91av

AS REVOLUTIONS go, this one has been rather lacking in revs. For the past decade or so, there have been confident predictions that gas-guzzling cars driven by accident-prone humans would soon be on the slip road to oblivion. The future of mobility was to be all-electric – and all-autonomous.

Electric cars are already on the move, although we must go much further and faster if we are to meet climate goals. Meanwhile, however, the “autonomous” bit seems to be stuttering, to say the least.

To be sure, some of the latest commercially available cars come with ever more computing smarts, such as adaptive cruise control, which allows for occasional hands-free use in very specific road conditions. But beyond a few small-scale tests of truly autonomous vehicles, drivers must keep their eyes and minds on the road at all times. A future where the average motorist can sit back, relax, even take a nap and let the car’s computer get them all the way from home to work and back, say, seems barely on the horizon.

Some observers are now openly saying the dream of full autonomy is a mirage: creating robot vehicles able to tackle any kind of road or traffic situation is just too tough a nut to crack. Are they right? And if so, what exactly is keeping down the self-driving car?

The dream of all-electric robo-taxis grew from a series of autonomous vehicle challenges launched by the US Defense Advanced Research Projects Agency more than a decade and a half ago. They showed that sensors and computers could guide driverless cars through a couple of hundred kilometres of terrain, including on mocked-up city streets. Rapid advances in machine vision, artificial intelligence, inexpensive radar and digital mapping fed the optimism that consumer versions were just around the corner.

There has been some progress. Since 2009, self-driving vehicles made by Waymo, part of Google’s parent company, have driven some 30 million test kilometres on public roads in over 25 cities, more than 100,000 of them autonomously in the true sense of that word – without a safety driver to take control. , accelerators or even human-operated brakes. Since 2017, they have been carrying passengers in Phoenix, Arizona. Results from these tests and from other developers, among them universities, car-makers such as General Motors (GM), Toyota, Honda and Ford, and the ride-sharing companies Uber and Lyft, have been used to refine computer driving systems.

In commercially available cars, meanwhile, driver-assist features such as lane keeping, which keeps a car between lane markings without a driver’s intervention, are becoming widespread. This sort of automated system, which changes either a car’s direction or speed, but not both, and still requires machine involvement, equates to level 1 on a established by US-based association SAE International, formerly the Society of Automotive Engineers. At the other end of the scale, level 5 is full autonomy, anywhere, under any conditions.

In 2015, Elon Musk’s electric car manufacturer Tesla started offering a level-2 automation option – one that can change a car’s direction and speed simultaneously – called Autopilot. All the 500,000 cars Tesla delivered in 2020 have the feature, which allows the vehicle to steer, accelerate and brake automatically. Despite the name, Tesla warns that . That means eyes on the road and hands still on the wheel, even when the on-board computer is doing the steering.

In 2018, GM introduced a level-2 “Super Cruise” option in its premium Cadillac line. The company says it allows drivers to take their hands off the wheel on certain company-checked, limited-access, multi-lane, divided highways in the US and Canada. The system tracks drivers’ eyes and warns them if they look away from the road for more than a few seconds. Super Cruise will be available on 22 GM models by the end of 2023, the company says. “Once you have used it, you can’t live without it,” says Super Cruise assistant chief engineer Jeff Miller.

Other car-makers, including Honda, Toyota and Volvo, plan similar systems for highway driving in private vehicles starting in 2023. In Japan earlier this year, Honda even delivered what it says is the first street-legal level-3 self-driving car. Level-3 autonomy is when the driver can relinquish control in certain circumstances and do other stuff, like read a book, while the computer does its thing. The Legend Hybrid EX with Traffic Jam Assist can manoeuvre itself with a human ready to take control with minimal notice, but , following another car in the same lane at speeds up to 72 kilometres an hour.

“The first obstacle to driverless carsis safety, or at least the perception of it”

Audi announced a similar feature in 2017, but European regulators have yet to be persuaded to allow such autonomy on the road. . That speaks to the first major roadblock to making higher levels of autonomy work: safety, or at least the perception of it.

A series of high-profile accidents involving fatalities both inside and outside driverless cars has shaken faith in the idea that they are safer overall (see “How safe are self-driving cars?”). Perhaps the most unsettling incident came in March 2018 when a cyclist who was walking her bicycle across a street in Tempe, Arizona, died after an Uber vehicle testing on-road autonomy with a safety driver present hit her. A concluded that the car detected something 5.6 seconds before the impact, but couldn’t identify it, and that the safety operator was looking away from the road for an extended period at the time. The safety driver is awaiting trial for negligent homicide over the death. .

Sam Abuelsamid at consulting firm Guidehouse, an engineer who formerly developed car safety systems, says a general problem with driverless technologies is that developers often have “no background in safety critical systems”. Instead they “came from technology businesses, where ‘move fast and break things’ works well”, he says. Nobody dies if a photo app fails – but it’s a different story when a ton of metal crashes at highway speed.

But blaming cars and their developers masks a more fundamental problem with level-3 autonomy. A study at the University of Southampton, UK, in 2017 showed that people took an average of about 5 seconds to take control of an autonomous vehicle, with . A car at a speed of 100 kilometres an hour travels something like 140 metres in 5 seconds; even at a low speed of 30 kilometres an hour, a 5-second reaction time corresponds to more than 40 metres, and you have to add braking distance to that.

