We were promised self-driving cars, but what we got were delays. In 2016, Tesla CEO Elon Musk promised drivers a car that could drive “hands-free” from New York to Los Angeles within a year. That same year, Carlos Ghosn, the former executive leading Renault, Nissan, and Mitsubishi Motors, promised self-driving cars by the end of the decade. Lyft co-founder John Zimmer predicted most Lyft rides would be AVs as early as 2021 and private car ownership would be phased out of major US cities. Ford Motor CEO Mark Fields told CNBC in 2017 that self-driving vehicles with “no gas pedal, no steering wheel, and the passenger will never need to take control of the vehicle in a predefined area” would hit the streets by 2021.
In 2019, it’s clear that none of this is happening on schedule.
Autonomy, alongside electrification, promises to be one of the most profound changes to human mobility since the advent of the automobile. Legacy automakers are racing to transform themselves from metal benders to suppliers of mobility “services” as the world begins to shift away from personal car ownership. Silicon Valley companies are trying to replace the incumbent automakers by developing the technology faster than rivals in Detroit.
It’s not going as expected. Everyone thought self-driving cars were a data problem. But data was only a piece of the puzzle. It is actually a design problem. Companies are rethinking transportation from the hubcaps up: new business models, new drivetrains, and a new definition of personal mobility. The arrival of our autonomous future will take far longer and be a lot different than we expected.
For years, self-driving car companies touted the distance their vehicles traveled or the distance simulated in massive data centers. They calculated their progress on the self-driving puzzle in miles. Some still do. In July, Waymo CTO Dmitri Dolgov told a TechCrunch conference the company’s cars had driven more than 10 million real-world miles, and 10 billion miles in simulation. Tesla’s Elon Musk, whose customers have driven more than a billion miles on autopilot, claims “it’s extremely difficult…. to catch up when Tesla has a hundred times more miles per data than everyone else combined,” according to a transcript from Sentieo.
The view that more miles will make self-driving cars suitable for the open road traces back to an influential 2009 white paper by Google researchers, The Unreasonable Effectiveness of Data (pdf). The paper showed how datasets containing billions or even trillions of objects were enabling computers to comprehend language and photos for the first time, problems that had flummoxed computer scientists for decades. Massive image and language libraries meant computers no longer needed to be told exactly what to do, or even how to identify images or words. Using machine learning, algorithms known as neural networks and modeled loosely on the human brain trained themselves by feasting on all that data.
From this process emerged a system that was competent at identifying cats and responding to “OK Google” requests. Eventually, it began to beat humans at complex tasks such as the board game Go. The theory was that sending cars outfitted with sensors and computers down enough roads, enough times, would ultimately account for almost all the possibilities needed to render self-driving cars safe.
But it hasn’t worked out that way. While stunning progress has been made designing cars that can drive themselves under typical conditions, the open road is a far more complex, chaotic environment than anticipated. There’s essentially an infinite number of edges cases, and any number of them could be deadly. “People understand ‘miles driven,’ and it sounds like progress,” says Noah Zych, chief of staff for Uber’s Advanced Technologies Group, which has reduced its reliance on miles as a benchmark. “But some quick math will show you can’t test all the things that can happen.” A paper by Israeli researchers estimates it would require roughly 30 billion miles of real-world testing, or around 1,000 years of testing for a fleet of 100 cars to guarantee an acceptable likelihood of traffic fatalities. Toyota has estimated it will need at least 8.8 billion test miles (including simulation) before safely deploying self-driving cars.
Even that isn’t enough to catch all the cases. “When you’re 90% done you still have 90% to go,” says Sacha Arnoud, a director of engineering at Waymo. As reality has settled in, the company’s CEO, John Krafcik, has started lowering expectations. “Autonomy always will have some constraints,” he admitted to Wall Street investors last year. “It’s really, really hard. You don’t know what you don’t know until you’re actually in there and trying to do things.”
Those challenges blow up the industry’s timeline.“Everyone is pushing out autonomy much further than predicted,” says Robert Siegel, a lecturer at Stanford’s Graduate School of Business. Not just by years, but by a decade or more.
