Five years ago you couldn’t move for it, or at least its promise. Autonomy (aka the self-driving car) was going to take over, save lives, save the planet, make cars uncrashable, put drivers out of work and change everything we knew about individual mobility. Audi staged a piloted drive in one of its large saloons; all 550 miles from Silicon Valley to Las Vegas. Surely this technology was just around the corner, wasn’t it? This year the autonomy buzz at the Consumer Electronics Show (CES) in Vegas was less certain; that’s the way with consumer electronics, it changes with the wind. “The personal and mass-mobility bubble, with or without flying saucers, has been well and truly popped,” says Oxford-based autonomy expert StreetDrone in a recent report, describing the hype as “the Jetson phenomenon” and suggesting that even the autonomous technology community admits “that maybe ambition might be outstripping capability”. With, according to StreetDrone, Waymo (formerly Google’s self-driving car project) scaling back the scope of its Phoenix launch program and Uber’s progress only “creeping forward”, there looks to be at least a new-found realism in the business. “Yes, it was over-hyped then,” says Alejandro Vukotich, senior vice-president of autonomous driving at BMW. “We only learned how complex it was to do once we’d started doing it.” Mercedes boss Ola Källenius agrees, pointing out that the 2014 hype “has been replaced by a sober realism. “Producing a full-scale autonomous robo [robot] taxi,” he says, “working in a complex and crowded urban environment, is the most significant technical challenge we face if we want to do it safely with 99.9 recurring per cent proof against accidents.” Level Three Autonomy: the next big thing? Despite what some companies might tell you, there hasn’t yet been a SAE Level Three (L3) autonomous car sold to the public, although we are expecting this year’s new Mercedes S-class flagship to introduce L3. “We’re working on it,” says Källenius gnomically. But BMW’s Vukotich questions just what kind of autonomy customers would be offered with the first L3 cars. “Early L3 could be sold to customers as a feature, a Highway Pilot for example,” he says. “So the car will offer, ‘Hey I can drive for you’, but if it meets something it doesn’t understand, hasn’t seen or can’t do, it will hand it all back to the driver; sleeping [at the wheel] will not be an option.” Gill Pratt, chief executive of the Toyota Research Institute, explains that what we might see is a qualified form of L3. He explains that L3 is “a mode where the car can drive on its own and there’s a guarantee that it will give adequate warning to the driver before a hand-back is going to be required - so the driver can take their eyes off the road... [But] doing L3 might be just as hard as doing L4.” This “hand-back” process is when, within five to 15 seconds, a driver has to wake up, understand what’s going on and retake control. “I think that’s quite a long time,” Pratt says, pointing out that on a motorway you’ll have travelled a long way in that time, while in a city a lot can happen in a very short period. Which is why L3, when it is offered, might only operate in a highly prescribed set of conditions such as motorways, with no vulnerable road users along with well-defined lanes, as well as low speeds. “All you want to do is not to hit the car in front,” says Pratt. “So can you push a button and have the car take over the drive while you’re going at five to 10mph - the answer is probably yes. And I think that not only Toyota, but most car manufacturers, will come out with L3 systems for traffic-jam assist on the highway within the next few years.” Scientists and engineers call it the Operational Design Domain (ODD) and that’s the key here; the ODD for these early L3 systems is highways and slow speed. “It’s the same thing for L4,” says Pratt, “probably at low speeds in the city [but] my guess is also at high speed on the highway with trucks, those are likely to be the two kinds of beachheads where L4 will start.” But does the public really want to go further than just a stop/start device for traffic jams? The wind of opinion appears to be on the change. In March 2018, for example, a Deloitte-sponsored survey found the public’s hostility to self-driving cars was “diminishing”, yet by November last year a comparethemarket.com-sponsored survey found that 80 per cent of the public have “safety concerns” over self-driving cars, citing “cyber hacking and crashes”. To be fair, no one knew they wanted a Biro, a mobile phone or even an automobile until they became available. Glen De Vos, chief executive of Aptiv, the auto parts and autonomy specialist formerly known as Delphi Automotive, says: “As consumers come to experience and rely on advanced active safety technologies, we believe their trust and openness for autonomous vehicles will grow. “Trust will also come from the industry aligning on an answer to the question, ‘How safe is safe?’ ” Crashing and hacking: how safe is safe? After an Uber self-driving test vehicle struck and killed a woman on a pedestrian crossing in Arizona in March 2018, the industry has been highly sensitive to autonomy and safety; “Robot Car Killers” is the headline no car maker wants to see. “How can we prove that a [self-driving] car is as safe as a normal driver?” asks BMW’s Vukotich. “This needs discussion and regulation in society in general, but however you strive to make things safe, things happen…” Part of the problem, as Pratt explains, “is that the numbers are against us. There are many, many cars in the world and they travel trillions of miles each year. When you multiply even a very small chance of a crash, at such huge numbers, you end up with a significant number of crashes where there’s no human driver to blame “Now maybe it’ll also be the fault of other people that helped create the accident situation to begin with,” he says, “or some other factor such as inadequate maintenance, but the part that will be missing is our natural inclination, I feel, to empathise with a human being that made a mistake.” It’s hard to feel sorry for a machine and, even if it is statistically