Sterling Anderson, Aurora co-founder, speaks to Lex Fridman's MIT self-driving cars class
Lex Fridman hosts Aurora co-founder Sterling Anderson for a guest lecture at MIT on autonomous vehicles.
Summary
Sterling Anderson, co-founder of Aurora and former head of Tesla's Autopilot program, delivers a guest lecture in Lex Fridman's MIT self-driving cars course. He traces his career from his PhD work at MIT on shared human-machine vehicle control, through his time at Tesla where he led both the Model X program and the Autopilot team, to the founding of Aurora in December 2016 alongside Chris Urmson (formerly of Google's self-driving car group) and Drew Bagnell (Carnegie Mellon). Anderson describes Aurora's partnership model with Volkswagen Group and Hyundai Motor Company, and argues that the economics of self-driving deployment in fleet settings already work even at current sensor costs — meaning the primary remaining challenge is technological, not financial. He identifies the forecasting of other road users' intent and future behavior as the hardest unsolved problem in the field.
Key Takeaways
FULL TRANSCRIPT
Introduction and Career Overview
Lex Fridman: Today we have Sterling Anderson. He's the co-founder of Aurora, an exciting new self-driving car company. Previously he was the head of the Tesla Autopilot team that brought both the first and second generation Autopilot to life. Before that he did his PhD at MIT working on shared human-machine control of ground vehicles — the very thing I've been harping on over and over in this class — and now he's back at MIT to talk with us. Please give him a warm welcome.
Sterling Anderson: Thank you. It's good to be here. I was telling Lex just before that it's been a little while since I've been back at the Institute, and it's great to be here. I want to apologize in advance — I've just landed this afternoon from Korea via Germany, where I've been spending the last week, and so I may speak a little slower than normal. Please bear with me if I become incoherent or slur my speech; somebody flag it and Lex will try to make corrections.
So tonight I thought I'd chat with you a little bit about my journey over the last decade. It's been just over ten years since I was at MIT. A lot has changed — a lot has changed for the better in the self-driving community — and I've been privileged to be a part of many of those changes. I wanted to talk with you a little bit about some of the things that I've learned, some of the things that I've experienced, and then maybe end by talking about where we go from here: what the next steps are both for the industry at large and for the company that we're building, which, as Lex mentioned, is called Aurora.
To start out with, there are a few key phases or transitions in my journey over the last ten years. As Lex mentioned, when I started at MIT I worked with Carlo Ratti, Emilio Frazzoli, John Leonard, and a few others on some of these shared adaptive automation approaches — I'll talk a little bit about those. From there I spent some time at Tesla, where I first led the Model X program as we both finished the development and ultimately launched it. I then took over the Autopilot program, where we introduced a number of new active safety features as well as enhanced convenience features — from Autosteer to Adaptive Cruise Control — that we were able to refine in a few unique ways. And then from there, in December of last year — 2016 — I co-founded a new company called Aurora. I'll tell you a little bit about that.
The Intelligent Co-Pilot: MIT Research
So to start out with, when I came to MIT it was 2007. The DARPA Urban Challenge was well underway at that stage, and one of the things that we wanted to do was find a way to address some of these safety issues in human driving earlier than potentially full self-driving could. And so we developed what became known as the Intelligent Co-Pilot.
What you see here is a simulation of that operating. I'll tell you a little more about that in just a second, but to explain the methodology — the key approach that we took that was slightly different from traditional planning and control theory — was that instead of designing in path space for the robot, we instead found a way to identify, plan, optimize, and design a controller subject to a set of constraints rather than paths.
What we were doing was looking for homotopy classes in the environment. Imagine for a moment an environment that's pockmarked by objects — other vehicles, pedestrians, and so on. If you were to create the Voronoi diagram through that environment, you would have a set of unique, continuously deformable paths — homotopy classes — that will take you from one location to another through it. If you then take the dual of that, which is the Delaunay triangulation of that environment, presuming that you've got convex obstacles, you can tile those together rather trivially to create a set of homotopy classes and transitions across which those paths can stake out a given set of options for the human.
