If Matt Johnson-Roberson ever wondered why so many autonomous vehicle developers do their testing in Arizona, he got a fuller understanding last winter, riding around Michigan on a tandem tricycle. Sitting side by side and bundled up to ward off the cold, he and his University of Michigan colleague Ram Vasudevan pedaled to keep up with their robot, which was plying the streets of Ann Arbor on its own. One handled the steering while the other worked the laptop that oversaw the REV-1, the autonomous vehicle they created to mimic human bike messengers.
Alex Davies covers autonomous vehicles and other transportation machines for WIRED.
Johnson-Roberson and Vasudevan, who jointly direct the University of Michigan and Ford Center for Autonomous Vehicles, cofounded Refraction AI, the latest self-driving outfit to announce plans to change the way people and their things move about the planet. While a juggernaut like Waymo can take on everything from robotaxis to trucking, this 11-person startup is focused on the local food-delivery market. “Trying everything would be a death sentence,” Johnson-Roberson says.
He has been making robots since 2003, when, as an undergraduate at Carnegie Mellon, he worked on the first Darpa Grand Challenge, a seminal event in the self-driving space. Sixteen years on, with self-driving vehicles still struggling to enter commercial service, he’s eager to see robots have a real role in the world, beyond the Roomba that vacuums his house. “It feels like a bummer that we don’t have anything,” he says.
So Refraction, which came out of stealth mode last week, will avoid the hard parts of driving by acting not like a car, but like a bicycle. The three-wheeled REV-1 is 4 feet tall and 32 inches wide, about the profile of an adult on a bike. It uses bike lanes where available (which should be most of the dense urban areas Refraction has in mind) and hugs the shoulder everywhere else. That confers a few advantages. At just 100 pounds (not counting cargo) and driving at 10 to 12 mph, it can stop in about 5 feet, reducing the need to spot obstacles hundreds of feet ahead and mitigating the damage of any crash. (A prototype hit Johnson-Roberson a few times in his lab, and he walked away unharmed.) That lets the REV-1 get around with a relatively modest and inexpensive sensor suite. It has a pair of solid-state lidar scanners but relies more heavily on readily available radar and cameras than many bigger AVs. It also uses ultrasonic sensors (the ones that beep when you’re about to back your car into a lamp post), whose limited range isn’t a problem here.
Keeping the REV-1 to low speeds and out of the way of cars should help Refraction move to market. It’s now working with two Ann Arbor restaurants, making deliveries to the startup’s employees and hoping to expand to the general public in the coming months. To make that happen, Johnson-Roberson and Vasudevan have a teleoperation setup that lets them control the vehicle remotely, using a system designed for a racing videogame. When one of their five REV-1s encounters an unprotected left turn, someone at the office will take over and handle it manually (the vehicle can also make three right turns to avoid the left).
Same goes for zebra crossings, where pedestrians have the right of way and there’s no light to control traffic. These are complex situations, and Refraction’s cofounders don’t want to take the time to solve them completely before getting to work. “It feels years off,” Johnson-Roberson says. After all, reports have said that Waymo, which has been at this tech longer than anyone, still has trouble with unprotected lefts, while Cruise boasted with a May video of its cars handling the turns.
Like its tech, Refraction’s business plan is streamlined. It’s starting with food deliveries, sticking to dense urban areas, and running routes between .5 and 2.5 miles. This is where the inexpensive sensor suite really helps: The REV-1, which is made mostly of fiberglass and uses an ebike motor for power, costs $4,500 to build, and its designers think they can get it down to $3,500. Johnson-Roberson figures that if Refraction can make four to six deliveries a day, each between $35 and $40, while taking a 10 to 15 percent commission from the restaurant, it can pay off the cost of a vehicle in a few months. Deliveries will be free for customers, to help them get over the fact that they’ll have to walk to the curb and punch a code into the REV-1’s screen to get their grub. From there, Refraction might move into pharmacy delivery, toting Advil and toilet paper.
Refraction’s move into this niche—local food delivery within bikeable areas—is just the latest example of how the self-driving industry has expanded in recent years. Making a robot that can safely do all the driving, even in a limited area, has proven harder than backers hoped. After a decade of work and billions spent, even Waymo hasn’t taken the human safety operator out of its vehicles for its taxi service in Chandler, Arizona. So smaller operations have become savvier, singling out driving tasks they think will be easier to master in the short term, while bringing in revenue.
The flip side of that approach is that limiting ambition can also limit growth. Beyond urban downtowns, bike lanes are hardly common, pushing the REV-1 into the street. And even with a cheap vehicle, it may be hard to compete commercially with the inexpensive humans who already deliver food via services like Grubhub and Caviar. Plus, the autonomous delivery space is starting to look crowded: Nuro recently raised nearly $1 billion and struck a deal with Domino’s. Amazon and Postmates are just two of many companies working on sidewalk delivery bots, while Doordash is working with Cruise to move its food.
Sticking to the bike lane, meanwhile, brings up a novel challenge for Johnson-Roberson and Vasudevan: Operating well enough to avoid turning Ann Arbor’s cyclists against them, and avoiding the sort of attacks that Waymo’s cars have suffered in Chandler. Both men are cyclists. They know how frustrating it is to have cars blocking precious biking real estate, and they want to avoid slowing traffic by cluttering the lane with a stopped or slow robot. “We’re trying to emulate what it is to be a cyclist,” Johnson-Roberson says. That means stamping out false sensor readings that can make the vehicle stop for no reason, while making sure it brakes when it should.
Along with his desire to see the robots on which he has long worked succeed, Johnson-Roberson has another motivation for moving quickly. Winter is coming, and he’d really like to get the REV-1 to the point where it can be safely monitored from the teleoperation setup in the office, so he can skip the freezing ride-along.