Cars That Drive Empty: Emilio Frazzoli on Why Self-Driving Needs Rules, Not Just Data

Frazzoli, introduced as the CTO of nuTonomy and inventor of the RRT* algorithm, returns to MIT where he spent 15 years as a student and then faculty member. He recalls that he first worked on autonomous cars after the DARPA Grand Challenges simply because it was cool: airplanes and cars that fly and drive themselves without hands. The company, he says, was essentially born at MIT out of the research that followed.

The turning point came in 2009, when a team was assembling a proposal on future urban mobility in Singapore. Frazzoli admits he wanted to join mainly because he wanted to go to Singapore. When the organizer asked what autonomous cars had to do with urban mobility, he improvised an answer on the spot: call a car from a smartphone app, ride, and let the car move on to the next passenger. She bought it, and the improvised idea led him to ask why we really want self-driving vehicles.

Frazzoli lists the usual reasons for self-driving cars: safety, since most accidents involve human error and many victims are in their 20s and 30s; convenience; access for people who cannot drive; efficiency; and lower environmental impact. His objection is that all of them improve the status quo rather than change it. What interests him is changing how we think about mobility.

He then runs a rough calculation, which he invites listeners to question. Road accidents cost about 300 billion dollars in the US plus roughly 600 billion in societal harm, nearly a trillion a year. Congestion adds about 100 billion, and its health costs about 50 billion more. But multiplying half the median wage by the hours Americans spend driving gives about 1.2 trillion dollars a year, more than the safety benefit.

The remaining half of his pie chart is the value of making car sharing finally convenient. Bike and car sharing are, in his words, concepts "everybody loves but nobody uses" because of friction, such as no bike available or no place to park. A car that drives itself removes both problems, and with a sharing factor of four, where one shared car replaces four owned ones, he estimates roughly 2,000 dollars a year of benefit.

Frazzoli walks through the SAE levels of automation and says he thinks numbering them was a horrible idea, because a sequence invites the belief that you simply progress through levels 0 to 5. Levels 2 and 3 require a human to supervise and take over, which he says goes against human nature. As a former aeronautics professor he points to aviation: autopilots were expected to reduce accidents but introduced new failure modes, from fading situation awareness to lost reaction skills, and pilots are highly trained, unlike everyday drivers.

Full automation, he argues, is in some ways easier, and it is essential to capture the technology's value. Safety is still an open question, since a few million autonomous miles with a small number of accidents is not statistically significant when humans drive far more miles, and every software change has to be validated again. Time savings also need a car that does not ask for your attention: the better it drives, the harder it is to stay alert. Car sharing needs a car that can arrive with nobody inside.

His conclusion is blunt: level 4 or 5 is the only path that matters, and the game-changing feature is that cars can move with nobody inside. Anything less is "just a nice gadget" to show off to friends.

Frazzoli sketches two industry paths on a chart of scope against automation level. Automakers follow the OEM path, adding driver-assistance features through levels 1, 2 and 3 across millions of cars and eventually having to cross a band that requires human supervision. The other path, which he says nuTonomy, Waymo and Uber are following, starts with full automation in a small geofenced area and then scales up the operation.

The difference matters because a consumer product has to work everywhere, while a service provider can choose location, weather and traffic conditions. Asked when autonomous vehicles will be common, Frazzoli says buying a self-driving car at a dealership is at least 20 years away, but riding one in a new city is a matter of a couple of years. The economics differ too: a buyer would pay at most about 20,000 dollars for the autonomy package, while an operator compares it to roughly 100,000 dollars a year for the three or so drivers needed for 24/7 service.

Other apparent obstacles shrink for a fleet operator. High-definition maps are needed only for the service area, and he says their cost scales with the square root of the customer base. A fleet of vehicles collects the data to make and maintain its own maps, and he expects maps to be "a dime a dozen" within a few years. A professional crew can handle sensor calibration and updates that no ordinary consumer would. Removing the driver, about half of taxi costs, creates a strong margin that can also be passed on to customers.

The usual reaction is that taxi and truck drivers will lose their jobs. Frazzoli recalls being summoned by Singapore's Ministry of Manpower and fearing a shutdown, only to learn the concern ran the other way: mobility services worldwide are manpower-limited, and Singapore would like more buses but lacks people willing to drive them. The same is true for trucking and taxis.