It turns out this tends to be a more intuitive way of imposing certain constraints on human operation. Rather than enforcing that the ego vehicle stick to some arbitrary position within some distance of a safe path, you instead look to enforce only that the state of the vehicle remain within a constraint-bounded dimensional tube in state space — those constraints being spatial. Imagine the edges of the roadway, or circumventing various objects in the roadway. Imagine them also being dynamic — limits of tire friction, imposed limits on side-slip angles.
Using that, what we did was find a way to create those homotopy classes, forward-simulate the trajectory of the vehicle given its current state and some optimal set of control inputs that would optimize its stability through that — we used model predictive control in that work — and then take that forward-simulated trajectory and compute some metric of threat. If the objective function is to minimize wheel side-slip, for instance, then wheel side-slip is a fairly good indication of how threatening that optimal maneuver is becoming. What we did was use that in a modulation of control between the human and the car, such that should the car ever find itself in a state where that forward-simulated optimal trajectory is very near the limits of what the vehicle can actually handle, we would transition control fully to the automated system so that it can avoid an accident, and then transition it back in some manner. We played with a number of different methods of transitioning this control to ensure that we didn't throw off the human's mental model, which was one of the key concerns.
What you see here is a simulation that was fairly faithful to the behavior we saw in test drivers up at Dearborn, Michigan. Ford provided us with a Jaguar S-Type to test this on. There's a blue vehicle and a grey vehicle. In both cases we have a poorly tuned driver model — in this case a pursuit controller with a fairly short look-ahead, shorter than would be appropriate given the scenario and the dynamics. The grey vehicle is without the Intelligent Co-Pilot in the loop. You'll notice that the driver becomes unstable, loses control, and leaves the safe roadway.
The co-pilot, remember, is not interested in following any given path. It doesn't care where the vehicle lands on this road, provided it remains inside the road. In the blue vehicle's case, it's the exact same human driver model, now with the co-pilot in the loop. You'll notice that as the scenario continues, what you see on the left in the green bar is the portion of available control authority being taken by the automated system. You'll notice that it never exceeds half of the available control — which is to say that the steering inputs received by the vehicle end up being a blend of what the human and what the automation are providing — and what results is a path for the blue vehicle that actually better tracks the human's intended trajectory than even the co-pilot understood. Again, the co-pilot is keeping the vehicle stable and keeping it on the road; the human is steering toward the centerline of that roadway.
There were a lot of very interesting things that came out of this. We did a lot of work in understanding what kind of feedback was most natural to provide to a human. Our biggest concern was that if you throw off a human's mental model by causing the vehicle's behavior to deviate from what they expect it to do in response to their control inputs, that could be a problem. So we tried various things. One of the key questions we had early on was: if we couple the computer control and the human control via a planetary gear and allow the human to actually feel a backwards torque to what the vehicle is doing — so the car starts to turn right and the human will feel the wheel turn left and see it start to turn left — is that more confusing or less confusing for the human?
It turns out it depends on how experienced the human is. Some drivers will modulate their input based on the torque feedback they feel through the wheel. A very experienced driver expects to feel the wheel pull left when they're turning right. However, less experienced drivers, in response to seeing the wheel turning opposite to what the car is supposed to be doing, found it a rather confusing experience. So there were a lot of really interesting human interface challenges that we were dealing with here.
Application to Unmanned Ground Vehicles
We ended up working through a lot of that and developing a number of applications for it. One of those — at the time Gill Pratt was leading a DARPA program focused on what they called time-maximal mobility and manipulation — we decided to see what this system could do in application to unmanned ground vehicles. In this case, what you see is a human driver sitting at a remote console, as one would when operating an unmanned vehicle in a military context. What you see in the top left is the top-down view of what the vehicle sees, with bounding boxes around various cones.
