Rupert Stadler, CEO of Audi, said they will have fully autonomous cars deployed by 2025.
Mark Fields, former CEO of Ford, indicated 2021.
Elon Musk, CEO of Tesla, said next year.
With headlines like these, it’s hard not to get excited about autonomy and self driving cars. After all, we’ve seen the cars in Minority Report, Total Recall, and iRobot, and thought to ourselves: “When can we finally get into those cars?”
Truth be told, it may be quite a while before we’re actually there.
There’s a general misalignment between what the public think is “fully autonomous” versus what these executives are actually saying. Elon Musk’s 2018 goal is to have a self driving car that’s safer than a human driver. And considering the 20–50 million people injured every year in car accidents, being better than human is a pretty low bar — not at all what the public envisions to be “truly autonomous.”
Surely Ford will have better news in 2021? The truth is…well, complicated.
There are 5 levels of SAE (Society of Automotive Engineers) automation which represents the extent of human involvement in driving a vehicle.
Ford hopes for SAE level 4 autonomy by 2021. But according to the Wall Street Journal, the vehicle will “only be self-driving in the portion of major cities where the company can create and regularly update extremely detailed 3-D street maps.” Like with Tesla, this is hardly the kind of autonomy we imagined of the future.
Even with the recent advancements in machine vision, sensors, and mapping technologies, we’re only really at level 2 automation and slowly moving towards level 3. Just this summer, Audi announced the 2019 A8 — advertised as the first company to sell level 3 self driving car — with a hefty $100,000+ price tag. It can park on its own, and at speeds less than 37 mph, be autonomous enough to fulfill the level 3 criteria.
Meaning: You no longer have to hold the steering wheel (or be prompted after a short duration) during traffic jams and low speed travel. Everything else? Well, you’re still out of luck.
So the natural question becomes: how can we get to level 4 and level 5 within a reasonable time frame? There are actually a couple of key barriers that the industry will have to overcome in order to implement the cars we’ve seen in futuristic Hollywood movies.
In a lot of ways, the development of autonomous vehicles in the U.S. mirrored the development of the internet — concurrent development that’s fragmented and isolated.
You have major automobile companies with deep pockets and massive R&D departments trying to get ahead of the autonomous wave. Uber, Google, and Ford; Volkswagen, Mercedes-Benz, and Tesla, the list goes on and on. Each of these companies are developing independently from one and another and holding their cards very close to the chest. The technologies that work on their vehicles may not necessarily work on others, and the wheel, literally, keeps getting reinvented. And for us, the consumer, that means waiting that much longer for the car of our dreams.
By no means should we advocate for strict standardization of technologies so that monopolies emerge. A car should not have access to just one kind of sensor to install. That would actually stifle innovation. Rather, it is more so important that a concrete framework emerge for how these technologies should interact with each other. That way, no matter the supplier or vendor, technologies can communicate with each other. This is key in the acceleration of research and development.
And interestingly enough…this is exactly what is happening in China.
Baidu, the massive Chinese web services company valued at $80 billion, is investing heavily in the autonomous driving industry. Specifically, its new open source Apollo program hopes to propel China and its manufacturers into the world pole position on autonomous vehicles. According to the Apollo website, Apollo “will provide an open, complete and reliable software for its partners in the automotive and autonomous driving industry.” And what’s even crazier? The entire project, right now, is available on GitHub.
In an article published by Wired on the Baidu Apollo program, Qi Lu, Baidu’s chief operating officer, comments, “With our code base that we released on July 5, [we will make it possible for] one person to assemble a vehicle in three days that can do autonomous driving in limited forms and start on R&Ds.”
Isn’t that unbelievable? Anyone, literally anyone, could download the code from GitHub and get started. Of course, in order to reap the full benefits of Apollo, they ask for users to contribute their data as well — a symbiotic relationship which inevitably accelerates innovation. And you’ll also need your own sensors.
Baidu is able to keep everything open-sourced because of unique market conditions. In the article, Mr. Lu elaborates, “China is highly, highly fragmented. There’s more than 250 car OEMs [original equipment manufacturers]. None of the OEMs will have the full capabilities to build out deep R&Ds.”
