I ordered robot takeout on two campuses with wildly different results
I learned a lot ordering a donut and a muffin from delivery robots.
I’d been following the robot for about five minutes when it seemed to get hopelessly lost.
The four-wheeled vehicle, the size and shape of a large cooler, was navigating the campus of Howard University in Washington, DC. A digital display on the front showed a pair of pixelated cartoon eyes, but the robot was struggling to understand its surroundings.
The robot repeatedly stopped, turned around, retraced its steps, and turned again. When it reached an intersection, it seemed afraid to cross the street. Instead, it turned around and went back for some 200 feet before freezing once again.
Finally, a young woman rolled up on a scooter and scooped up the confused robot. I identified myself as a reporter and told her I’d recently ordered a muffin for robot delivery. I learned that the robot, made by a company called Kiwibot, doesn’t handle vehicular traffic well. It needed help crossing the street so it could pick up my muffin. So the woman zipped off on her scooter with the robot balanced in front of her.
I caught up with her and watched her put my muffin in the robot, put the robot back on her scooter, and ferry it across the street once again. The robot then made its way to my chosen delivery point, where I was able to retrieve the muffin without further difficulty.
This did not seem like a viable business—at least not yet. Robots are supposed to save human labor, and this robot wasn’t doing that.
But another sidewalk robot company seems to be much further along. Two weeks before my Howard visit, I drove to Fairfax, Virginia, where dozens of robots were delivering food on the campus of George Mason University. I got one of them to deliver me a donut without any difficulties—and I saw lots of robots cross streets unassisted.
The George Mason robots come from a startup called Starship. Ryan Tuohy, the company’s chief commercial officer, told me that the company has 2,000 robots in operation worldwide, up from 700 two years ago. And he said the company is planning for rapid growth in the coming months.
“Our unit economics”—that is, the per-delivery cost of running the service—“are now at the point where I'm allowed to expand as fast as we can,” Tuohy told me in a phone interview. “We already know what the next campuses are, what the next cities are, all this stuff. It's a matter of robots coming off the line and hiring people for these areas.”
After more than five years reporting on self-driving technologies, I’ve learned to be skeptical when companies tell me they’re on the verge of large-scale commercialization. But after seeing Tuohy’s robots in action, I believe him.
Over the next couple of years, I expect a lot more college campuses to feature Starship robots. And over the next decade or two, I expect technology like this will become widely available off-campus too.
The future of delivery is in Fairfax
Whenever possible, I like to try new technologies on my own, without pre-arranging the experience with the company. This gives me a more realistic picture of how a service works for regular customers.
I took this approach for both the Starship and Kiwibot services, and the results couldn’t have been more different. The Starship app was easy to use, and my robot arrived in about 10 minutes. In contrast, I found the Kiwibot app confusing, and I had to wait more than 30 minutes to get my muffin.
On the day I visited George Mason, a lot of people seemed to be getting robot deliveries. I spent time in a plaza at the north end of campus that featured a Panda Express, an Einstein Bros. Bagels, and a Manhattan Pizza. Each restaurant had a line of Starship robots waiting outside, and I saw employees come out to put food in the robots at least three times.
I talked to several George Mason students who received robot deliveries. One told me she used the Starship service regularly to order pizza or a burger for lunch. She said deliveries typically take 10 or 20 minutes except during the lunch rush, which can take longer.
On the other hand, I couldn’t find any evidence that people were using the robots at Howard. I talked to half a dozen Howard students, and none had used a delivery robot. A couple of students said they didn’t even realize it was an option.
Around lunchtime, I followed a couple of the Howard robots in hopes of seeing a delivery in action. But it soon became clear they were just driving around in circles.
Kiwibot did not respond to emails seeking comment for this story.
The long road to large-scale commercialization
Starship has been testing its robots here in the DC area for more than five years. In 2017, Starship launched a pilot project to deliver restaurant meals near Dupont Circle in partnership with Postmates. Back then, every robot had a person trailing behind, ready to answer questions or intervene if the robot got into trouble.
A lot changed over the last five years. The robots no longer have human followers. And at least in the US, Starship shifted its focus from cities to college campuses. There are also a lot more robots than there used to be.
In August 2019, Starship announced it completed 100,000 deliveries. It made its millionth delivery in January 2021, and the company expects to reach 4 million lifetime deliveries this month.
The robots “require less and less human intervention, and almost no remote human intervention,” Tuohy said. He added that Starship crosses “well over 100,000 roads per day,” and the “vast majority” occur without asking a remote operator for permission.
According to Tuohy, some campuses have no full-time Starship employees. “What we're trying to do is hire students with interest in autonomy and engineering,” Tuohy said. Part-time student workers plug the robots in at night, wipe them down, and then unplug them in the morning.