safer, we’re almost hard wired to trust human judgement over that of the robot. Pratt explains that it’s partly how autonomous systems work; they just don’t see things the way we do. “What goes on inside of the technology stack in the automated vehicle is perception, prediction and then planning,” he says. “At the perception stage, it figures out what’s out there and that’s at human levels of perception now, actually a bit beyond, because the sensors are in better places and they never get tired. The last stage, planning, is very much the same way that the computers are very good at playing chess, they can figure out a reaction plan to what’s going on the road at superhuman levels of sophistication, so that’s also good. “It’s the middle where the difficulty lies,” he says. “How do you predict what human beings are going to do? And one of the factors to think about is that these failures are likely to be weird because machines understand the world differently than we do... it’s just pattern matching, not actually thinking.” As far as the hacking of self-driving cars is concerned, Olivia Rudgard’s recent article posits the notion that cars could be taken over externally, though Pratt isn’t quite as concerned. “I think that this is a serious kind of issue,” he says, “and it’s particularly serious when a intelligent vehicle is trying to take advantage of communication with vehicle-to-infrastructure, or vehicle-to-vehicle communication, but I think there are technical answers to that with various types of cryptography, trusted keys and so on, where we can ensure that our vehicles aren’t spoofed by inaccurate information.” There are aspects of vehicle security that concern him, more, however. “What we can’t guarantee is that our vehicles will not be susceptible to what’s called a denial of service attack. And that means that the vehicles have to be safe, even if they’re totally cut off from any communication with the outside world; it doesn’t have to be a malicious thing, it can happen in natural disasters also.” An Uber self-driving car drives down 5th Street on March 28, 2017 in San Francisco, California Uber"s self-driving trial was back on the road in the US only a week after one of its test vehicles struck and killed a woman on a pedestrian crossing in Arizona in March 2018 CREDIT: Justin Sullivan/Getty Images As far as insurance is concerned, the UK trials are covered by the Automated and Electric Vehicle Act of 2018, which means that drivers of autonomous vehicles are required to have third party insurance cover, which should cover self-driving collisions, although even I can think of some holes in that legislation. Operators don’t require a permit, or surety bonds in the UK, but a driver, in or out of the vehicle, has to be ready and willing to resume control. There is a disconcerting feeling that the UK might have biased its legislation more in favour of autonomous research than it has road-user protection. It’s a feeling further confirmed by the fact that the Law Commission is in the middle of a three-year review of the legal framework as it applies to automated vehicles. In a statement to The Telegraph about this article, George Freeman, the Future of Transport minister, said: “Intelligent vehicle technology is a fast-growing global sector with big opportunities for UK industry, road safety and traffic management. Through our world class science and research, global role in setting standards and in vehicle insurance and regulation, the UK can play a big role in setting the global rule book for the AV sector. That’s why we’re investing more than £250m in trials and safe testing zones.” Court papers filed in its legal battle with Uber showed that Google had spent more than £850 million on its self-driving project between 2009 and 2015, Uber has spent a similar amount between 2016 and 2018, which is about a third of its entire R&D budget, and last year it secured another £766 million in investment for the project. Self-driving is very far from a cheap technology... Autonomous cars, robo taxis or trucks: which comes first and where’s the profit? Which is why, while most early projections were that autonomous passenger cars would be introduced before commercial vehicles, the reverse might yet turn out to be true. “Mass deployment [of autonomous cars] is a big financial commitment,” says Mercedes’ Källenius. “That’s why we’ve invested in truck hub-to-hub [autonomy], because it’s one we can make money on - we’ve reversed, switching the order of introduction.” “Autonomy will happen,” said Hildegard Wortmann, Audi’s sales and marketing board member, during the CES show. “Probably in commercial vehicles at first. In cars it will take time.” Cambridge-based research consultancy IDTechEx has recently identified autonomous robot shuttles as an important new form of transportation. Think boxy, upright cargo vans or minibuses carrying eight to 20 passengers. IDTechEx examined 36 manufacturers with robot shuttles deployed across 20 countries, identifying nine of excellent design, being successfully trialled, with strong backing and imaginative multi-purposing. It does, however, identify problems such as the huge amounts of electricity required to operate such vehicles and the money required to develop build and operate them. “That will be overcome with a tsunami of new technologies,” it concludes with Panglossian optimism. Or maybe not. Robo taxis are “challenging on the side of complexity”, says Vukotich. “The challenge is the sheer complexity of the environment” and, in a nod to Roy Schneider’s Chief Brody character in Jaws, says: “We’ll need more sensors.” This “sensor dilemma” was becoming clear at CES. Mircea Gradu, senior vice-president of product and quality at Velodyne Lidar, a respected sensor supplier, gave a thoughtful presentation pointing out just how simple it is for car makers to claim they are offering advanced driver assistance when in practice such a system operates at the speed of a slug with the forward vision of a mole. Even the industry-standard NCAP tests of driver aids represent only a small proportion of real-world conditions, says Gradu, posited a rating system