Autonomous driving: are we there yet?
Written by Silvio Memme
On my last day working at autonomous driving (AV) technology company Waymo, I handed in my badge and wasn’t able to hail my usual Waymo ride home, so I jumped into an Uber. The timing is almost comical, but within about 1km from leaving work in my first non-AV ride in months, we were inches away from getting in a serious head on collision because of a distracted driver.
We have all heard the stats — an estimated 38,680 died in motor vehicle accidents on US roads in 2020. I think you will be hard pressed to find people who will argue that self-driving doesn’t have the potential to make a huge difference to safety on the roads and to our daily lives.
But the AV market, which is filled with so much promise, investment and quite frankly so much hype, has unfortunately been plagued by seemingly little commercial progress. Over the past year or so, the industry could be characterized by a figurative pumping of the brakes after years of companies over-promising and under-delivering. Responses to the “when can we expect AVs out in the wild?” question like this from prominent leaders in the space tempering expectations are fairly commonplace. Even Elon hasn’t been as bullish as usual!
While it can sometimes feel like the AV market is in a trough of disillusionment, real progress has been made and there are technical and commercial reasons to be excited. I speak from the perspective of someone with a history as an engineer at a traditional OEM, who has watched the inception and development of the AV space with skepticism. But I’ve also had deeper exposure than many to technical developments in the AV space, through some of the most-respected companies in the space. With more than a decade of progress, billions of dollars invested in R&D, and some fresh new approaches, I believe strongly that there is plenty to be optimistic about.
What a ride….
We’ve come a long way from Stanley and the original DARPA challenge that saw the birth of most of today’s modern AV programs.
I had the privilege of watching the early days of the AV space as an insider during my days at Fiat Chrysler Automobiles (FCA) and I thought it would be useful to reflect on the AV journey so far, from the perspective of someone who’s seen both sides of the industry.
As an engineer at a traditional OEM, who spent thousands of hours in vehicles assessing engine calibrations, drivability and emissions (and who loves driving!), I was incredibly skeptical of AVs. Being immersed in the car-centric cultures of the US Midwest and Motor Valley in Italy where the vehicle was more than just a way of getting from A-B, but a key part of your identity, it was impossible to imagine a world in which people would give up their keys to be chauffeured (pun fully intended) around by a robotic car. This skepticism wasn’t helped by the fact that these cars were being built by the same people who put my phone together. In my mind, they were nothing more than a pet project for the Silicon Valley billionaires who didn’t know what else to do with their money. “Let the experts who have been doing this for a century take care of this” was the refrain in my own mind.
But in about 2017, things started to change. To their credit, FCA was one of the first to officially partner with a major AV provider, Waymo. The moves by GM and Ford to make 10 figure investments in Cruise and Argo, respectively, represented tectonic shifts in the automotive space. It demonstrated that in the eyes of the incumbents, autonomous driving was at worst, part of their future, and quite possibly a real disruptor to the business model they had been applying for the past century. The large funding and partnership announcements began to come in hot and heavy, and it became difficult to ignore the noise. While initially it seemed like a hedge or an interesting science project, as top level talent from the traditional OEMs began their migration from the US Midwest to the Valley, I started to think this might be more than just a passing fad…
Welcome to the world of AVs
I got my first real taste of the AV world in during the summer of 2019 when I joined Waymo as part of the business development team. I had the pleasure of taking an AV to and from work every day and the experience blew my mind. In fact, that experience alone convinced me that widespread AV use is completely plausible and has the potential to revolutionize the world. Prior to joining Waymo, I was driving about 1000km a week for the 5 years I worked at FCA. Sitting in the back seat of an AV regularly opened my eyes to a few important things:
- I actually don’t love driving. The whimsical drives in the foothills of the Alps in an Italian supercar we fantasize about is very different from the reality of being stuck in gridlock on a snowy day. I much rather have someone else drive for me
- I am tired of close calls in cars
- I (admittedly) suffer from the odd case of road rage
- I have a heavy foot and have paid a hefty price for that in the past
The Waymo driver, in 2019, was nowhere near ready to be put in the hands of a consumer. During my daily commute, I experienced uncalled for emergency stops, infinitely long waits for unprotected left turns, got stuck in messy construction, and relatively little attention was paid to driver comfort. The well publicized case of the Waymo vehicle going haywire demonstrates that technically we are not there yet. But sitting in that car was like picking up your first smartphone and realizing how world changing the technology had the potential to be.
Just in the past few months, announcements from industry giants like Cruise and Waymo have shown how much amazing progress has been made. Anyone who has driven in driven in SF or NY realizes what a huge feat this is. But is the AV ready to be let out into the wild anywhere? We still have a ways to go….
So why aren’t we there yet?
An AV is a (very, very) complicated system. As with any complicated system (happy to chat about the magical mix of machine design, combustion, thermodynamics, heat transfer and controls of an IC engine — where the difference between success or failure is measured in microns and milliseconds), there is no silver bullet that’s going to solve the AV problem (be it a better lidar, a better camera, a more powerful neural net, a more powerful compute stack). All of these things need to work in harmony. But while it pains me to say it, as a hardware engineer through and through, most will admit that the true value and IP of the AV is the software stack.
So focusing on software, it’s safe to say that today’s traditional robotics tech stack won’t hold up forever. The DARPA challenge architecture dates to the early 2000s. Is there any other software architecture that has held up for two decades? Of course, algorithmic breakthroughs in computer vision, etc have happened, but inevitably these techniques are shoe-horned into existing architecture which means they are hamstrung by the deficiencies of the broader system.
