BY MARTIN RUDIGIER
ME 302C - The Future of the Automobile - Mobility Entrepreneurship is a Stanford course taught by Reilly Brennan and Stephen Zoepf. The Spring 2018 course will feature a series of guest speakers across the spectrum of the mobility industry, with Zoox, May Mobility, Waymo Drive.ai, Waybots, Phantom Auto featured among others.
Drive.ai is a Silicon Valley startup that was founded by mainly former lab mates out of Stanford’s Artificial Intelligence Lab in 2015. In the past three years, the company has raised a total of more than $77 million in venture funding to ‘build the brain of self-driving vehicles’ using deep learning, which the company believes is key to make autonomous cars a reality. Drive.ai was one of the first startups to be issued an Autonomous Vehicle Testing Permit in the State of California, and now operates a fleet of vehicles that can drive fully autonomously even on a rainy night. And the company just announced that they will offer a self-driving car service for public use in Frisco, Texas, starting in July, 2018. You can read more about the launch in this article by Drive.ai board member Andrew Ng.
With Waymo launching a self-driving car serivce in Phoenix this year and GM’s announcement to launch its robotaxi service in 2019, the race to deploy large fleets of robotaxis in metropolitan areas is gaining momentum. I think it is quite impressive that a relatively small startup like Drive.ai is competing with the deep pockets of Google (Waymo) and GM (Cruise) to be one of the first companies to launch an autonomous ride-hailing service. I was also impressed by the the company’s focus on the interaction between humans and automated vehicles. As seen in the picture below, their autonomous Nissan NV200 vans have large screens that can display messages to pedestrians to communicate their intent. Since most people will first interact with autonomous cars as a pedestrian, cyclists, or as another driver (vs. as a passenger), I believe that these interactions will be crucial for the public acceptance of autonomous vehicles.
According to the LA Times, Drive.ai will offer rides free of charge, and its vans will operate only between predetermined pickup and drop-off spots. The service is limited to residents, employees and patrons of the buildings participating in the pilot program and the cars will have human backup drivers or chaperones riding in them at all times.
For this post, I had the chance to interview Carol Reiley, one of the co-founders and board member at Drive.ai. As Reilly put it in his introduction of Carol, ‘there are few people who have touched as many applications of robots as her.’ With 18 years of both academic and professional experience in robotics and artificial intelligence (AI), Carol has always tackled highly regulated applications such as space, underwater, and medical. At Intuitive Surgical, for example, she worked on the da Vinci robot as one of the leading scientists in the field of teleoperated surgical robots. This fascination with building products and software in highly regulated fields eventually led Carol to self-driving cars.
How difficult was it to start Drive.ai?
We first made the decision to start the company in 2015. Some things were easier than I thought, some things more challenging. We were fortunate to start with a large and extremely talented technical team which had already worked together in the past. With such a great team right from the start, there was no sense of loneliness that some entrepreneurs experience when they set out to start their own company. I also had many successful startup friends, and much of the basic company infrastructure, introductions to OEMs and VCs, and advice came from them. We turned down several offers of acquisition and VC seed funding at this stage.
To move quickly, I seed funded the company, so that we would not waste time, and found us free office space at an incubator. On a personal side, it was the year I was getting married and I did not want to spend a year of my life planning it. So I decided to use my ‘wedding fund’ money to seed fund Drive.ai. I didn’t take salary for the first year so we could pay the engineers and extend our runway. Nvidia also donated over $100,000 worth of GPUs to us, which enabled us to get started right away.
Six months later, we raised a relatively large Series A of $12 million and had our first prototype. Knowing how hard the problem we were trying to solve was, we raised a large amount that gave us a longer runway to revenue and more time to hit significant milestones. It was very challenging to sell Sand Hill Road on an idea that seemed completely crazy back then, as self-driving technology was still mostly research. They were offering $3 to $5 million investments but we stuck to our target of $12 million. At the time, many investors had never heard of deep learning and Cruise had not been acquired by GM yet, so the general investor sentiment for the autonomous vehicle space was much less optimistic and aggressive than it is today. We ended up hosting workshops for interested investors, and were ultimately able to close our Series A as planned.
The initial feedback we got was often discouraging, because the VCs were hoping we would pivot into something quickly profitable. We also started with too narrow of a focus: a perception solution. We soon pivoted to owning the self-driving car stack, user experience, and operations after getting funded. Within that time, we negotiated a few deals and definitely felt like a startup underdog with a unique approach that the industry wanted and needed to survive this transportation revolution.
In the context of Cruise, Waymo, and other prominent companies in this space, what makes Drive.ai different?
First of all, we are one of the veterans in this space. Students in Stanford’s Artificial Intelligence Laboratory had been experimenting with self-driving vehicle software since 2003, and the founders brought deep expertise in this field.