And soon, even that may be unnecessary. The company now has some “hubs where the robots can charge themselves and drive off in the morning with no human there,” Tuohy said. “We've got a remote-controlled door to secure access” to the charging area.
Starship has also been working to improve the physical reliability of its robots. Keeping a robot running 18 hours a day “is harder than it looks,” Tuohy said. “You don't know if a design spec is right or not until the robot is driving up and down hills. Then you figure out which of your little parts need to be swapped out for something a little more reliable.”
All of these incremental improvements mean a steady reduction in the per-delivery cost of operating the robots, putting the company on the path to profitable expansion.
Starship has also benefited from word-of-mouth among its restaurant partners, Tuohy said. “When we go to a new potential partner, they're able to speak with our existing partners. We benefit from the years of service and trust that we've built.”
Starship hasn’t gotten a lot of attention so far, perhaps because its robots seem too small and slow to be significant. But in addition to restaurant meals, people make a lot of short trips to convenience stores or bodegas to pick up a loaf of bread or a gallon of milk. In the long run, a lot of those short trips could be replaced by sidewalk delivery robots.
At a minimum, this seems like a big market opportunity for Starship. And if some of these deliveries replace car trips, it could also be good for pedestrian safety and the environment.
Lessons for larger robots
One reason I’m interested in Starship’s progress is that it may provide a preview of how self-driving companies with larger robots might develop in the future. You can think of Starship’s sidewalk robots as a smaller, slower cousin of the road-based delivery robots that companies like Nuro and Udelv are developing—or the robotaxis being developed by Waymo, Cruise, and other companies.
Because they operate on public streets, these robots have to move much faster than Starship’s robots. Larger size and higher speed make safety a much more pressing concern. If a Starship robot runs into you at its maximum speed of 4 mph, it probably won’t kill you or even cause a serious injury. Getting run over by a full-sized car from Waymo or Cruise—or even a smaller but still substantial Nuro robot—would be much more serious.
So the companies behind these larger robots need to be extremely cautious as they move toward commercialization. Waymo, widely viewed as the leading robotaxi company, has had an agonizingly slow rollout process over the last five years, gradually expanding its service footprint and reducing the level of human oversight. In March, Waymo announced that it would begin fully driverless testing in San Francisco—but that testing would exclude San Francisco’s dense downtown. Dense urban areas are especially challenging for software to navigate, but inconveniently, they are often where taxi passengers want to go.
In the last couple of years, I’ve heard people suggest that the slow progress of self-driving technology indicates that the technology is fundamentally flawed—or at least won’t be ready for commercial deployment for many years. But I think that’s the wrong lesson to draw.
The better lesson is that it simply takes time to validate and scale up a service like this. Starship is trying to solve a much simpler problem, technologically speaking, and it still took five years for it to reach the fairly modest scale it has achieved today. A big reason for this is that Starship is building a service, not a manufactured product like an iPhone.
Before Starship can expand to a new campus, it must do a significant amount of logistical work. It needs to get approval from local authorities, map the roads and sidewalks, hire local staff, and sign restaurant partners. It was slow going for the first few years, but it should get faster and faster as the company builds up its staff and technology infrastructure.
Most of the companies working on larger delivery and taxi robots haven’t even started this process—they’re still trying to build robots that can reliably get to their destination without killing anyone. Even industry leaders like Waymo are still operating their driverless services within tight geographic limits that may limit their commercial potential. Starship’s experience shows that even after a self-driving company gets its technology to work, it still takes years of additional work to turn it into a profitable business.
And one competitor that makers of larger robots may face is Starship itself. Tuohy told me that with the sidewalk robots reaching maturity, Starship is starting to think about other robots it might build—robots that are “maybe not sidewalk but probably land-based."
Tuohy wouldn’t say more than that, but it sounded to me like Starship is working on a Nuro-style delivery robot that operates on roads instead of sidewalks. If that’s true, Starship would be getting a late start. But if Starship’s sidewalk robots are as successful as I expect them to be, it could give Starship deep enough pockets—and a deep enough bench of engineering talent—to become a credible player in the road-based robot game.
I’d love to have a few more questions for my forthcoming mailbag post. If you’ve got a question to ask, please leave a comment here.
they’re still trying to build robots that can reliably get to their destination without killing anyone.
Got a good chuckle out of that one!
Because Chess computers beat humans, Go computers are close behind. Discuss.
If you don’t understand the mechanics, the algorithms, the limitations, the challenge, it can seem like one follows another inevitably.
Chess computers used brute force. Go required a completely different approach, neural networks and reinforcement learning.
Perhaps delivering pizzas in metro DC is a completely different challenge from other "similar" problems. I don’t know.