What we did was set up about twenty test subjects looking at this control screen and operating the vehicle through a track. We set this up as a race with prizes for the winners, and penalized them for every barrel they hit — a five-second penalty for knocking over a barrel, a one-second penalty for brushing one — and they were to cross the field as fast as possible. They had no line-of-sight connection to the vehicle. We played with things on their interface: we caused it to drop out occasionally, we delayed it as one would realistically expect in the field, and then we either engaged or didn't engage the co-pilot to try to understand what effect that had on their performance and experience.
What we found was, not surprisingly, the incidence of collisions declined by about 72% when the co-pilot was engaged versus when it was not. We also found that even with that 72% decline in collisions, speed increased by somewhere in the range of 20 to 30 percent.
Finally, in perhaps the most interesting finding to me: after every run I would ask the driver — and again these were blind tests, they didn't know if the co-pilot was active or not — how much control did you feel like you had over the vehicle? I found that there was a statistically significant increase of about 12% when the co-pilot was engaged. That is to say, drivers reported feeling more in control of the vehicle — 12% more of the time — when the co-pilot was engaged than when it wasn't. And the statistics show that the average level of control the co-pilot was taking was 43%. So they were reporting that they felt more in control when in fact they were 43% less in control, which was interesting. I think it bears a little bit on the human psyche — they were reporting that the vehicle was doing what they wanted it to do, maybe not what they told it to do. That was a fun observation.
From MIT to Tesla and the Founding of Aurora
From there, my collaborator and I looked at a few different opportunities to commercialize this. Again, this was years ago and the industry was in a very different place than it is today. We started a company first called Gimlet, then another called Ride. At the time we intended to roll this out across various automakers. At the time, very few saw self-driving as a technology that was really going to impact their business going forward. Even ride-sharing at the time was a fairly new concept that was, I think, to a large degree viewed as unproven.
So as I mentioned, in December of last year — 2016 — I co-founded Aurora with a couple of folks who have been making significant progress in this space for many years: Chris Urmson, who formerly led Google's self-driving car group, and Drew Bagnell, now a professor at Carnegie Mellon University, who was an exceptional machine learning researcher and one of the founding members of Uber's self-driving car team, where he led autonomy and perception.
We felt like we had a unique opportunity at the convergence of a few things. One: the automotive world has really come into the full realization that self-driving, ride-sharing, and vehicle electrification are three vectors that will change the industry — something that didn't exist ten years ago. Two: significant advances have been made in machine learning techniques, in particular deep learning and other neural network approaches, in the computers that run them, in the availability of low-power GPU and TPU options, and in sensing technologies — high-resolution radar and a lot of the lidar development. It's really a unique time in the self-driving world. A lot of these things are really coming together now.
We felt that by bringing together an experienced team, we had an interesting opportunity to build from a clean sheet — a new platform, a new self-driving architecture that leverages the latest advances in modern machine learning together with our experience of where some of the pitfalls tend to be down the road as you develop these systems. Because you don't tend to see them early on — they tend to express themselves as you get into the long tail of corner cases that you end up needing to resolve.
So we've built that team. We have offices in Palo Alto, California, and Pittsburgh, Pennsylvania. We've got fleets of vehicles operating in both locations. A couple of weeks ago we announced that Volkswagen Group, one of the largest automakers in the world, and Hyundai Motor Company, also one of the largest automakers in the world, have both partnered with Aurora. We will be developing — and are developing — with them a set of platforms, and ultimately will scale our technology on their vehicles across the world.
One of the important elements — Lex asked me before coming out here what this group would be most interested in hearing — was what does it take to build a new company in a space like this. One of the things that we found very important was a business model that was non-threatening to others. We recognized that our strengths and our experience — in my case over a decade, in Chris's case almost two — really lies in the development of the self-driving systems, not in building vehicles. Our feeling was: if our mission is to get a technology to market as quickly, as broadly, and as safely as possible, that mission is best served by playing our position and working well with others who can play theirs. Which is why you see the model that we've adopted, and you'll start to see some of the fruits of that through the partnerships with some of these automakers.
