Since Google launched its self-driving car project in 2009, the biggest challenge has been one of technology: can it be safe enough to deploy at scale?
That dispute is over. Google’s project, now branded Waymo, has experienced only minor incidents – about once every 210,000 miles – since 2019 when it began operating a driverless service in Phoenix, Arizona. Cruise, its GM-backed rival, received a permit last month to begin commercial operations in its home city, San Francisco. Both groups are valued at more than $30bn (€25.5bn) by the most reputable names in venture capital and tech.
What these rivals are not doing, however, is conquering one metropolis after another the way Uber deployed in 100 cities within four years of launch. The costs are just too exorbitant, the testing hours too prolonged and it remains unclear whether there is really a business case for shared “robotaxis”.
Meanwhile, a new threat has emerged: suppliers of advanced driver-assistance systems, or ADAS – a bottom-up approach to building autonomous technology – are making massive strides. They already have a great business case, generating profits as they sell their tech to carmakers, constantly upgrade their systems and save lives along the way.
Several experts say this is a better pathway to scaling driverless tech. If they are right, then the central risk for the robotaxi hopefuls is not whether full autonomy can succeed, but whether an entirely different approach to the problem will get there first.
“There’s no more dispute around whether robotaxis are real: they are real today,” says Karl Iagnemma, chief executive of Motional, the autonomous driving unit of Hyundai and Aptiv. “The question is whether the other guy can come along and do the same service, the same product, but at half the price. If you’ve got a competitor who’s in that position, you’re in big trouble.”
Driverless groups such as Waymo, Microsoft-backed Cruise, Amazon-owned Zoox and Aurora, which announced plans for a public listing last week, are betting on a “moonshot” solution with no plan B. They plan to offer full autonomy – albeit ringfenced to certain locations – or nothing at all. In regulatory jargon, this is called Level 4, in which a robot driver requires no input from passengers. Level 5, the highest step, would allow the vehicle to go anywhere.
This “go big or go home” approach stands in direct opposition to the step-by-step path of the ADAS players led by suppliers Mobileye, Aptiv, Magna and Bosch, which work with all the major carmakers. Their advances mean most new vehicles already have partial automation – Levels 1 and 2, including cruise control and automated braking. Tesla’s AutoPilot System is the best-known Level 2 system.
The notion that this low-cost, evolutionary track could experience a butterfly-like transformation to offer a fully driverless experience has long been dismissed by the Level 4 groups.
Chris Urmson, Aurora chief executive, put it eloquently in 2015 when he was Google’s leading driverless engineer: “Conventional wisdom would say that we’ll just take these driver assistance systems and we’ll kind of push them and… over time, they’ll turn into self-driving cars,” he said. “Well, that’s like me saying that if I work really hard at jumping, one day I’ll be able to fly.”
Urmson’s logic felt sound at the time: ADAS was rudimentary, whereas robotaxis seemed just a couple of years away from mass deployment. Waymo prepared to order 82,000 such vehicles in 2018, Uber projected it would have 100,000 on the road by 2020 and Lyft divined that a “majority” of its rides would be autonomous by 2021.
But none of that happened. Instead, the closer they got to a consumer-facing product, the more complex the problem was understood to be.
At the same time, the traditional automotive industry has been galvanised by these efforts and has evolved ADAS into a multi-faceted feature set capable of hands-free highway driving, automated lane changes and robotic valet parking.
Iagnemma says the dramatic improvements in ADAS were unforeseen, and he now sees the two technology curves converging: Level 4 groups are desperately trying to bring costs down to boost their business case, while ADAS suppliers are accelerating their performance to achieve maximum safety.
“In 2015 I would have agreed with Chris [Urmson],” Iagnemma says. “Every intelligent observer in the industry believed that was the right path forward. But what that failed to anticipate was the increase in performance, in part enabled by deep learning and other advances, that would allow us to do things with radars and cameras that I would not have thought possible in 2015.”
Nevertheless, the Level 4 groups continue to dismiss ADAS as a threat. Shortly before stepping down as Waymo chief executive earlier this year, John Krafcik said: “There is really no path from L2 to L4 – there’s a huge chasm. It’s a completely different development mindset.”
So if the evolutionary approach to building driverless technology proves successful, the upshot would be startling: the world’s biggest, most sophisticated companies – Alphabet, Apple, Amazon and Microsoft – would have all backed the wrong horse for a future technology widely expected to earn revenues in the trillions of dollars.
This would entirely flip the script from half a decade ago, when carmakers felt under siege from tech companies that were going to displace them at the top of the value-chain.
“There was this general fear that the entire auto industry was going to get taken over by tech – and that totally didn’t happen,” says Austin Russell, chief executive of Luminar, which began as a supplier of visual sensors but now offers a highway autonomy solution for partners such as Volvo.
“That whole ‘shoot for the moon’ approach really didn’t play out. That’s why they are in the testing stage and we’re in series production,” he adds. “The traditional automakers still absolutely control all the volume, all the production, everything that goes out there. They are the only ones that can provide a real business case for any of this stuff.”
Today’s driverless car groups are burning through untold amounts of cash. Cruise alone has raised $10bn and just opened a $5bn credit line to build more vehicles. Waymo was funded by Alphabet for a decade before raising $3.2bn in 2020, but it still needed to raise another $2.5bn last month. Zoox was so close to running out of money that it sold itself to Amazon early in the pandemic, while Apple has been working on autonomy since 2014 without so much as a prototype to show for its efforts.
Several big companies have already thrown in the towel: Uber in effect paid rival Aurora to absorb its 1,200-person team last year, while Lyft sold its ambitiously-named Level 5 unit to a division of Toyota in April.