Second, we consider ourselves a ‘deep learning first’ company. Deep learning is the closest algorithm to how the human brain learns. Consider how young (or old) people learn to drive: First, they need to learn certain rules ‘by the book’, for example the meaning of different road signs. Then, they need to gain on-the-road experience. Similarly, autonomous driving algorithms can not be purely rule-based. Instead, we feed our deep learning algorithms with a lot of different examples — what is good driving behavior, what maneuvers are safe, what is a car, what is a pedestrian, and so on. It then starts to generate its own set of rules on how to navigate the road, much like a new driver would.
And third, we recognize that autonomous vehicles don’t just need to be as ‘smart’ as humans, but their intentions need to be crystal clear. Autonomous vehicles will be the first robots that people will interact with, and these first interactions will be crucial for the general sentiment towards self-driving cars and their adoption. And for most people, the first interaction with autonomous vehicles will not be in the car (i.e. as a passenger), but rather as a pedestrian, bicyclist, motorcyclist, or another driver. We spend a lot of time thinking about different ways in which cars can communicate with the world through audio, visual, and motion cues. For example, today’s cars have a monotone horn sound, but there might be value in redesigning the horn to feature a friendly honk vs. an alerting honk etc. For self-driving cars to gain the trust of the public, they will have to communicate their intentions and interact with human-operated vehicles and pedestrians effectively.
There are quite a few articles that describe Drive.ai as a retrofit company. But you just announced that you are launching an autonomous ride-hailing service in Texas. Did you change the strategic direction of the company?
Well, we never wanted to build a car, but we also did not want to focus solely on retrofitting cars. Retrofitting was a short term solution to get going. While we started by focusing mainly on perception, we expanded our focus to motion planning, fleet management, app development, and operations, for example.
Today, we work directly with businesses and governments. We’re not in the business of making and selling consumer cars.
What do you think is the timeline for L4/L5 autonomy in (a) passenger cars, (b) ride-hailing services, (c) trucks, and (d) logistics (particularly last-mile)?
There is a lot of appetite for autonomous vehicles in ride-hailing services, mostly because L4/L5 autonomy is well suited for this use case (e.g. geographically limited to certain cities) and would fundamentally change the unit economics of the ride-hailing business. Thus, I think that we will see autonomous vehicles transporting people in this space the soonest. Drive.ai is already doing this now. We have really hit a big milestone running a pilot without a driver in the car. We have accomplished this by working with city officials and getting their support. Over the next several years, I expect more and more pilots in controlled settings to be rolled out.
For logistics and trucks, I think the timeline will be longer, let’s say 5 to 15 years. The main issues I see here are public perception, cybersecurity, and regulation. If you think about 40 ton trucks (incl. trailer) driving autonomously on highways at 65 mph, there is less room for error. Also, different loads can significantly affect the ‘behavior’ of the trailer, which makes it even more complex to develop this technology for trucks.
Lastly, I believe that L4/L5 autonomy in passenger cars is still decades away. No one will want to buy a car that is geographically limited to the San Francisco area, for example. And we don’t even have great internet coverage between the major cities in the United States, even though this is a technology that has been used for decades.
In this context, what do you think is the greatest misconception about self-driving cars right now?
The greatest misconception is that the technology is going to be fully solved out of the box and ‘ready’ once it passes a test set with a high level of confidence. But the sheer variety of possible driving situations based on geography, weather conditions, etc. is is large, that this is a chicken and egg problem. While deep learning is extremely powerful, I think that many people have unrealistic expectations of this technology and often make it sound like artificial intelligence is the solution for everything. We need to let the technology be released ‘into the wild’ and then iterate. And it needs to happen soon. Over 40,000 Americans die every year from fatal car accidents, and these numbers have increased over the past couple of years. Self-driving vehicles will be better and safer.
The other common misconception is that self-driving cars are going to mix in with how human-driven cars drive today. I believe that there are going to be completely new car designs, new fueling systems, new technologies (e.g. V2X communication), different speed limits, etc. and these changes will fundamentally change the whole transportation ecosystem, not just the cars itself. As an analogy, we moved from dirt roads used for horses and carriages to concrete roads and complex infrastructure with the invention of the car more than a century ago. I believe that autonomous cars are going to bring a similarly dramatic change to how we think about transportation today.
Your background is in computer science and robotics, but you were initially focused on different applications of this technology. What surprised you most about the automotive space when you first got into self-driving cars?
What really fascinated me about self-driving cars was the impact that this technology could have (e.g. significant decrease in the number of road deaths and injuries, more ‘productive’ hours for millions of commuters, etc.). I always wanted to build things that have as large of an impact as possible. And I believe that we don’t even realize the full scale of potential benefits that autonomous cars will bring to the world. I am pretty sure that in the early days of the internet, no one could imagine all the potential applications that would one day be enabled by this technology, such as Netflix or Facebook or Youtube stars. Similarly, I think that we are now in the (very) early days of autonomous cars, and we can’t even imagine the full range of benefits that this technology will bring.