The newest member of Yelp’s security team wakes just after 8 pm, ready to begin its rounds.
It traverses the lobby, gliding over polished concrete toward a small recess in the corner, where it inspects the emergency exit tucked inside. Last year, burglars tried to breach the office by rending the door from the building, frame and all. The low-resolution camera mounted in the lobby saw nothing.
“We couldn’t see what was going on inside the alcove,” says Rick Lee, Yelp’s head of security, who joined me on my late-night visit to one of the company’s San Francisco offices. “But there was daylight and cold air coming through where they’d bent the door frame. A high definition camera would have spotted the light. A directional mic would have heard the noise. A FLIR infrared sensor would have flagged the temperature delta,” Lee says, with the punctilious air of a man who has spent more than a decade overseeing security at Silicon Valley firms like Uber, Apple, Google, and Amazon. “Cobalt has all three,” he says, gesturing toward his recent hire as it about-faces from the alcove and resumes its patrol. “It’s also mobile.”
Cobalt is one in a growing class of autonomous robots developed for spaces like malls, museums, and offices—the kinds of places that are more structured and less cluttered than, say, an apartment, but more dynamic and unpredictable than a warehouse or server room. Tug, an autonomous medical robot, delivers food and medications to hospital patients. Tally and Bossanova audit the shelves of grocery stores. And then there’s the Henn-na Hotel, an uncanny lodging in Japan’s Nagasaki Prefecture staffed entirely by androidal assistants. Experts agree that commercial bots will soon take over many of the world’s blue-collar, high-turnover jobs. Less clear is how these robots should look and act, and the role we meatbags will play as stewards to our robot replacements.
Cobalt, developed by the Palo Alto startup of the same name, distinguishes itself from its peers in both respects. Each of its robotic sentries stands 5-feet 1-inch tall, and tapers from a big, blue ovular foundation to an ivory-colored top the size and shape of a badminton racquet head. Tetris-ed between its base and its crown lie more than 60 sensors and the computer parts required to run them, along with (what else?) AI and the machine learning and computer-vision algorithms to recognize people, places, and, yeah, temperature deltas. “We have the equivalent of what’s running in an autonomous car running on this robot,” says Cobalt Robotics CEO Travis Deyle.
However—and this is significant—you don’t see most of this hardware. A soft, mesh, cobalt-blue fabric envelopes the robot, hiding it all. Cobalt designed Cobalt to look capable, not humanoid; authoritative, but nonthreatening. On this front, the company’s collaboration with Yves Béhar—the A-list industrial designer behind products like Samsung’s gorgeous new television The Frame and, more infamously, the Juicero juicer—is evident. Lee says it reminds him of a piece of Herman Miller furniture. “Classic and understated,” he says, as we watch it scoot around Yelp’s 17,000-square-foot office. “You notice it, but it doesn’t dominate the room.” To me, it looks like a traffic cone with fashion sense.
Either way, that congenial vibe is one of two big things differentiating Cobalt from other commercial robots—especially ones from Knightscope, another Silicon Valley startup and the current leader in developing security bots. Knightscope’s flagship, the K5, started patrolling malls, offices, and schools in 2015, and looks pretty intimidating when it’s not nose-diving into fountains. Aggressive, even—like a cross between a Dalek and Bullet Bill. The upshot is that Knightscope’s robot resembles hired muscle, while Cobalt Robotics’ looks like the kind of machine you might ask for directions.
The second thing Cobalt has going for it is telepresence. Most of the time, Cobalt patrols autonomously, detecting people, dodging obstacles, and flagging anomalies on the fly. But if it spots something unusual, it alerts a remote operator. The robot can then ask its human partner questions like “Is this a person?” to improve its computer-vision algorithms, or allow that carbon-based lifeform to assume control of its movements and display their face on its forward-mounted touchscreen.
This is exactly what happens around 8:30 pm, halfway through the patrol route. Cobalt spots a woman sitting quietly at the far end of the office, staring at a computer screen. The robot alerts Shiloh Nordby, Cobalt Robotics’ lead pilot, who is overseeing an undisclosed number of robots from the company’s headquarters in Palo Alto, some 35 miles away.
Lee and I watch from perhaps 40 feet away as Nordby takes control of the robot. He approaches the woman slowly, parking Cobalt a few feet to her right, just beyond her field of view.
“Hi there,” Nordby says. No response. Maybe she didn’t hear him.
“Hi there,” he says again, louder this time.
“Oh, hi! How are you,” the woman replies, turning quickly in her chair.
