After Mastering Singapore’s Streets, NuTonomy’s Robo-taxis Are Poised to Take on New Cities

An AI alternative to deep learning makes it easier to debug the startup’s self-driving cars

Advertisement
Top Tech 2017 logo

Take a short walk through Singapore’s city center and you’ll cross a helical bridge modeled on the structure of DNA, pass a science museum shaped like a lotus flower, and end up in a towering grove of artificial Supertrees that pulse with light and sound. It’s no surprise, then, that this is the first city to host a fleet of autonomous taxis.

Since last April, robo-taxis have been exploring the 6 kilometers of roads that make up Singapore’s One-North technology business district, and people here have become used to hailing them through a ride-sharing app. Maybe that’s why I’m the only person who seems curious when one of the vehicles—a slightly modified Renault Zoe electric car—pulls up outside of a Starbucks. Seated inside the car are an engineer, a safety driver, and Doug Parker, chief operating officer of nuTonomy, the MIT spinout that’s behind the project.

The car comes equipped with the standard sensor suite for cars with pretensions to urban autonomy: lidars on the roof and around the front bumper, and radar and cameras just about everywhere else. Inside, the car looks normal, with the exception of three large buttons on the dashboard labeled Manual, Pause, and Autonomous, as well as a red emergency stop button. With an okay from the engineer, the safety driver pushes the Autonomous button, and the car sets off toward the R&D complex known as Fusionopolis.

By the end of this year, nuTonomy expects to expand its fleet in Singapore from six cars to dozens, as well as adding a handful of test cars on public roads in the Boston area, near its Cambridge headquarters, and one or two other places.

“We think Singapore is the best place to test autonomous vehicles in the world,” Parker tells me as the car deftly avoids hitting a double-parked taxi.

One-North offers a challenging but not impossible level of complexity, with lots of pedestrians, a steady but rarely crushing flow of vehicle traffic, and enough variability to give the autonomous cars what they need to learn and improve.

Riding in an autonomous car makes you acutely aware of just how many potentially dangerous behaviors we ignore when we’re behind the wheel. Human drivers know from experience what not to worry about, but nuTonomy’s car doesn’t yet, so it reacts to almost everything, with frequent (and occasionally aggressive) attempts at safety. If the car has even a vague suspicion that a pedestrian might suddenly decide to cross the road in front of it, it will slow to a crawl.

This mistrust of pedestrians as well as other drivers was designed into the software. “Humans are by far our biggest challenge,” Parker says.

Over the course of 15 minutes, our car has to deal with people walking in the gutter, cars drifting across the centerline, workers repairing the road, taxis cutting across lanes, and buses releasing a swarm of small children. Even a human driver would have to concentrate, and it’s unsurprising that the safety driver sometimes has to take over and reassure the car that it’s safe to move.

To handle these complex situations, nuTonomy uses formal logic, which is based on a hierarchy of rules similar to Asimov’s famous Three Laws of Robotics. Priority is given to rules like “don’t hit pedestrians,” followed by “don’t hit other vehicles,” and “don’t hit objects.” Less weight is assigned to rules like “maintain speed when safe” and “don’t cross the centerline,” and less still to rules like “give a comfortable ride.”

The car tries to follow all of the rules all the time, but it breaks the less important ones first: If there’s a car idling at the side of the road and partially blocking the lane, nuTonomy’s car can break the centerline rule in order to maintain its speed, swerving around the stopped car just as any driver would. The car uses a planning algorithm called RRT*—pronounced “r-r-t-star”—to evaluate many potential paths based on data from the cameras and other sensors. (The algorithm is a variant of RRT, or rapidly exploring random tree.) A single piece of decision-making software evaluates each of those paths and selects the path that best conforms to the rule hierarchy.

By contrast, most other autonomous car companies rely on some flavor of machine learning. The idea is that if you show a machine-learning algorithm enough driving scenarios—using either real or simulated data—it will be able to figure out the underlying rules of good driving, then apply those rules to scenarios that it hasn’t seen before. This approach has been generally successful for many self-driving cars, and in fact nuTonomy is using machine learning to help with the much different problem of interpreting sensor data—just not with decision making. That’s because it’s very hard to figure out why machine-learning systems make the choices they do.

“Machine learning is like a black box,” Parker says. “You’re never quite sure what’s going on.”

Formal logic, on the other hand, gives you provable guarantees that the car will obey the rules required to stay safe even in situations that it’s otherwise completely unprepared for, using code that a human can read and understand. “It’s a rigorous algorithmic process that’s translating specifications on how the car should behave into verifiable software,” explains nuTonomy CEO and cofounder Karl Iganemma. “That’s something that’s really been lacking in the industry.”