More than once in my time in the industry have I heard senior engineering professionals bemoan the monstrosities the existing AV tech stack has become, as more features are added to modules to account for a new sensor or the latest edge case that needs to be addressed. As Oliver Cameron of Cruise said at the Information’s AV conference earlier this year, the question is a bit of a philosophical one each company wrestles with: hardening existing tech to deliver a product vs looking for a new and better way to do something. It is a tradeoff any engineering organization has faced, and in an industry where so many billions have been spent and so few products have been delivered, it is perfectly understandable companies are looking to harden tech and focus on commercialization. But does that mean that the path has been defined to AV success? I’m not convinced it has.
But it is surely too late to be trying something new?
While billions have been and continue to be poured into this problem, there is still a way to go. While some big name and very well-funded companies have staked claim to technical and commercial leadership in the AV trucking and robotaxi space, it’s difficult to know whether today’s traditional approach to building AVs will eventually get us to the reality we all hope for. But I think we need to be willing to explore other options. In a highly regulated, safety critical industry like transportation, government will undoubtably continue to play a key role in the rate of technology adoption/deployment. Safety has, and will continue to be a primary driver of government’s willingness to allow technology providers to fully deploy autonomous vehicles on the road. To-date, only 16 states allow for the deployment of AVs — governments are still writing the rules of the game that technology providers will be forced to adapt to and abide by. While early AV players have spent millions on lobbying and educating governments and regulators on the potential safety benefits of autonomy, regulation will ensure that a verifiably safer technological solution will have a faster road to adoption.
What are we going to need to get there?
Brian Salesky from Argo (who I have always admired for his pragmatic and forthcoming approach to leading an AV company) clearly pointed out that there are lots of things to tackle, but nothing that can’t be overcome. New and more powerful compute capabilities through the likes of AI chip makers Untether AI, Tenstorrent, and Nvidia will continue to make the software requirements for AVs faster, cheaper, and more efficient, ultimately driving AV systems closer to production readiness. And just like any other hardware problem, scale will bring cost reductions and performance improvements as technologies evolve.
I believe one of the biggest challenges the industry will face is giving itself permission to free itself from years of technical debt. As an engineering org, the most painful moment is the realization that you have created a patchwork of technology that is unrecognizable and “starting from scratch” is really the only way to accelerate the pace of innovation. I have lived this firsthand and it is no fun, but a necessary evil. While it is difficult to define “start from scratch” from afar, anecdotally, I sense this is the point we may have approached when it comes to AVs.
The industry needs to be able to iterate and scale faster.
While the AV problem is an enormous software challenge, having to test new code on the road takes time and money. Being able to shorten testing cycles without reducing the accuracy of your models — ideally through simulation — is the key to making AVs safer and cheaper to develop. I spent a decade of my life running experimental engine tests, scoffing at the idea of being able to simulate the complexities of an internal combustion engine. I compare those boundary conditions to what an AV has to be able to react to on the simplest of drives and I realize I was shooting fish in a barrel. I never had to worry about testing for how my engine would react if a child turned left instead of right, nor were the consequences of being wrong anywhere near as severe.
There must be a smarter way to attack the infinite variety of potential edge cases, and intelligent simulation is the only hope of getting these things done faster, safer and more cost effectively. Amazing companies like Applied Intuition are being built on synthetic data and while all AV companies have leveraged simulation for years, light is being shed on how important they are to their efforts through announcements like this and this showing how serious AV players are taking developing their simulation capabilities.
There’s a new(ish) game in town
Speaking to a colleague who works at one of the major AV players and knows the space intimately, he believes an end-to-end trainable solution, is the way things should be going, but lamented the fact that no one respectable was actually creating one. Enter Waabi.
We participated in Waabi’s Series A fundraise earlier this year and I have had the honour to work with Raquel as she builds out an amazing team rethinking the AV tech stack. The company’s approach has the promise to allow for more robust, safer and higher performing AV systems, using an AI-first approach. By pairing this approach with state-of-the-art simulation capabilities, Waabi can create a system that is a fully trainable end-to-end AV software stack that can more readily be tested/verified against edge cases with realtime, high fidelity, closed loop simulation. The ability to tie these key tenets together creates a virtuous cycle that results in more efficient model development, faster computation, more robust verification, and ultimately a more scalable system that will allow for an increased rate of innovation in the space.
No one knows what the right technical approach is. In any startup, there will always be arguments to be made for why it is too hard, too late, too expensive… and the AV commercialization challenge is no different. While at heart it is an enormously complicated engineering problem, the amount of money, attention and yes, hype, around the industry has sometimes turned this into a philosophical discussion that everyone from your local barber to your friendly neighbour wants to chime in on.
That being said, Raquel Urtasun, the founder of Waabi, has seen the self-driving world from a multitude of different perspectives and knows that diversity of thought is key. When you have one of the brightest minds in the space doing something, you jump on it.
And there we have it — I’ve now joined the thousands of people who have shared thoughts, opinions, comments, and theories on the future of autonomous vehicles. My goal is always to move the conversation forward, and I hope that by sharing insights from someone who’s been on both sides of the table, that I’ve contributed in a small way to doing so. I’d welcome your thoughts, comments and ideas. Please don’t hesitate to reach out!