“Good, thanks. My name is Shiloh. I’m with Cobalt Robotics. Do you mind badging in on the screen?”
Nordby scoots the robot to within arm’s reach of the woman, who grabs her employee ID, raises it toward the screen, and pauses. She seems confused about where to put her badge.
“Oh, just place it right below the screen,” Nordby says.
“OK, right here? Oh. Oh! Wow, OK,” she says, as the screen gives the all-clear.
“Thanks,” Nordby replies. “Have a good night.”
“Thanks, you as well.”
“Thank you!” Lee shouts at the woman, alerting her to our presence.
“Oh, no problem,” she says. “That was awesome!”
Lee swears Cobalt’s badge-in interaction wasn’t staged. In fact, in the four months Cobalt has been monitoring Yelp’s offices, he’s never seen it do this in person. Lee occasionally pops in unannounced to join the robot on patrol, but he mostly keeps tabs on its performance through a weekly “instance report,” prepared by Cobalt Robotics. The dossier includes details like the number of people detected and authenticated, the robot’s schedule and patrol routes, and detailed descriptions of significant incidents from the preceding week.
Like a recent report of a rogue pair of microwave thieves. In timestamps, video stills, and transcribed conversations, Cobalt told Lee how it spied a couple of Yelp staffers schlepping an appliance from the office. Lee showed me a still of the video Cobalt captured of the incident. The pair did look suspicious, but they had merely burned something in the microwave and were moving it outside to air it out.
Lee says he’s quite satisfied with the ‘bot. “I wanted a super-capable machine, full of sensors, coupled with telepresence to help with customer service, creative problem solving, and negotiating all sorts of things people usually deal with every day,” he says. “So far, it’s functioning exactly the way I envisioned. It operates on its own, and the employees, once they’ve interacted with it a few times, barely notice it.”
Friendly—and subtle—human interaction is a bigger deal than you might think. Especially for robots, and especially right now. “I think that 20 years from now, robots moving among us will be very common—people aren’t even going to notice whether they’re there or not,” says roboticist Manuela Veloso, who oversees Carnegie Mellon University’s machine learning department. But first, she says, machines must earn humanity’s trust.
One way to do that is through something Veloso calls symbiotic autonomy, a relationship in which robots recognize their limitations and ask humans for assistance. Veloso coined the term to describe the behavior of semi-autonomous robots developed in her own lab, which can ask passersby for help with, say, riding an elevator. Cobalt works in a similar fashion, asking its remote operator to identify people and objects it doesn’t recognize, or bystanders to help open or close a door.
It bears mentioning that symbiotic autonomy is different from the concept of human-in-the-loop, whereby the human is the one asking the robot to do something. Cobalt uses both techniques in concert, to powerful effect. The exchange between Cobalt, Nordby, and the employee working late? Impressive. Nordby’s ability to not only control Cobalt but address the employee’s confusion on the robot’s behalf leant the whole exchange a surprisingly pleasant, human quality.
Which is ironic, because Cobalt seems purpose-built to eliminate human guards, save for a few remote operators. Deyle won’t tell me what his company’s services cost, but he says Yelp pays a fraction of what it would employing human guards. He puts the figure between $30,000 (the annual cost of monitoring an office full of security cameras) and $300,000 (the annual cost of a flesh-and-blood security detail).
It’s also not clear just how many robots a remote operator could realistically oversee. “The toughest thing today is quickly getting a sense of what’s going on around each robot, when switching between different units in the fleet,” Nordby says. Cobalt Robotics would not say how many robots Nordby manages on a given night, but it’s not unreasonable to assume the ratio of robots to pilots would increase with time, as the latter train the former to become increasingly self-sufficient.
But I’m getting ahead of myself. It’s early days for Cobalt, and the robot has yet to prove its worth in a range of security settings. Voloso, for her part, would like to see it navigate busier, less predictable environments. (Right now, Cobalt patrols only at night, when foot traffic is minimal, but the company’s promotional video shows the robot easily navigating a naturally illuminated office bustling with people.) “I’m skeptical, not doubtful,” she says. “Show me! When Cobalt’s robots have logged more than 1,000 kilometers without a collision, we’ll talk.”
Deyle says his company is working on that. And that promotional video? “It’s definitely representative of today’s capabilities,” he says, adding that, if it’s something customers want, Cobalt robots could begin daytime operations within six to 12 months.
I’ll believe that when I see it. But for now, color me optimistic. At three points in my visit, I stepped abruptly into the robot’s path, and it always avoided me with little fuss. At this small but significant task, at least, Cobalt has earned my trust.