At the end of the day, our aspiration and our hope is that this technology — which is so important to the world in increasing safety, improving access to transportation, and improving efficiency in the utilization of our roadways and our cities — reaches as many people as possible. This may be the first talk I've ever given where I didn't start by rattling off statistics about safety and all these other things. If you haven't heard them yet, you should look them up — they're stark. The fact that most vehicles in the United States today have on average three parking spaces allocated to them. The amount of land taken up across the world housing vehicles that are used less than 5% of the time. The estimate that somewhere between six and fifteen million people in the United States don't have access to the transportation they need, either because they're elderly or disabled or one of many other factors. This technology is potentially one of the most impactful for our society in the coming years. It's a tremendously exciting technological challenge, and the confluence of those two things is a really unique opportunity for engineers — and others who are not engineers — who really want to get involved and play a role in changing our world going forward.
So with that, maybe I'll stop and we can go to questions.
Q&A: Lidar vs. Camera Sensor Strategy
Audience member (Wayne): Thanks for coming. A lot of self-driving car companies are making extensive use of lidar, but you don't see a lot of that with Tesla. Did you have any thoughts about that?
Sterling Anderson: I don't want to talk about Tesla too much in terms of anything that wasn't public information. I will say that for Aurora, we believe that the right approach to getting to market quickly and safely is to leverage multiple modalities, including lidar.
These are all just Aurora videos of our cars driving on various test routes, just to clarify what's running in the background.
Q&A: The Car Enthusiast Market and AVs
Audience member (Luke): A lot of customers have visceral connections to their automobile. I was wondering how you see the car enthusiast market being affected by autonomous vehicles, and vice versa — how AVs will be designed around those types of customers.
Sterling Anderson: It's a good question, and I am one of those enthusiasts. I very much appreciate being able to drive a car in certain settings. I very much don't appreciate driving in others. I remember distinctly several evenings almost literally pounding my steering wheel sitting in traffic in Boston on my way somewhere. I do the same in San Francisco.
I think the opportunity really is to turn personal vehicle ownership and driving into more of a sport — something you do for leisure. A gentleman some time ago asked me, "Don't you think this is a problem for the country — if people don't learn how to drive? That's just something a human should know how to do." My perspective is it's as much of a problem as people not intrinsically knowing how to ride a horse today. If you want to know how to ride a horse, go ride a horse. If you want to race a car, go to a racetrack or go out to a mountain road that's been allocated for it. I certainly agree there is an important place for driving as an enthusiast activity — I'm very much a vehicle enthusiast myself — but there is so much opportunity in alleviating some of these other problems, particularly in places where it's not fun to drive, that I think there's a place for both.
Q&A: Business Model and Open Source
Audience member: Congratulations on the partnership that was announced recently. I have a two-part question. First, we heard last week from someone talking about how long they've been working on autonomous car technology, and you've ramped up extremely fast. Is there a licensing model you've taken? How are you able to commercialize the technology in one year?
Sterling Anderson: Just to be clear, we're not actually commercializing yet — I want to distinguish that from broad commercialization of the technology. We are partnering and developing vehicles, and we may be running pilots, as we announced with the Moia shuttles. I don't want to get too much into the nuances of that business model. I will say that it is done in very close partnership with our automotive partners, because at the end of the day they understand their cars, they understand their customers, they have distribution networks, and they are fairly well positioned — provided they have the right support in developing a self-driving technology — to roll it out at scale.
Audience member: The second part of my question: looking at the pace of adoption and the maturity of technology, do you see an open-source model for autonomous cars?
Sterling Anderson: I am not convinced that an open-source model is what gets to market most quickly. In the long run it's not clear to me what will happen. I think there will be a handful of successful self-driving stacks that will make it — nowhere near the number of self-driving companies that exist today, but a handful.