“We overestimated the arrival of autonomous vehicles,” admitted Ford Motor’s CEO Jim Hackett in April. Ford no longer plans to deliver a “fully autonomous vehicle” in 2021. Lyft suggests its rides won’t be mainly autonomous until 2029—at the earliest. This year, Cruise delayed its own robo-taxi service until some time in the future. (“Developing and deploying self-driving vehicles at massive scale is the engineering challenge of our generation,” Dan Amman, the CEO of Cruise, said in a statement in May.) Only Tesla has continued to promise robo-taxis by next year, but experts are doubtful: “They’re not even close,” says Steven Shladover, a research engineer at the University of California Berkeley.
The AV race is now setting off in another direction. Miles are still important, but autonomous car companies are now adopting the aviation and nuclear industry playbook, and designing in safety from the start. This “safety case” approach breaks down every element of the system into its core components. For driverless cars, that means identifying every possible point of failure (cyberattacks, collision risk, misidentification, etc.) Each one is verified by math and logic, and then put through software simulation and hardware validation. Only then does it enter the real-world for testing.
Uber, which suspended its autonomy program after one of its self-driving vehicles killed a pedestrian in Tempe, Arizona, released its own safety case this summer in the hope it will become an industry standard. Safety associations and standards are springing up all over. Ford, GM, and Toyota formed the Automated Vehicle Safety Consortium this year. BMW, Jaguar Land Rover, and self-driving startup Zoox formed another.
The auto industry is reaching the same conclusion: Every autonomous vehicle must be as safe as the next. One rogue company could wreck the entire effort. “There’s been a lack of transparency within the industry about how the tech works and why we think it’s safe,” says Jack Weast who leads the autonomous vehicle standards effort at Intel’s Mobileye. “A lot of the answer has been its proprietary magic, trust me…The auto industry has competed on safety for over a century. This is not a space to compete. Bad actors would ruin it for all of us.”
Automakers don’t need prior approval for any design. They certify their vehicles to certain standards. Government labs test them. Ultimately, a car can stay on the road until problems are caught and show up in the crash and mortality statistics. But just one pile-up caused by AVs would likely set the industry back years. “We need to arrive at a common definition of what a safe AV is, and more importantly how to measure and assess it,” Weast says.
In that sense, the car industry is now starting to look more like the aviation industry, in which the world’s two biggest rivals, Boeing and Airbus, share all their safety data and must by law. There’s not yet a safety framework in place to mandate data sharing for autonomy and no government or industry yet agrees on just what self-driving cars are considered “safe.”
But the industry is reorganizing. It’s creating deep alliances and partnerships. Ford and VW. Honda and GM. Waymo and Fiat-Chrysler. Competition remains fierce, but it’s no longer (just) which firms build the better car. Automakers must decide what they can do best in an uncertain world in which they sell mobility, not engines and chassis.
Asking which company will “win” the autonomous race is not the right question. The victors (and there will be a few) will not be the ones to cross an arbitrary self-driving milestone first. It will be a multi-decade marathon sprint. Firms that generate the expertise and revenue to stay in the race—selling safer products people want at affordable prices—may live to cross the finish line. “The horserace is irrelevant because the time horizon has moved out,” says Robert Siegel. “[Autonomy] is such a complex problem there is no single metric of who’s ahead.”
Three forces are remaking the car industry: electrification, autonomy, and shared transport (a la UberPool or Lyft Line). Each one is enough to upend the auto industry’s century-old hierarchy. Together, a radical shakeup is unavoidable. It will take decades and billions of dollars (trillions once the cost of infrastructure is counted) to transition from a world dominated by personally-owned, mostly analog vehicles to self-driving vehicles owned mainly by fleets.
It’s tempting to think of this as old versus new, Silicon Valley versus Detroit. Yet both sides need each other more than ever. The auto industry has the manufacturing expertise and capacity. Silicon Valley has the technology, talent, and experience to build software and support billions of users on its networks. The future belongs to companies that can position themselves—through acquisitions, partnerships, and their own core competencies—in critical roles of this new ecosystem. No one can do it alone: The expertise to build world-class software, manufacture cars at scale, craft a superior user experience, and execute viable transportation business models is too diffuse across companies and geography. “If your strategy is do everything, you have no strategy,” says Stanford’s Seigel.
With the end of the old car economy in sight, legacy players are busy tossing out old assets to finance the new, from car-sharing (GM’s stake in Maven, BMW’s bet on DriveNow and Daimler’s ownership of Car2Go) to ride services (Ford’s acquisition of Chariot and bike-share programs, GM’s $500 million investment in Lyft, and Volkswagen stake in dispatch service Gett).
