After a long wait, it seemed that 2023 could be a turning point year for autonomous cars. Last August 10, the California commission which is also responsible for issuing permits for public transport granted Waymo and Cruise (two of the main companies in the sector) the license to operate their robotaxis without human personnel in the city of San Francisco.
The Bay Area city would thus join Phoenix and Austin, where for some time these fully autonomous taxis have, in some neighborhoods, actually been in service (in all other cities are still in the testing phase). However, the hope that the moment of self-driving cars had finally come did not last very long: in the days immediately following the green light from the Californian commission, a Cruise car crashed into a truckwhile another remained stopped at an intersection for 20 minutes, causing a serious traffic jam.
Mobility What are the 5 levels of autonomous driving and what level are we at now by Emanuele Capone 12 May 2022
It was inevitable that something would go wrong, considering that from 2021 to today, over 600 events have been documented in which autonomous cars have committed various types of errors: from sudden and unmotivated braking in the middle of traffic to the most serious accidents involving pedestrians or workers working on the street. The final blow for Cruise came last October 17, when the national body responsible for road safety announced that the company, controlled by General Motors, was under investigation for risks posed to cars and pedestrians. A week later, the license allowing Cruise to operate in San Francisco was revoked.
What made the picture even worse was a recent article from The Interceptwhich revealed how Cruise was aware of the difficulties of his autonomous cars in recognizing large potholes in the middle of the road (for example those present in road construction sites) and even in correctly identifying the presence of children (which require precautions further, due to possible unpredictable behavior).
Again in internal documents, it reads like the difficulty recognizing children is linked to a “poor exposure to vulnerable road users and consequently to the few events from which to derive a risk estimate”. In short, and as we will see further on, the fact that robotaxis encounter a relatively low number of children prevents algoritmi di machine learning to learn to recognize them correctly (and therefore to behave with caution around them).
After the controversy that followed these revelations, General Motors has decided to suspend the entire national fleet of Cruise robotaxis. It is yet another setback in a sector that for years has been announcing the imminent arrival of autonomous cars (in the development of which 75 billion dollars invested), but which has so far struggled a lot to keep its promises. Worse still: according to a pattern published on X by computer programmer Filip Piekniewskiout of 11 companies active in the sector, 6 are already defunct (including Otto from Uber, Argo and Optimus Ride), 3 are dying (TuSimple, Nuro and Cruise) and only two are still in good financial health (Waymo and Aurora).
Own Waymo (owned by Alphabeti.e. Google) represents a case of particular interest: although it records losses amounting to hundreds of millions of euros, its robotaxis are the only ones to offer a commercial service (therefore not experimental, as happens elsewhere) to San Francisco, Austin e Phoenix.
Taking a look at the capital of Arizona it is also clear why this is one of the few cities in which autonomous cars are becoming a reality: the wide avenues, the streets that always run parallel or perpendicular and the limited traffic are all elements that make the work of self-driving cars much easier. They do the same weather conditions: in a desert city where the sun practically always shines (on average there are 17 days of rain a year), the sensors mounted on the robotaxis do not risk having visibility problems due to the rain, or being confused by the leaves carried by the wind or going haywire due to fog.
And elsewhere instead? To understand the enormous difficulties that one can encounter in a more complex city than Phoenix, just think that (according to the reconstruction of the computer scientist Gary Marcus) it is possible that multiple rules concerning each individual road section have been manually entered into the software of the Waymo autonomous cars active in San Francisco: “If this were true – explained Marcus – it means that what works on a certain street in San Francisco it may not work so well elsewhere.”
It’s a big deal, because it certifies the carenze del deep learning (which learns autonomously) in this area and because the manual insertion of the rules should be repeated for each city in which these robotaxis are to be put on the road, enormously complicating the work.
If you are familiar with what the streets of most large US cities are like, try to imagine what would happen if these autonomous cars landed on the streets of the center of Milanoin traffic Romain the rain of London or in the chaos of Istanbul. Will self-driving cars ever be able to deal with the very complicated European cities, if at the moment they can almost only get around in a city like Phoenix?
On the one hand, it might just be a matter of time: we start from the simplest tasks and then tackle the more difficult ones. On the other hand, the suspicion is that the obstacle could also prove insurmountable. Autonomous cars are equipped with a combination of refined sensors that create a detailed picture of their surroundings, thus identifying roads, cars, pedestrians, bicycles and obstacles and road signs and evaluating the correct behavior to follow thanks to deep learning algorithms.
Automotive Apple brings augmented reality to the car: the windshield of the future in a patent by Vincenzo Borgomeo 30 August 2023
By definition, these algorithms find correlations in a sea of datapredicting which action has the greatest probability of being correct (for example, stopping before the pedestrian crossing if there is a person about to cross it, because a different behavior would have a high probability of causing an accident).
Beyond do not guarantee the correctness of the action undertaken, the probabilistic calculation underlying the decisions of artificial intelligence must deal with the immense quantity of unknowns and variables that arise on urban roads. A number so high that it becomes very difficult for the software to identify a reference pattern. As Phil Koopman, professor of Engineering at Carnegie Mellon, explained, “if you are trained to only manage situations you have already faced, the problem is that there is an infinite number of situations that your car won’t be able to handle until it encounters them for the first time. As a result, deep learning is fundamentally unsuitable when human safety is at stake.”
And therefore, no autonomous cars? Certainly not for years to come, especially in cities that are less easy to navigate. In the meantime, however, driving assistance software will become capable of managing a greater number of extra-urban scenarios (for example, motorways in conditions of low traffic and good visibility) and of intervene when needed (if we have fallen asleep or risk an imminent collision).
Perhaps we have reversed the factors for a long time: we are not the ones who should leave the steering wheel to the algorithms (if anything, acting as supervisors), but it is the algorithms that must supervise us, intervening more and more effectively when something is going wrong. It may not be the dream of self-driving cars without even a steering wheel, but it’s not that small either.