Q&A: Technological and Economic Bottlenecks
Audience member: In any new product development there are typically two types of bottlenecks — a technological bottleneck and an economic bottleneck. I'd be interested to hear what you would say is the current thing that, if it were ten times better, would unlock the most progress. And on the economic side, I'd be interested to know what cost reduction would be most transformative.
Sterling Anderson: Let me start with the economic side, as that's a slightly quicker answer. The economics of operating a self-driving vehicle in a shared network today would close — that business case closes even with high costs of sensors. That is not what's stopping us. And that's part of why, if your target is to initially deploy these in fleets, you would be wise to start at the top end of the market, develop and deploy a system that's as capable as possible as quickly as possible, and then cost it down over time.
There's no unobtainium in lidar units today. There's no fundamental reason that a lidar unit should lead you to a seventy-thousand-dollar price point. However, if you build anything in low enough volumes it's going to be expensive. Many of these things will work their way into the standard automotive supply chain, into Tier 1 suppliers, and when they do, the automotive community has shown themselves to be exceptional at driving those costs down. So I expect them to come way down.
To your other question — technological bottlenecks — one of the key challenges of self-driving remains that of forecasting the intent and future behaviors of other actors, both in response to one another and in response to your own decisions and motion. That's a perception problem, but it's something more than a perception problem. It's also a prediction problem, and there are a number of different things that have to come together to solve it. We're excited about some of the tools that we're using — interleaving various modern machine learning techniques throughout the system to do things like project our own behaviors, learned for the ego vehicle, onto others, and assume that they'll behave as we would had we been in that situation.
You assume nominal behavior and you guard against off-nominal. But it's very much not a solved problem. As you get into that really long tail of development — when you're no longer putting out demonstration videos but instead just putting your head down and eking out those fine gains — that's the kind of problem you tend to deal with.
Q&A: Security and Malicious Actors
Audience member: This question isn't necessarily about the development of self-driving cars but more of an ethics question. When you're putting human lives into the hands of software, isn't there always the possibility for outside agents with malicious intent to use it for their own gain? And if you do have a plan, how do you intend to protect against that?
Sterling Anderson: Security is a very real aspect of this. It's a constant game of cat and mouse, and so I think it just requires a very good team and a concerted effort over time. I don't think you solve it once, and I certainly wouldn't pretend to have a plan that solves it and is done with it. We try to leverage best practices where we can in the fundamental architecture of the system to make it less exposed — and in particular to make key parts of the system less exposed to nefarious actions of others. But at the end of the day it's just a constant development effort.
Q&A: How Self-Driving Changes Vehicle Design and Society
Audience member: Thank you for your talk. My question is about what opportunities self-driving cars open up, since driving has kind of been designed around a human being at the center since the beginning. If you put a computer at the center, what society-wide differences — and maybe even individual car differences — open up? Could cars go 150 miles an hour on the highway? Could cars look different when a human doesn't need to be paying attention?
Sterling Anderson: I think the answer is yes, and that's something very exciting. One of the unique opportunities that automakers in particular have when self-driving technology gets incorporated into their vehicles is that they can differentiate the user experience. They can provide services — augmented reality services, location services, many other things. It opens a new window into an entirely new market that automakers haven't historically played in, and it allows them to change the very vehicles themselves, as you've mentioned.
The interior can change. As we validate some of these self-driving systems and confirm that they do in fact reduce the rate of collisions as we hope they will, you can start to pull out a lot of the extra mass and other things that we've added to vehicles to make them more passively safe — roll cages, crumple zones, airbags. In a world where we don't crash, there is much less need for passive safety systems. So yes.
Q&A: Go/No-Go Testing and Statistical Confidence
Audience member: I have a question about the go/no-go tests that you conduct for certain features. When you decide to launch a feature, how do you know it's definitely going to work in all scenarios, because your data set might not capture everything?