And the need to master self-driving technology has triggered a flurry of acquisitions and investments in the new car economy too. Venture capitalists are pouring money into the sector at an unprecedented rate: $10 billion last year. The number of deals soared from 8 in 2009 to more than 200 in 2018, with Uber, Faraday Future, Lyft, Aurora, Nio, and Xpeng nabbing the most.
Silicon Valley firms are spinning out self-driving units to take outside investment, while automakers are coalescing into globe-spanning alliances. Volkswagen has invested $7 billion with Ford in the startup Argo AI to develop autonomous cars, and merged its own $1.6 billion self-driving division to create Argo’s European headquarters. Uber has spun off its self-driving car unit (now valued at $7.25 billion) to take on $1.5 billion from SoftBank, Toyota, and Denso. Toyota teamed up with automotive suppliers Aisin Seiki and Denso to invest $2.8 billion in the Toyota Research Institute-Advanced Development (on top of its previous $1 billion investment). Volvo has a $300 million joint venture with Uber, while GM invested $500 million in Lyft.
GM’s bet on self-driving cars is Cruise and it’s perhaps the most illustrative. The Detroit titan bought the startup in 2016 for $1 billion just two years after it was founded. What happened next is a microcosm of what’s happening in the entire car industry. Last year, GM announced it was closing five North American factories and cutting 14,700 jobs (including 15% of its North American white-collar workforce). At the same time, Cruise’s headcount in San Francisco grew from 36 to more than 1,500 people. Cruise’s co-founder Kyle Vogt has said it will keep growing by 40% every quarter as the company hires hundreds of software and hardware engineers.
It’s easy to understand GM’s actions looking at one of Cruise’s cars. Most of a self-driving Bolt’s value resides in its sensors, electronics, and lithium-ion battery pack. One report estimates a fully autonomous electric Bolt costs around $200,000, about six times more than a conventional one. None of this is the car industry’s traditional expertise. On each roof sits a metal crown of sensors with cylindrical lasers known as LIDAR, radars, and cameras. Electronics overflow out trunks packed with cellular routers, servers, antennae, flashing LED lights and meters of wires. If this is the brain of an electric car, it looks like an unruly teenager with a lot to learn. And adult supervision is still necessary. Whenever the Bolts leave Cruise’s underground garage, there’s a safety driver and engineer sitting in the front seats. Between them is a big red button, ready to push, in case the car needs to make an emergency stop.
But this car is GM’s great hope. The fleet gives Cruise’s 1,600 employees rides throughout San Francisco, a prototype for the robo-taxi service it plans to roll out globally. The company’s Orion assembly plant in Michigan is now producing them on the same assembly lines as GM’s conventional retail model. If GM can crack San Francisco’s tough streets, it thinks it can quickly scale up to seize the autonomous opportunity. “It’s the most engineering-intensive thing ever attempted,” one automotive executive told Design News about AVs. “And you need lots of the world’s best engineers to do it. I’m not talking about tens or hundreds of engineers. It’s in the thousands.”
Investors appear to think it’s a smart bet. The 110-year-old GM, which hasn’t seen its stock price budge in the last decade, has a market cap of $56 billion. Cruise is already worth about a third of that, an estimated $19 billion, according to PitchBook, and has attracted more than $7.5 billion in investment from Honda, SoftBank, and T. Rowe Price.
The distant future is more clear than the near future. We know cars will ultimately drive themselves. When, and whether we’ll even own them, is the big question. To get them on the road now, engineers have found a way to solve the “edge case” problem—how to account for highly unlikely but potentially lethal situations—for two specific applications: trucks on the highway, and cars on very specific routes. Both solutions are now being tested on public roads on their way to market.
- Truck platooning: Put humans in the loop to manage difficult conditions
- Transit shuttles and robo-taxis on routes: dramatically simplify the driving environment
Truck platooning: Automated trucks are the holy grail of shipping. Tractor-trailers that carry more cargo at lower cost are poised to transform an industry that historically operated on punishing margins below 4%. Engineers at companies like Peloton Technology are doing it by relying on what they call the world’s best sensor: human drivers. Instead of eliminating human drivers, driverless trucks follow closely behind them. This practice, known as platooning, allows one or more trailing vehicles to use automatic steering and braking systems to keep every truck in the platoon at a safe distance and speed. A few trials are underway. Trucks in Germany hauled cargo across 35,000 km (21,748 miles) between Munich and Nuremberg in a seven-month platooning trial. Drivers only overrode the automated system once every 2,000 km (1,242 miles). Doubling or tripling drivers’ loads with platooning is now on the horizon in the next two years. When will trucks drive themselves without any human assistance? “We think those are a long way off,” said Peloton CEO Josh Switkes at the Automated Vehicle Symposium in Orlando last month.
