The Long-Ignored, Most Obvious Path to Autonomy
Why an L3 system is the secret to putting true self-driving on the road faster.
Imagine a world where cars drive themselves. People are finally free to do more with their time, and the world is a happier and more productive place.
Sadly, there is still no clear path to that future. The world has bifurcated on two distinct approaches to autonomy, defined by their level of automation. On one side, auto companies are aiming for L2: driver assistance systems that still require you to drive. And on the other side, robotaxi companies are aiming to go straight to L4/L5: cars with no drivers that transport people door-to-door like self-driving taxis. L2 systems are not that useful, and L4/L5 are still many years away from broad commercialization.
But what about the middle ground? Companies have largely ignored L3, when a car can truly drive itself without driver oversight in constrained areas (e.g. highways).
L3 just might be the fastest path to transportation nirvana.
The Simple Case for L3
1) It’s true self-driving. Your car drives itself, so you can do something else. No “just in time” human backup required. If something happens that the car can’t handle, it gracefully puts you into a safe position. This is a transformative technology. Better than anything we have today for mass consumption, with the potential to free up billions of hours of human capital every single day. Any system that’s lower than L3 is not very useful, because it still demands your constant attention — the requirement to touch the wheel every 15 seconds is hardly magic, so you might as well be driving yourself.
2) There is a licensed driver behind the wheel, massively simplifying the problem. Having a licensed driver in the car gives the computer flexibility when and where it can self-drive (e.g. freeways or suburban Phoenix). The computer is no longer required to self-drive perfectly everywhere. The computer just needs to drive perfectly somewhere. Perfection somewhere is still a real challenge, but it is orders of magnitude simpler than boiling the ocean.
A licensed driver also eliminates the potential dangers of remote driving. This concept has been popularized lately as a bridge to self-driving, with the idea that somewhere call centers full of remote drivers will fill in all the blanks where computers fail. This is actually a terrible idea and should never come to market because it is both economically infeasible and incredibly dangerous.
3) It’s how every other technology succeeds. Technology that changes the world starts by solving a real problem in the simplest possible way. The path is straightforward: Build the minimum viable product that solves a problem, sell it to people who need it, then expand and improve product functionality with real customers and revenue. The gospel of the MVP is repeated so often in Silicon Valley, it’s almost impossible to imagine that tens of billions of dollars of investment forgot the rules. Companies promised a car that could go everywhere, but in a decade failed to commercialize a car that could go somewhere. What happened?
Why Aren’t More People Building L3?
A brief interlude on defining levels of automation. The SAE (formerly the “Society of Automobile Engineers”) is a trade organization that has defined technical standards for the automotive industry since the early 20th century. Their “Levels of Automation” (chart below) has become the de facto standard for categorizing autonomous vehicle capabilities.
Auto companies have been stuck on incremental innovation. These companies dominate L1 and L2 products. L1 was introduced 50 years ago with cruise control. L2 has come to include all of the basic driver assistance features from the past two decades — lane assist, blind spot detection, emergency braking — as well as those that promise pseudo self-driving functionality like Tesla Autopilot.
The common theme in all L2 features: they require your attention. Auto companies build and market features as L2 because they know that their software isn’t very good and that they might need you to take over the wheel in a moment’s notice. You become the back up system when things inevitably go wrong. Anyone who markets their product as L2 or L2+ is telling you, in no uncertain terms, “I’m not sure if this works,” leaving you holding the bag.
But perhaps this is no surprise. Self-driving is a software problem, and OEMs are manufacturers first and marketers second. At their core they are not software developers, not staffed or structured to do it well. The exception of course is Tesla, based in Silicon Valley and built with software engineering at the core from the beginning. But they too seem to be stuck in L2 with Autopilot. Since the product debuted in 2014, Autopilot has not improved much, and ultimately drivers still need to pay attention.
Instead of perfecting Autopilot on the highway to sell “true self-driving” (L3 and beyond), Tesla seems to be opting for a horizontal L2 strategy. They have added summon and parking features, and are teasing “full-self driving,” adding Autopilot capabilities beyond just the highway. But none of these are true self-driving, capable of eliminating human oversight. By “full self-driving” they must mean “L2 in more places” — they have not demonstrated perfection anywhere, so it is hard to believe perfection everywhere will suddenly appear.
While the auto companies went incremental, tech companies went all or nothing. Robotaxi makers like Waymo, Cruise and Uber are all trying to build L4/L5 vehicles. These are considered “fully autonomous,” capable of doing complete trips end-to-end. This is of course incredibly hard, and has yet to work.
Interestingly, Google built an L2/L3 product in 2013, but quickly shelved it because they observed their test drivers failing to constantly monitor the system for error. They learned that they could not rely on people as backup. This observation seems correct — it is critical to take the human out of the loop in a real-time driving system. But the remedy seems extreme — instead of perfecting the driving system to not require a human to intervene at a moment’s notice, they decided to build something that would not require a human at all, ever. Their response is a bit like observing a leak in your roof at home and deciding that the only solution is to move to the desert where it never rains.
The decision to remove human drivers entirely broadens the scope of the problem dramatically. An L4/L5 vehicle must go door-to-door by definition. Now every scenario across every road type must be perfected before commercialization. Eight years later, they still have not reached a broad audience. It has proven simply too ambitious a starting point.
L2 to L3 is the leap — from people driving to computers driving. This is the thing that promises to change our lives. But there are two critical technical challenges.
The first challenge is knowing when and where the computer can’t drive. We have all encountered moments like this on the road — whiteout blizzards, massive downpours, unusual terrain, crazy drivers — the moments where we simply decide to pull over until we get more clarity. For autonomous vehicles it is no different — they have to know their boundaries exactly, and switch to a much simpler, safer mode whenever they recognize that they have reached one of these boundaries.
The second challenge is gracefully transitioning to a safe position after the computer decides that it can’t drive. This requires a simpler, secondary driving system that can always guide the vehicle to a safe place to stop. Critical here is the idea of “not wrong” driving behaviors. While there are surely optimal behaviors for every driving scenario, there are also “not wrong” behaviors that can keep every car on the road safe — humans (and computers) can and should slow down or pull over in a torrential rain storm.
Solving both of these problems makes it possible to remove human oversight and the need for real time human intervention. This is the critical barrier to achieving true self-driving at L3, but also L4/L5. The minimum viable product for a self-driving car is an L3 system that can drive and gracefully transition to a safe position in one narrowly defined area (e.g. highways). From there, adding more areas is a natural incremental path to L4/L5 capabilities, eventually going door-to-door.
L3 is true self-driving, the first step towards unlocking the potential of billions of drivers stuck behind the wheel. It drives your car, and does not require any immediate human intervention. It knows its limits, and can gracefully transition to a safe position when it cannot drive. It fits our world as it is today, and massively improves it. This is the right path.
John Hayes is the CEO and cofounder of Ghost. He founded Pure Storage in 2009, taking the company public in 2015 (NYSE: PSTG). Follow him on twitter @ghosthayes. Learn more about Ghost at www.driveghost.com.