems that already exist
Developing automatic systems to stop a vehicle from running in an emergency is complicated, and the recent tests conducted on the automatic braking systems of some traditional automobiles indicate that there is still work to be done in the perspective of integrated solutions in self-driving vehicles.
The American Automobile Association (AAA), roughly the US equivalent of our ACI has tested traditional cars, but equipped with emergency braking and pedestrian detection systems, which should offer greater guarantees for drivers and for those who walk on the street, reducing investments. For their tests they used dummies, with a mechanism to make them move and cross a street while a driver drove the test cars towards them. The results were rather disappointing.
In 60 per cent of the cases the vehicles hit the dummies, even in optimal conditions of visibility and with a speed of around 30 kilometers per hour. Things went worse in tests with mannequins with dimensions comparable to those of children: investments occurred in 89 percent of cases. By simulating more difficult conditions, including those at night, the researchers obtained even worse results: none of the cars tested were able to detect the presence of a pedestrian on the road at night.
For the tests they are Four cars were used for sale starting this year in the United States and beyond: Model 3 (Tesla), Camry (Toyota), Accord (Honda) and Malibu (Chevy). The investment rate reached 100 percent in the scenario where the mannequin crossed the road immediately after a sharp bend, without the car sensors being able to detect its presence.
Some of the companies involved automakers admitted that their systems don't always work in all situations, due to some technical limitations, but still remembered that in different circumstances they can make a difference, helping to compensate for the slower reaction times of drivers . Governments and institutions are lobbying to incentivize auto manufacturers to incorporate emergency automatic braking systems into their vehicles. Last April, the European Parliament approved new rules to make them standard by 2022, while in the United States it was the manufacturers themselves who committed themselves to do the same. Meanwhile, several car models are already sold with automatic emergency braking.
Road deaths
According to the World Health Organization, every year over 270,000 pedestrians die from road accidents. Deaths from people walking on the street make up 22 per cent of all deaths from road accidents, with some countries reaching 60 per cent. Automatic emergency braking systems can help reduce the problem, but they are not enough if you do not work to improve road safety, ensure safe pedestrian crossings and areas where cars can only travel at low speeds.
Automatic emergency braking is considered one of the main innovations of recent years as regards active safety systems, even if, as we have seen, the specific ones for recognizing and avoiding pedestrians are not always effective. Some manufacturers have begun to supply these solutions as standard, while in the meantime they have dedicated themselves to the development of other safety systems such as: sensors to automatically maintain the gear in a lane, cruise control that automatically adjusts to limits and road traffic, parking autonomous.
Levels of autonomous driving and safety
The prototypes of self-driving cars already use all these solutions and they integrate other, even more sophisticated ones, to detect what is around them, distinguish fixed objects from moving ones, recognize horizontal and vertical signs, automatically follow paths without human intervention. The combination of all these systems offers greater guarantees regarding safety, but the level of reliability varies a lot and only in recent years with the improvement of algorithms and artificial intelligences there have been significant progress.
To sort out the various classifications of autonomous vehicles, in 2014 SAE International, an institution that deals with coordinating rules in the automotive sector, identified 6 different levels for automatic driving, based on how much a human being has to intervene to compensate for car choices: level 0 is that of most cars on the road, while level 5 is the one in which vehicles will have achieved full automation.
Although there are exceptions and the classification is often debated by individual manufacturers, we can say that today most self-driving cars are between levels 2 and 3.
A level Full 3 requires that, under normal conditions, the car handles all major aspects of driving, choosing direction and accelerating and braking. The presence of the driver is in any case necessary to resolve situations in which the car software demonstrates uncertainties or if conditions arise that make driving safely problematic. With the “Autopilot” option active, the Teslas currently on sale meet part of the requirements identified in level 3 and, on multiple occasions, have been shown to be able to avoid accidents with much faster response times than those of humans.
Overall, the risk of accidents and investments will tend to decrease as the last two levels of the classification are reached, but it will not be easy to get there. In fact, the number 4 provides that the car is able to handle any eventuality and unforeseen event, except for the extreme conditions determined for example by bad weather. Level 5 will be reached when cars can manage themselves completely, thus making it possible to move them without anyone on board to supervise their driving.
The current safety assessment of autonomous systems compared to those traditional is carried out on a statistical basis, calculating the number of fatal accidents per kilometers traveled. Traditional cars are involved in about 1.2 deaths per 160 million of kilometers; Waymo's cars had driven 8 million kilometers as of February 2018.
The only known fatal accident involving a Tier 3 car so far was last year's Arizona one. Five other accidents involved as many Tesla with active autopilot: in all accidents the driver died from his injuries. According to Tesla, the accidents would have occurred even in the absence of autopilot, due to the external causes that caused them. Tesla is the only car company to already sell an advanced automatic system, and as a result it is common for it to log more accidents, driving far more miles than companies like Waymo still have vehicles in testing.
To date, level 5 still seems to be a long way off, but according to the most optimistic, the objective could be reached in a relatively short time. A push forward was provided by the evolution of artificial intelligence systems, which have become more accessible thanks to the greater computing power of the processors and models optimized to manage them, even with limited resources. The large amount of data collected from prototypes already on the road for years, or from models already on sale such as Tesla, offers further opportunities to integrate the new systems and make them more “intelligent” and safer.