Transit buses: It’s not sexy, but slow-moving buses are the immediate future of autonomous vehicles. Amid the hype over driverless Teslas, projects like May Mobility’s shuttle service in Colombus, Ohio are quietly shuttling thousands of people at top speeds of 25 miles per hour. May Mobility’s six-seat electric shuttle has completed more than 10,000 rides along its three-mile route by the Ohio River. The shuttle fulfills a crucial role in urban transit: regular last-mile service between transit and final destinations. Cities such as Detroit and Grand Rapids in Michigan, and Providence, Rhode Island have already contracted with the company (which still uses safety drivers). May has succeeded this quickly because it simplified the driving audience with a regular route and defined stops, and contracted directly with cities on well-trafficked routes, says May Mobility COO Alisyn Malek. May’s not alone. A number of companies are elbowing in on the market: Optimus Ride, Coast Autonomous, France’s Navya, and Drive.ai.
Robo-taxis: We’ll see robo-taxis on the streets in the near future, but not in the way you might think. Rather than roving cars driving us anywhere in a city, cars will traverse carefully chosen routes at low speeds under 40 MPH in a handful of cities. City roads remain far too complex, geographically distinct, and unregulated for any company to guarantee safety without a driver along for the ride. Intel’s MobileEye looked at future autonomy scenarios and concluded that all roads lead through robo-taxis. “We came to the realization we must go through a robo-taxi phase because of regulation,” said Amnon Shashua, CEO of Mobileye at the TechCrunch Mobility conference this July. “We also came to the realization that this robo-taxi phase is not going to be short.”
Waymo’s trials in Arizona, Cruise’s tests in San Francisco, nuTonomy’s service in Singapore, and Argo’s fleets in Detroit, Pittsburgh, Miami, and Washington DC all face years of refinement. The simpler the route, the faster a company can ditch the driver, but the more constrained the system. While platooning trucks can scale up across the nation as soon as they master highways, truly free-range robo-taxis may take a decade or more to arrive because each city is difficult in its own idiosyncratic way.
Incidentally, autonomy is not just for cars. It’s also flying drones, hauling mining ore, moving planes around airports, logistics barges, and other vehicles. Expect to see these areas make huge leaps in autonomy, perhaps even leapfrogging passenger vehicles, in the coming years.
How fast companies move, and how well their strategies match the market, will make the difference. The rumors that Silicon Valley will displace Detroit are overblown. First, nations are unlikely to watch the automotive equivalents of Google and Facebook dominate the world’s mobility market. Automakers are national champions in many countries (see the US auto industry bailout or Japan’s arrest of former Nissan chief Carlos Ghosn to see how that works). This national importance, plus their under-appreciated ability to mass-produce cars, means legacy automakers are not going anywhere soon.
“Building cars is more difficult than building self-driving technology,” says one industry insider who has worked with technology firms and carmakers. “That’s blasphemy in Silicon Valley. It’s like saying God doesn’t exist in a church. But [they] don’t get it.” The idea of Silicon Valley taking over auto manufacturing wholesale, he says, is “as foolish as GM starting a search engine.”
But the capital needed for the car industry to reinvent itself on the fly has forced hard decisions. In 2018, as GM was closing five factories, Ford gave up building unprofitable passenger cars (except for its Mustang) so it could focus on pick-up trucks, autonomy, and electrification. “We’re going to feed the healthy parts of our business and deal decisively with the areas that destroy value,” Hackett said at the time. “We’re starting to understand what we need to do and making clear decisions there.” Apple and Alphabet opted not to make their own cars. (Waymo refused to front the costs for Ford to retool factories for its autonomous vehicles).
The various autonomous camps can be viewed through the lenses of three different strategies: Build it yourself, license and sell, and partner up.
Build it yourself: The DIY-approach means everything is designed in-house and through close, exclusive relationships with suppliers. Apple perfected this strategy while manufacturing the iPhone (Taiwanese suppliers assemble all iPhones and Apple partners with its supply chain to conduct R&D). Ford, GM, VW, Tesla, Daimler, and Zoox are all trying a version of this strategy with autonomous cars. Many are convinced it’s the only way to deliver a nearly perfect autonomous experience because it’s impossible to get suppliers to provide what’s needed off-the-shelf.