Sterling Anderson: It's a statistical evaluation every time. You're right — you will never have comprehensively captured every case and every scenario. I think that's an unbounded set — it may in fact be bounded at some point, but I think it's essentially unbounded. So you'll never have characterized everything. What you will have done, hopefully, if you do it right, is established with a reasonable degree of confidence that you can perform at a level of safety that's better than the average human driver. And once you've reached that threshold and you're confident that you've reached it, I think the opportunity to launch is real and you should seriously consider it.
Q&A: Positive and Negative Implications of Self-Driving
Audience member: Thank you for your talk. Self-driving seems to be able to ultimately take over the world to some extent. Just like other technologies, it opens up new opportunities but also brings adverse effects. How do you respond to the unintended effects that may come, and what do you see as the positive and negative implications of future self-driving?
Sterling Anderson: The positive ones are well listed — you'll find them in your favorite press article. The negative ones: in the near term I do worry about the displacement of jobs. This will happen. It happens with every technology like this. I think it's incumbent on us to find a good way of transitioning those who are employed in some of the transportation sectors that will be affected into better work. There are a few opportunities that are interesting in that regard, but I think it's an important thing to start discussing now, because it's going to take a few years, and by the time we've got these self-driving systems on the roads really starting to displace that labor, I'd really like to have a new home for it.
Q&A: Partnering with Multiple Automakers
Audience member (Kasha, Sloan School): My question was more about your business model — partnering with both Volkswagen and Hyundai. What was your thought process about that? Did not one of them want to go exclusive with you?
Sterling Anderson: Our mission, as I mentioned, is to get the technology to market broadly, quickly, and safely. We have been and remain convinced that the right way to do that is by providing it to as much of the industry as possible — to every automaker who shares our vision and our approach. We were pleased to see that both Volkswagen Group — and I'm assuming you all know the scope of Volkswagen; this is a massive automaker — and Hyundai Motor, also very large across Hyundai, Kia, and Genesis, both shared our vision of how we should do this. They both shared a keen interest in making a difference at scale through their platforms. Volkswagen has a very admirable set of initiatives around vehicle electrification and a few other things; Hyundai is doing similar things. For us it was important that we enable everyone, and that was kind of what Aurora was started to do.
Q&A: Vehicle-to-Vehicle Communication and Open Networks
Audience member: Now that a lot of companies are coming up with self-driving cars, most of the technology is bound only to the car. Would we see something like an open network where cars communicate with each other regardless of which company they come from, and would this increase the safety or performance of vehicles?
Sterling Anderson: I think you're getting at vehicle-to-vehicle and vehicle-to-infrastructure communication. There are efforts ongoing in that area, and it's certainly only positive — having that information available to you can only make things better. The challenge has historically been that vehicle-to-vehicle and in particular vehicle-to-infrastructure communication doesn't scale well, and it's been much slower in coming than our development. So when we develop these systems, we develop them without the expectation that those communication protocols are available to us. We'll certainly plan for them and they'll certainly be a benefit once they're here. But until then, many of the hard problems that I would have welcomed a beacon on every traffic light to solve ten years ago — just telling me the state of the light rather than having to perceive it — are now less significant, because we've kind of worked our way through a lot of the problems that would have solved.
Q&A: Cooperative Control of Self-Driving Vehicle Fleets
Audience member: What's your opinion about cooperation of self-driving vehicles? If you can control a group of self-driving vehicles at the same time, you can achieve a lot of benefits to traffic.
Sterling Anderson: That is where a lot of the benefits come from — in infrastructure utilization and in ride-sharing with autonomous vehicles. Specifically, the better we understand demand patterns, people movement, and goods movement, the better we can optimally allocate these vehicles to locations where they're needed. That coordination is where, as I mentioned, these three vectors of vehicle electrification, ride-sharing, and autonomy — or mobility as a service — really come together with a unique value proposition.