Gill Pratt, CEO of the Toyota Research Institute, agrees that “the promise of formal methods is provable correctness,” while cautioning that it’s “more challenging to apply formal methods to a heterogeneous environment of human-driven and autonomous cars.”

nuTonomy is quickly gaining experience in these environments, but it recognizes that these things take time. “We’re strong believers that this is going to make roads much, much safer, but there are still going to be accidents,” says Parker. Indeed, one of nuTonomy’s test vehicles got into a minor accident in October. “What you want is to be able to go back and say, ‘Did our car do the right thing in that situation, and if it didn’t, why didn’t it make the right decision?’ With formal logic, it’s very easy.”

The ability to explain what’s happened will help significantly with regulators. So will the ability to show them just what fix you’ve made so that the same problem doesn’t happen again. Effective regulation is critical to the success of autonomous cars, and it’s a challenging obstacle in many of the larger auto markets. In the United States, for example, federal, state, and local governments have created a hodgepodge of regulations related to traffic, vehicles, and driving. And in many areas, technology is moving too fast for government to keep up.

A handful of other companies are testing autonomous taxis and delivery vehicles on public roads, including Uber in Pittsburgh. The motive is obvious: When robotic systems render human drivers redundant, it will eliminate labor costs, which in most places far exceed what fleet operators will pay for their autonomous vehicles. The economic potential of autonomous vehicles may be clear. But what’s less clear is whether regulators will approve commercial operations anytime soon.

In Singapore, the city-state’s government is both more unified and more aggressive in its pursuit of a self-driving future. “We’re starting with a different philosophy,” explains Lee Chuan Teck, deputy secretary of Singapore’s Ministry of Transport. “We think that our regulations will have to be ready when the technology is ready.” Historically, Singapore has looked to the United States and Europe for guidance on regulations like these, but now it’s on its own. “When it came to autonomous vehicles, we found that no one was ready with the regulations, and no one really knows how to test and certify them,” says Tan Kong Hwee, the director for transport engineering of the Singapore Economic Development Board.

Singapore’s solution is to collaborate with local universities and research institutions, as well as the companies themselves, to move regulations forward in tandem with the technology. Parker says that these unusually close ties between government, academia, and industry are another reason nuTonomy is testing here.

Singapore has good reason to be proactive: Its 5.6 million people are packed into just over 700 square kilometers, resulting in the third most densely populated country in the world. Roads take up 12 percent of the land, nearly as much as is dedicated to housing, and as the population increases, building more roads is not an option. The government has decided to make better use of the infrastructure it has by shifting from private cars (now used for nearly 40 percent of trips) to public transit and car shares. Rather than spending 95 percent of their time parked, as the average car does today, autonomous cars could operate almost continuously, reducing the number of cars on Singapore’s roads by two-thirds. And that’s with each car just taking one person at a time: Shared trips could accommodate a lot more people.

Over the next three to five years, Singapore plans to run a range of trials of autonomous cars, autonomous buses, autonomous freight trucks, and even autonomous utility vehicles. The goal will be to understand how residents use autonomous vehicle technology in their daily lives. Beyond that, Lee says, Singapore is “about to embark on a real town that we’re developing in about 10 to 15 years’ time, and we’re working with the developers from scratch on how we can incorporate autonomous vehicle technology into their plans.” Building new communities from scratch, such as One-North, is a Singaporean specialty.

In this new town, most roads will be replaced with paths just big enough for small autonomous shuttles. For longer trips, on-demand autonomous cars and buses will travel mostly underground, waiting in depots outside the city center until they’re summoned. It’s a spacious, quiet vision, full of plazas, playgrounds, and parks—and practically no parking spaces.

To begin to meet this challenge, nuTonomy has partnered with Grab, an Asian ride-sharing company, making autonomous taxi services available to a small group of commuters (chosen from thousands of applicants) around One-North. Testing the taxis in a real application like this is important, but equally important is understanding how users interact with the cars once they stop being a novelty and start being a useful way to get around. “People very quickly start to trust the car,” says Parker. “It’s amazing how quickly it becomes normal.”

If all goes well, Parker adds, the company should be ready to offer commercial service through Grab—to all customers, not just preapproved ones—around the One-North area in 2018. At first, each taxi will have a safety driver, but nuTonomy is working on a way to allow a human to remotely supervise the otherwise autonomous car when necessary. Eventually, nuTonomy will transition to full autonomy with the option for teleoperation.

“The whole structure of cities is going to change,” Parker predicts. “I think it’s going to be the biggest thing since the beginning of the automobile age.”

This article appears in the January 2017 print issue as “Hail, Robo-taxi!”

Advertisement