for autonomous braking, blind-spot monitoring, adaptive cruise control and lane keeping (all the panoply of the self-driving car), which would allow the public to more fully understand just what their cars can and can’t do. He admits, however, to get to SAE L4 is going to mean the gold standard (actually his ratings are based on diamonds) and that would mean several thousand dollars worth of laser light-surveying (Lidar) equipment on the car. But the biggest hurdle is still our sheer unpredictability. A few years ago, Kiyotaka Ise, Toyota’s former director of research and development, suggested that autonomy might mean that bicycles, motorcycles, horses, even pedestrians, should not be allowed to mix with autonomous vehicles and “everything on the road will have to obey the speed limit”. As Aptiv’s De Vos says: “Technology development continues to be the primary element on the critical path - specifically the ability to develop and demonstrate that the vehicle can reliably handle all of the required use cases for a given Operational Design Domain (ODD).” And what self-driving car developers are beginning to understand is that we do weird things. “Predicting human behaviour is the largest hurdle,” says Toyota’s Pratt. “So if you have a situation where there’s mixed traffic between automated vehicles and human-controlled vehicles, or pedestrians or cyclists, how do you predict what people are going to do? It’s very difficult because we’re not mindreaders and all we can do is consider that people will be benevolent and follow the rules. “We can use a lot of data to train machine learning networks to try to pattern match and say, ‘Well I’ve seen people do this kind of thing before in this situation, so it’s likely that this particular person will do it again’, but of course that’s not always a perfect prediction.” Try walking a three-year-old on a busy pavement... Pratt thinks that the sort of urban shuttles in campus conditions could provide a gateway into full autonomy, however, provided they’re going slowly. “I think we’re going to start to see these kinds of L4 systems, but for a rather restricted ODD; this means slower speed. I think we will begin to see either urban shuttles or taxis that go pretty slow, as the kind of first step in this sort of thing - and it’s potentially a big area that most people have not paid much attention to.” So, too, does De Vos, whose company recently signed a joint venture with Hyundai to begin full driverless systems this year with a view to initial commercial deployment in 2022. “The first deployments will be in the commercial market (robo taxis), driven in limited, well-defined, urban, geo-fenced areas,” he says. He anticipates that the market for these sorts of vehicles will reach “meaningful scale” by 2025, but economic, regulatory and technical challenges will mean we won’t see consumer self-driving before 2030 - and even then in very limited quantities. After you: how aggressive will the self-driving car have to be? The lane running past my house is narrow and muddy, with high banks and poor visibility. The verges are a graveyard of expensive door mirrors smashed off by drivers travelling too fast and not knowing the width of their vehicles. Progress only works with give-and-take, eye contact and gestures, and even then it’s slow. Just like the original series of Darpa challenges for autonomy, it’s easy to envisage an autonomous car getting stuck there forever, uncertain if it should move or not. Pratt laughs. “Yeah, what people do is they engage in this kind of dance - they negotiate. Will you let me in? Will you slow down? Are you not going to be not so nice and not let me in? There’s a non-verbal dialogue where we use our car as a prop; we sort of stick the nose of the car out, we make eye contact, we motion with our hands. An automated car is handicapped with respect to that negotiation; it doesn’t have eyes, or a face, or arms. “There are lots of us thinking about what other ways we could use to indicate what it wants to do. Some kind of lights on the car, perhaps, or the behaviour of the vehicle itself.” What is also being considered is what Pratt terms “multi-hypothesis planning”, where an autonomous car has a variety of “what-if?” plans dependent on traffic and situations, perhaps even messaging surrounding cars as to its intentions. All of it, of course, will have to be done within the laws of the road including the speed limit, which an autonomous car could never break. “I don’t worry too much about this,” says Vukotich. “I can make the car more defensive or aggressive even within the law.” De Vos sees what he terms “Driver Behaviour” as one of the key development targets. “Finding the right balance between aggressive and passive behaviour is one of the key challenges to deploying automated vehicles. Aptiv uses structured AI [artificial intelligence], which enables us to continuously assess the potential path and velocity of the vehicle to find the optimal balance between safety and comfort. To be commercially viable, you have to solve for both.” A work in progress might be the best description of where this particular conundrum is right now, along with lots of the trials currently taking place. Robot minibuses are likely to be the first results but deployed into well-defined areas beyond which they can’t stray, with trains of lorries linked electronically the next. The day, however, when you can buy a fully autonomous car to sit on your drive is likely to be a long way off, at least 10 years, maybe 20. Which makes a farce of the sort of street and infrastructure planning that is going on at present in the UK to accommodate autonomy that is a generation off. We heard of one major road scheme where planners had assumed autonomy would be in common use by 2025, because the Government had said so... Besides, the next big thing at CES this year was vertical takeoff air taxis: Uber and Hyundai’s, Toyota and Joby Aviation. These things are apparently going to take over, save lives, save the planet, be uncrashable, put drivers out of work and change everything we knew about individual mobility. Well that’s the idea, for the moment at least...