GM will spend about $1 billion, or 13%, of its annual capital expenditures, on self-driving cars and services. Ford and VW have teamed up. Daimler is working with Bosch. “The big difference to other competitors is that we are conceptualizing our vehicle as a robo-taxi right from the beginning and not as a technology kit mounted on a serial vehicle,” Wilko Stark of Daimler and Mercedes-Benz said last year. “We will not have a makeshift solution.” For Zoox (which has raised $790 million, according to Pitchbook) and China-based Nio, two self-driving startups with just a few billion in the bank (or less), the challenge will be doing this on a far tighter budget.
It’s a tightrope act and not everyone will make it across.
License and sell: Google’s Android dominates mobile phones with a 74% market share. Companies like Aurora Innovation and Alphabet’s Waymo want to do the same for autonomous cars’ operating systems. Waymo has spent billions developing and testing its software, and it has made massive deals with automakers to supply its autonomy suite for thousands of cars. (Fiat Chrysler and Jaguar Land Rover will supply 62,000 Pacifica minivans and another 20,000 Jaguar electric I-Paces.)
Waymo appears to be replicating Google’s Android strategy. It buys hardware from suppliers (the equivalent to phone manufacturers such as Samsung and LG), while Alphabet owns the customer experience and data. That’s a nightmare for some automakers who fear being commoditized as hardware suppliers without access to lucrative premium markets. Honda, for example, walked away from a deal last year to supply Waymo (which refused to share its autonomy expertise). Klaus Froehlich, the head of development at BMW, which has about 1,000 people working on its research and development team, said last year that Apple and Google will never get into its cars.
But seeking to control the platform for autonomous driving is short-sighted, argues Sterling Anderson, co-founder of Aurora Innovation and the former manager of Tesla’s Model X. Aurora has struck deals to supply autonomous systems to Hyundai and Fiat-Chrysler. Yet Volkswagen walked away from its partnership with Aurora this year to buy a stake in Ford’s Argo. “Some automakers see autonomy as the next engine. If they don’t own the IP, they’re not in control of their own destiny,” Sterling said. “But others take a more pragmatic approach. There are benefits to participating in a platform not owned by a single automaker.” Namely, economies of scale. A bigger platform means faster learning, cheaper hardware, and lower development costs, if they can become the platform of choice. That’s Aurora’s play. It deploys its own vehicles, and licenses its AI Driver to partners. Nuro, Voyage, Optimus Ride, and Drive.ai are hoping to do something similar.
Partner up: Everyone has partners, but some have made partnerships their primary strategy. “People realize they need to partner,” says Stanford’s Siegel. “There will not be the Cambrian explosion [of companies] we have seen. The capital required is so huge, the only way to do things is to partner and work together on fewer things.”
Honda is a prime example. The company, which sells fewer cars and spends less on R&D than many rivals, moved slowly into autonomy. It’s now pouring money into the effort to catch up. Last October, Honda announced it would invest $750 million in Cruise and another $2 billion over the next 12 years developing the technology with GM. It’s doubling down on its core competency—designing and engineering an interior and exterior for Cruise’s robo-taxi—while revamping its entire lineup over the next decades through its “2030 Vision.” Similarly, Ford bought Argo AI in 2017, one year after it was founded. Despite being years behind Cruise and Waymo, VW announced it would invest $2.6 billion in Argo and it plans (pdf) to test its technology with delivery services and retool one of Ford’s Michigan factory to build autonomous vehicles.
Slow movers may still find their footing in the autonomy space. The auto parts firm Aptiv predicts (pdf) that affordable autonomous personal vehicles won’t arrive until 2030, years after robo-taxis have become well established.
A lot. Autonomous cars are still fooled by raindrops, masking tape, seagulls, bushes, hills, and snowflakes. And that’s the shortlist. Cruise and Waymo have had to cut back vegetation on regular routes because bushes and hedges still baffle self-driving cars. But autonomy is more than technology. It’s regulation, hardware, economics, and safety. Here’s a summary:
Regulation: Very little regulation has been decided in the autonomy space. That’s left automakers groping for regulatory reassurance to avoid expensive balkanization of legal regimes.