And that’s why Apollo is thriving. With over 50 partnerships in China and a standardized system for all manufacturers, China seems poised to surge ahead in the industry.
All of these developments in China remind me very much of the adoption of TCP/IP for the internet. And if history is any indication, I truly think there’s the potential for exponential growth — provided the world follows Baidu’s example.
Of course, standardizing protocols is half the battle. The technology itself must be good. What’s the use of technologies communicating perfectly with each other if these communications amount to nothing.
And here lies one of the core issues of automation. Reliable and accurate data across all conditions. This is something that the executives don’t mention. We’ve seen these executives talk about their lofty goals and aspirations but it’s always framed in the context of perfect driving conditions. They’re not talking about the rainy weather. The blizzards. The dark nights. The times when having a self driving car would be dramatically safer. The kind of weather we’d expect going to work on a daily basis (for everyone outside the Bay Area peninsula that is!).
With current technologies, companies in the U.S. are hiring hundreds, if not thousands, of labelers whose sole purpose is to label images with boxes. This data is in turn fed to the machines so they get smarter.
The bad news? According to Carol Reiley, President of Drive.ai, “For every one hour driven, it [takes] approximately 800 human hours to label.” Even with all the effort spent, the results are still sub optimal.
As an autonomous car, it’s not enough to cleverly interpret and draw boxes around what looks like a speed limit sign. Obviously, that’s a big step: being able to see that it is a sign in the first place. But what’s even more important is seeing, interpreting, and understanding what the sign means. Imagine a car approaching an intersection and correctly identifies a stop sign. The next logical step is to read what that stop sign says. Is it a four-way stop or a two-way stop? Has the sign been changed? What if the map isn’t up to date? These are the things that could result in tragic accidents. Things self-driving is supposed to eliminate or reduce.
This is where pixel perfect, precise, vision comes in. An innovative startup called DeepenAI hopes to solve these challenges. Automating the labeling service has the potential to change the game all together — saving countless man hours. Trained labor processing with pixel level quality allows for cars to detect, in real time, the objects in their field of view.
Pixel level quality is important because it’s not enough to know what separates road from the sidewalk. We want to know and see the potholes, the roadside kills, the infinite variation of things that may exist on the road and not depend on a set amount of data of labeled images.
Precise vision extends far beyond just the single car as well. There must be uniform precision across all aspects of autonomy — even at the map level. How many times have you used Apple Maps and Google Maps for the same route? Chances are…not very often. Yet, these two companies are duplicating and repeating processes in a way that’s inefficient for the consumer. What if Google and Apple had worked together on the same map platform? Likewise, with map data, there must be a centralized process in which companies collaborate and build off of each other.
Imagine a world where cars are identifying and labeling objects in real time. Let’s say a car notices a sign that’s been updated or deviates from the map data. Another car drives by. Same conclusion. After a hundred or so, the map updates automatically to reflect this change and this update applies to every single car in the U.S.
That’s the power of standardization and centralization, and DeepenAI is well positioned to accomplish such a monumental task. And if that is accomplished, then we’re one step closer towards the future.
Are self-driving cars hyped?
Short answer: Yes.
Long answer: It depends.
Shared riding platforms are going to be early adopters, and if you live in a city that’s mapped you could experience it ahead of everyone else — as long as you don’t expect it to be a fully level 4 or 5 platform. But for the rest of us, we have to wait longer than the industry executives may have us believe.
Of course, the issue of standardization vs non standardization also affects the timeline significantly. With programs like Apollo, I can see development rapidly progressing in China. The Singapore government has also reportedly already convinced Apollo to open an office there, hoping to be one of the early adopters.
Many in the transportation industry have jumped on the autonomous bandwagon, and for good reason, but the road ahead is long and arduous. Reaching our collective dream of level 4 and level 5 autonomy will require ingenuity, but more importantly, coordination and collaboration across the entire industry. Or else the true autonomous dream may end with “Do not pass Go.”
— Michael Guo & Li Jiang