This fundamental limitation seems to undermine any hope that human and vehicle can somehow share responsibility for reacting to unexpected situations. For that reason, many car-makers are hoping to bypass level-3 autonomy altogether and proceed straight to level 4, in which the vehicle has sole control in certain defined areas – so the driver could take a nap, for instance.

This is where obstacles to automation become a machine vision issue. Autonomous vehicles rely on computers analysing sensor inputs to identify objects, measure changes in their position, predict motions and steer the car to avoid dangers. The systems depend heavily on machine learning, a type of artificial intelligence that collects and analyses data to modify its own algorithms.

Most early self-driving test cars were topped by spinning laser-based radars, or lidars. Operating at much shorter wavelengths than conventional radars, these can measure speed and distance a lot more accurately than radar at distances of up to about 200 metres. Lidars were initially considered essential for real-time, three-dimensional tracking of the local environment, including other cars, pedestrians and cyclists. But at $60,000, early lidars cost more than most cars. Musk in particular took against them, deciding that conventional radar, cameras and improvements in computing technology would be enough.

Human superiority

But standard radars in fast-moving cars can’t spot static objects if other things clutter the field of view. In May 2021, Musk announced that Tesla was abandoning radar, leaving it heavily dependent on camera data. Camera-based systems can classify objects reasonably well, but have difficulty measuring distance, making it hard to predict other vehicles’ paths.

Lidar might still ride to the rescue. Prices have tumbled, with compact, integrated systems costing $500 for driver-assistance modules and $1000 for fully autonomous cars. have all added them into their plans for new models with autonomous features.

But even with lidar, autonomous car sensor systems will be “brittle”, says Abuelsamid. “They often fail on something that would be no problem for a human,” he says. , director of Duke University’s , is blunter. “People need to sit back and do a fundamental rethink of how to make self-driving cars,” she says. “The current way we are approaching computer vision is not going to scale and is not going to work.”

To her, the key problem is that machine learning depends on combining individual sensor inputs, lacking the power of top-down reasoning that humans develop from an overall understanding of the world. “Machine learning works only from the bottom up,” she says. We can instinctively tell, for example, whether lane markings are complete or dashed lines even if they are partly covered by snow, or that a stop sign remains a stop sign even if partially obscured, and instantly recognise the implications of an emergency vehicle heading our way. Machine learning has problems with them all, and even lacks the basic intuitive understanding of the laws of physics that people possess.

Safety demands that truly autonomous vehicles should take avoiding actions, slow down and stop in any unfamiliar situation that might be unsafe. But “all deep-learning systems have false alarms or mistakes”, says Cummings. Self-driving cars that stop at every false alarm would cause gridlock and any number of rear-end collisions.

That is an inherent limitation of machine learning. Whereas a few years ago the big question seemed to be how long it would take to reach the nirvana of level 5, or full, go-anywhere autonomy, now it is which section of a journey can be made autonomous.

GM’s CEO Mary Barra says the company’s goal is to extend Super Cruise to allow “hands-free transportation in 95 per cent of all driving scenarios”. That’s a tall order, however. The US alone has , only 300,000 kilometres of which meet current Super Cruise requirements. Of those that aren’t covered by the system, 4.2 million kilometres are paved, ranging from busy city streets and quiet, wide, well-maintained streets in affluent suburbs to lightly travelled two-lane rural byways without centre lines. The remaining 2 million kilometres are unpaved, lacking markings and often signs. The need to upgrade those roads to be robot-friendly “is a hidden cost most people are not thinking of”, says Cummings. In any event, the priority should be repairing the , she says.

That doesn’t mean truly autonomous vehicles have no part to play. Robots’ great strengths include their ability to deal with boredom on long, predictable stretches, and not needing to sleep. That has led to growing interest in autonomous trucks for freight transport, which can bypass many concerns dogging vehicles with human passengers (see “Keep on automated trucking”).

Navigating poorly marked roadworks, snow-covered surfaces or complex city environments full of pedestrians, cyclists, parked cars and traffic signals is a different matter. While truly autonomous consumer options seem to be on the road to nowhere, there are still near-term hopes for automated taxi services on clearly defined, computer-navigable roads. GM’s Cruise division is doing night-testing in San Francisco of a new autonomous ride-sharing vehicle called the , which would carry four to six people rather than single passengers. Boston-based firm is developing fleets of robo-cars that would provide transport in localised areas, with prime targets being older people who can no longer drive their own cars or have other mobility issues. Similar vehicles could provide shuttle services at airports, industrial parks or business campuses.

But others are pulling back. Uber and Lyft both sold their vehicle development groups to other companies in the past year, citing the costs: Uber to Aurora, founded by a former Waymo engineer, and Lyft to Toyota. Jody Kelman, who heads self-driving planning at Lyft, says robo-taxis might be deployed safely by 2023, but only on certain streets, at certain speeds and in good weather.

There is a growing sense that the phase of irrational exuberance that often characterises new technologies might be over for self-driving cars, replaced by a more limited vision in which automation doesn’t fully replace human drivers, but helps us drive better under certain circumstances. That’s still a revolution of sorts – just not the one, perhaps, we first thought was coming.