In the US, states have taken the lead (often co-opting cities) in deciding how AVs will be rolled out in their communities. At the federal level, the Trump Administration dissolved its AV advisory committee after just one meeting, and industry insiders say conversations in Washington suggest no rule-making will be coming out of this White House anytime soon. Last year, Congress failed to pass legislation on the topic, but legislators say they may make a second try this year at passing the AV START Act, a national safety effort for AVs. There’s more progress in Europe and Asia where governments have pursued firm regulatory frameworks: UK, Germany, South Korea, and Singapore all have national guidelines in place. Yet full AVs are still illegal under international law. The 1968 Vienna Convention on Road Traffic for international rules, updated in 2016 for vehicles that can drive themselves, doesn’t allow for fully driverless vehicles since humans must still able to take control if needed.
Sensors: Autonomous cars use a battery of sensors to see the world: lasers (LIDAR), radar, ultrasound and cameras. No one knows which will be the standard. At least two, perhaps more, will be needed for safety. But which ones? Until recently, the gold standard in perception has been LIDAR: millions of laser blasts paint a precise “point cloud” picture of their environment. At $75,000 per sensor, the cost was steep. Radar and cameras are far cheaper, but lack the precision or range. Now the price is plunging as the capabilities of all three types of sensors grow.
Tesla has publicly stated it will only rely on its existing suite of sensors without LIDAR for full self-driving capabilities in the future, putting it at odds with the rest of the industry. At the same time, companies such as Luminar have demonstrated LIDAR sensors with far greater range priced below $1,000. The radar sensor startup Vayyar says it has a $100 radar sensor that can rival LIDAR in its precision and even track a passenger’s body position and breathing inside the vehicle. Whatever solutions win will offer the optimal mix of price, performance, and safety. We may see all of them become standard equipment on future vehicles once the price drops low enough. Until then, the lack of standardization will raise the cost and complexity of manufacturing and regulating AVs.
Business model: Mobility subscriptions, not car ownership, are ostensibly the future. Car markets have likely peaked in the US and Europe and are set to decline sharply as ride-sharing and autonomous vehicles spread. Transportation modelers predict (pdf; p 68) a fleet of just a few thousand AVs could replace most cars in major cities. A small AV fleet could effectively replace 90% of personal vehicles in Lisbon, Portugal, for example. University of Texas researchers report a similar result in Austin, Texas, while 3,000 shared AVs could meet 98% of New York’s existing taxi service and still keep average wait times to under three minutes.
Once AVs become affordable for most, the Rocky Mountain Institute, a think tank, estimates urban robo-taxis will become a $120 billion industry in US cities. (It forecasts that day will arrive by 2025.) People will switch from personal to shared transport predicts Barclays (pdf) as the cost of a shared ride-hailing services drops from about $1.25 per mile today (Uber) to just eight cents per mile.
How to make money doing all this is another question. In one optimistic scenario, automakers like Ford and VW will look more like Netflix for buses, cars, scooters, and bikes. But this model has not been proven out. Neither Lyft nor Uber is profitable yet. (Uber posted a $5 billion loss last quarter.) Automakers have little experience making money on services rather than products. That question mark is at the heart of automakers’ future.
Safety: AVs must be better drivers than humans, says Aurora’s Sterling. “At the end of the day, it really is the only thing that matters,” he says. “If we can get there, we can prove to regulators with the data that deploying these systems in a moral imperative.” But we still don’t exactly know how to measure this. We know humans aren’t all that safe. Human error is at the root of 94% of all crashes in the US, estimates the Department of Transportation, causing about 40,000 deaths per year.
But there’s not yet a consensus on a rigorous baseline against which to test AVs. Sterling argues we need to collect extensive data on granular elements of driving—behavior at intersections, or detection of pedestrians, for example—and then rigorously assess each of these against how well humans perform in similar situations. Only once this has passed a certain threshold should AVs be allowed to drive unsupervised. At the moment, precise data is scarce on these fronts. AVs may outperform humans in some respects, yet new errors may be introduced. Unlike humans, for example, AVs are not yet adept at forecasting the intent of others, and then accurately predicting how others will respond to them. Its a feat humans perform almost effortlessly as social animals, but we’re only beginning to tackle in self-driving algorithms.
Will it take a breakthrough or mere refinement of existing technology to get there? “That’s the million-dollar question,” says Sterling who helped engineer an entrant in DARPA’s urban challenge for autonomous vehicles in 2007. “I’ve been through this enough times in the last decade I wouldn’t be surprised if we ended up encountering a few substantial improvements we need to make.”