Meanwhile, the global ADAS market has become a gold mine. Revenues last year were $25bn, according to BlueWeave Consulting, and are expected to nearly triple by 2027. Roland Berger, a consultancy, expects advanced automation features to be common for “nearly every new vehicle sold in the developed world” within four years.
Regulators are hesitant to give a green light to robotaxis, but they are cheerleaders for driver-assist tech. America’s highway regulator estimates that thousands of lives a year could be saved if all cars had partial automation features. The EU has mandated that all new cars must have lane-keeping assistance and advanced emergency braking by 2022.
Aside from Elon Musk, who in 2019 promised Tesla would have “operating robotaxis next year” no one is really suggesting the ADAS players are about to “unlock” fully driverless capabilities in the next few years. Nor does their current focus on highway driving necessarily translate well into urban autonomy. But they are under little pressure to make that leap anytime soon.
The business model of selling highway-only autonomy is solid, growing rapidly, and drivers have demonstrated they are willing to pay. Super Cruise, GM’s semi-autonomous feature set, must be purchased in a bundle costing $6,150, and when GM surveyed Cadillac owners last year it found that 85 per cent want it in their next vehicle. Tesla sells its “full self-driving” package, which requires no additional hardware, for $10,000 – an incredibly lucrative sum for the auto business where the average vehicle profit is less than $2,000.
So if ADAS players get stuck at highway-only autonomy, they are not in crisis. But if Waymo, Cruise, Zoox and Aurora delay their rollout, they don’t have a product.
Yet the likelihood that driver-assistance systems do keep moving forward, tortoise-like, is high, because as more cars are equipped with the tech, the faster the systems learn. This is where Tesla currently leads: more than a million of its vehicles are equipped with its AutoPilot system that is always on, in “shadow mode”, ready to upload snapshots to Tesla servers whenever the human driver makes decisions different from its own.
Tesla is, in effect, outsourcing its real world testing to owners who pay for the privilege. It is a stunningly more efficient model than most Level 4 groups, which bankroll hundreds of engineers to sit behind the wheel of robotic test fleets.
Willard Tu, senior director of automotive at US chipmaker Xilinx, says it’s only a matter of time before all the bigger automotive groups are performing similar feats. “Every company that’s doing ADAS is now trying to leverage, and create, machine-learning databases,” he says. “Just like Tesla…they are learning to manage troves and troves of data.”
Some already are. Navigation specialist TomTom has equipped more than 3 million vehicles with high-definition maps accurate to a few centimetres that get near real-time updates from crowdsourcing. “It’s a continuous stream of data that’s coming in, and that data is exploding – it’s exponential growth,” says Willem Strijbosch, TomTom’s head of autonomous driving.
The robotaxi groups’ counter argument that their data is superior is undoubtedly true. Waymo’s fifth-generation sensor suite, for instance, comprises 29 cameras and a suite of radar and lidar – light detection and ranging lasers that create 3D renderings of the environment. By contrast, a Tesla vehicle has eight cameras, most ADAS-equipped cars have even fewer, and virtually none today are equipped with lidar.
But as costs fall dramatically, carmakers will be able to add new layers of sensors to help cross the chasm into a Level 4 feature set. As head of visual sensor company AEye, Blair LaCorte may be biased, but he’s not alone in projecting that lidar will be in “almost all” new cars in half a decade.
“Lidar will be in almost everything that moves,” he says. “The question will be what business model gets it there.”
TomTom’s Strijbosch argues that the Level 4 cars are “overfitted” with technology because they let costs run amok and need to make up for their small sample sizes. But even if ADAS data is less robust, it can be multiplied by a far greater number of real-world hours and eventually close the gap with Level 4 tech. “By the time we’ve driven so many billions of kilometres, in millions of vehicles that have this whole sensor set-up, all corner cases will have been captured,” he says.
Mercedes expects to have regulatory approval in Germany later this year to allow its Drive Pilot system to take full control in certain dense-traffic highway situations at up to 60km/h. For this Level 3 system, the driver will be able to take her eyes off the road entirely and it is the German carmaker that would be liable for any accidents.
Georges Massing, a Mercedes executive, says “the next logical step” is to press ahead “with the development of Level 4/Level 5 technologies towards series production”.
The Level 4 groups could end this debate by rolling out robotaxi services in multiple cities at scale. But most experts believe that remains a distant prospect.
“It will take another 10 to 20 years until [they] go from the suburbs of Phoenix to something that runs across the country,” says Jan Becker, chief executive of Apex.AI, an automotive software group.
Tech isn’t really the problem; societal acceptance is. Amnon Shashua, chief executive of Intel-owned Mobileye, points out that if a Level 4 system can get to a point of crashing only once every 1m miles – two times better than a human driver – that would risk massive reputational blowback.
“If I drive 10 miles per hour, that means I crash once every 100,000 hours of driving,” he explains. “So if I deploy 100,000 cars, I’ll have a crash every hour. From a business perspective that is very, very challenging.”
Costs would also be prohibitive. “If you were to try and deploy 100,000 driverless vehicles that would cost you tens of billions of dollars,” Luminar’s Russell says. “The key distinction here is that we’re actually getting paid to put our stuff on cars, to deploy around the world and collect data.”
If robotaxis fail to scale up in the coming years, analysts say the ADAS advantage will become clearer. Today, the biggest difference between the two approaches is the tech itself: ADAS is low cost and limited; Level 4 is high cost and sophisticated. But in a few years, the biggest difference is likely to be cash flow: ADAS players will be raking it in; Level 4 groups will be burning it at even greater rates. – Copyright The Financial Times Limited 2021