He backs this with another calculation: if everyone in the world used Uber-like services, about one person in seven would have to be a driver. That cannot happen, because people also need to be teachers, doctors and firefighters. Today we all double up as drivers for ourselves, spending around one-seventh to one-eighth of our productive day behind the wheel. He concedes that more supply could push wages down, and notes that trucking accounts for 25 percent of job-related deaths in the US, so remote supervision from an office could help those drivers.

His timeline follows from this: fast adoption in shared mobility-as-a-service fleets, with private ownership much later.

Frazzoli is skeptical of impressive demos, noting that a German vehicle in the late 1990s drove hundreds of miles on highways with cameras and basic computer vision. Using deep learning and GPUs for the same thing, he says, is no progress over 20 years. What he finds meaningful is driving in dense urban traffic, which nuTonomy does daily in Singapore, including right turns across traffic on a left-driving road, and in Boston's Seaport area, where the city permits it.

Quoting the founder and CEO of Mobileye, he lists three challenges: perception, mapping and driving policy. Perception is improving fast and mapping is a logistical pain that will pass, so he sees decision-making as the remaining one. His example video shows a car at a green light yielding to a pedestrian, passing a parked truck, watching a motorcycle and then waiting when the truck starts to move.

Frazzoli argues that rules of the road exist precisely to minimize negotiation, as with everyone keeping right or stopping at red. The old way to encode them was if-then-else logic, which he used in the DARPA Urban Challenge and describes as miserable to debug. He shows the Caltech team's bug there, in which the car backed out of an intersection and was eliminated, as proof that smart teams still miss bugs.

The alternative many people propose is to feed a car enough data and let it learn to drive, as in end-to-end deep learning demonstrations. Frazzoli says he does not want to sound too negative in a deep learning course, but he lists problems. A developer's early traffic-light code effectively learned to speed up at a yellow light, and while that is sometimes right, other features of the situation matter. Engineers also need decisions to be explainable and actionable: knowing why something happened and how to fix it.

He also disputes that there are too many rules of the road. Every licensed US driver has to study the booklet and pass a written test, and the rules fall into about twelve categories, such as who can drive where, speed, yielding, signaling and parking. What is huge is the number of combinations of rules and actors. That makes it hard to code a response for every case and equally hard to learn one, because training data would be needed for every combination.

Frazzoli's central idea draws on the structure of NP-hard problems: finding a solution is hard, but checking a candidate is easy. His RRT* algorithm explores a large graph of reasonable trajectories, and checking one against the rules is linear in the number of rules. nuTonomy writes the rules in a precise formal language, translates them automatically into something like a finite state machine, and runs trajectories through both physical space and a logical space showing which rules are satisfied.

Because sometimes no trajectory satisfies every rule, the system needs a way to handle infeasibility. Frazzoli claims human rule systems are hierarchical, citing Asimov's Three Laws of Robotics, and in driving, not hitting people or cars outranks staying in lane, which outranks speed. Violating an important rule by a little is worse than violating a minor rule by a lot, which gives a total order of costs, and the car then solves a shortest-path problem. He shows a car crossing into the oncoming lane when it is safe, with no special-case scripting.

When he was younger, Frazzoli says, he assumed he could turn the rules of the road into formal language and be done. Then he read them and found them incomplete and inconsistent. His favorite example is the Swiss "fundamental norm" that users must behave so as not to endanger others who follow the rules, which does not say what to do when someone else breaks them. It would imply a fleet of vigilante cars that are technically in the right.

He finds most trolley problems silly but offers a meaningful version: a jaywalking pedestrian ahead, and a swerve that kills a bystander with some probability P. If P is one, you stay on course; if it is zero, you swerve; where the boundary lies, nobody knows. Perception adds a second dimension, because the car is only some percent confident a pedestrian is there. Frazzoli says it should not be him writing these answers but the community.

His closing claim is that the biggest challenge in autonomous vehicles is that we do not rigorously understand how we want vehicles, including human-driven ones, to behave. Right of way has no mathematical definition, and he knows of no numbers for the distances and speeds involved. He wants a sound theory of the rules, perhaps combining formal methods with learning from what people actually do. Such a theory would also derive requirements for sensing and planning, and once it exists, he says, building automated vehicles will be much easier. He ends by saying nuTonomy hopes to double in size in the next couple of years.