All cars have at least some automatic features
First up, there are various degrees of autonomy, ranging from 1 to 5. Practically every new car on the market today qualifies for Level 1: all it needs is adaptive cruise control (longitudinal assistance). Level 2 combines a lane-keeping system (lateral assistance) with longitudinal assistance. Level 3 is what most people have in mind when they think about autonomous cars: the car drives itself – but only in specific areas and circumstances (for example crawling along in heavy traffic). The driver, however, is still in charge and needs to be ready to take over at any time. In Level 4 autonomous vehicles, drivers are no longer in charge and the car does all the work – again, in certain conditions, which the industry calls ODDs for operational design domains (on a given motorway stretch, at nominal speeds and in specific weather conditions, for instance). The steering wheel, pedals and gear lever will only disappear when cars reach Level 5, i.e. drive themselves 100% of the time.
Forecasting is of course an intricate exercise and the overly bullish promises we have heard throughout the history of AI call for a spot of scepticism. Google announced that “ordinary people” would be buying autonomous vehicles by 2017(!). But there has been tangible progress notwithstanding the hype cycles: Waymo, Google’s company tasked with developing self-driving technology, recently logged 10 million miles on public roads in its test cars, amassing a wealth of new data and specific scenarios to feed its algorithms. The company has also introduced Waymo One, its autonomous taxi service, on the outskirts of Phoenix, Arizona. The forecasts vary somewhat from one player to another, but they all seem to agree that they will hit Level 4 at some point between 2022 and 2025, and won’t be reaching Level 5 before 2030. But they may follow different paths: legacy carmakers are planning to clear the levels in order whereas new players are aiming to skip straight to Level 5 with self-driving shuttle services (which, again, will operate in specific situations and at low speeds).
The obstacles ahead
Beyond the issues that may slow down IA development (for instance the likely demise of Moore’s Law, meaning that processing power will probably plateau), there are technical hurdles that autonomous vehicles still need to overcome. The first one is the amount of data available: IA software needs a huge amount of it to inventory all possible driving cases. And this data needs to be amassed around the world – driving conditions in Paris and Nevada, say, are quite different – then processed and analysed.
The second big issue relates to managing exceptions and ambiguity when making decisions. This is one of the problems that autonomous vehicle manufacturers are dealing with today: the fact that human and artificial drivers are sharing the wheel involves negotiating skills that are difficult to code. Driving onto a motorway from a slip road, for example, has more to do with compromise, rights of way and tact than with maths. Part of the solution will no doubt come from seamless real-time communication among the vehicles and with their surroundings (or V2X for vehicle-to-everything technology). That is at the core of augmented infrastructure development. The goal, here, is to enable infrastructure to gather data via roadside sensors and send it to vehicles, so as to complete and broaden their awareness of their situation and help them make smarter decisions about their position. The discussions about the communication standard to carry that data are ongoing, and the main question is whether it makes more sense to use Wi-Fi (which has shorter ranges but can be up and running sooner) or cellular telecommunications (future 5G networks, which will require more substantial development).
Cybersecurity and social acceptability – or, in other words, avoiding the disastrous consequences of autonomous vehicle hacking – are the last two thorny issues. Social acceptability relies on the perceived risk, which will only be low if security levels are high. This is already the case in other autonomous modes of transport (automatic metros and aircraft autopilots, for instance). Then, who will be accountable for an accident? Explainability is another crucial concern: it’s difficult to certify something that can’t always be fully explained. Discussions about future standards are trying to address this by certifying “Safety for the Intended Functionality” (SOTIF).
And the final question is whether legal frameworks will evolve as fast as technology. Article 43 of the French PACTE law (an action plan for business growth and transformation) is an important step in the right direction because it opens the door to testing autonomous cars on French roads. The country’s law on mobility, which is currently in the pipeline, should clarify the regulatory framework.