Visitors to Columbus, Ohio, have a new way to see the city’s downtown attractions. A fleet of electric, self-driving vehicles now shuttle passengers around a cluster of museums and parks, using sensors and software in lieu of engines and auto parts. The pilot project, which began in mid-December, belongs to a larger statewide effort to improve road safety and mobility in this car-dependent capital.

“What we’re looking at is, how do we apply technology to improve people’s lives in a transportation context?” says Jordan Davis, director of Smart Columbus, which spearheads the fleet project. “We want to keep stretching the technology of self-driving vehicles to solve real use cases in our communities.”

Smart Columbus, launched in 2016 after the city bested 77 mid-sized U.S. cities for a pool of “smart transportation” funding. The U.S. Department of Transportation granted Columbus up to $40 million, while Vulcan Inc., the private company of late Microsoft co-founder Paul Allen, will chip in up to $10 million. The program aims to develop advanced data and intelligent systems technologies to solve problems plaguing many urban transportation networks: traffic congestion, accidents, tailpipe pollution, commuter delays and inaccessibility.

On a January afternoon in downtown Columbus, three green-and-white shuttles loop around a 2.4-kilometer circuit, past a handful of skyscrapers and along the snaking Scioto River. Three more shuttles sit in a parking facility, charging their batteries or waiting their turn in the rotation. Michigan startup May Mobility operates the vehicles, which can seat six people and travel up to 40 kilometers per hour.

Jhane Gaines, a shuttle attendant, sits up front behind a digital dashboard. A T-bar for manual steering rests near her lap, along with a control panel of push buttons and an emergency hand brake. The sky is clear, but the sidewalks are still buried beneath the previous day’s snowfall. As a precaution, none of the shuttles are operating autonomously when I visit. Should the snow return, the sensors could interpret the flakes as obstructions and stop the vehicle, Gaines says as she drives.

“It’s still early, and we’re still learning and observing,” Zafar Razzacki, head of product at May Mobility, tells me earlier by phone. “With any self-driving system, how the system manages environmental changes, changes in precipitation, et cetera, those are things we’re watching very closely.”

Still, Gaines says “it’s the coolest thing” when she can sit back and let the shuttle drive itself. She especially likes when the vehicle turns down a side street, performs a U-turn, then turns left and continues along the circuit. “It does everything. I don’t have to do anything,” she marvels, though she admits it was “nerve-racking” at first to experience.

Razzacki says the “secret sauce” behind May Mobility’s self-driving shuttles is the software, which is built on a proprietary set of algorithms that the company calls “Multi-Policy Decision Making.” The work began in the lab of CEO and founder Edwin Olson, a robotics professor at the University of Michigan in Ann Arbor.

“The common approach to artificial intelligence is to train the system by feeding it tons and tons of data that represents many different scenarios, and then try to teach the system to react to those scenarios,” Razzacki says. By contrast, May Mobility’s software “is designed to understand situations at a much more granular level and actually be able to predict what’s happening on an agent-by-agent basis. Instead of ‘recognize and react’ we like to ‘understand and plan.’”

May Mobility equips its shuttles with a combination of cameras, radar and multiple LIDAR (light detection and ranging) modalities, which gives the system 360-degree vision at the high-resolution, mid-range level and up close, for rapid response. “We’ve tried to come up with the hardware stack that’s reasonable to put out on the road but not prohibitively expensive,” Razzacki says.

The startup operates another self-driving shuttle service in Detroit, which launched last summer but isn’t open to the public. May Mobility has plans for a third service in Grand Rapids, Michigan, and several yet-to-be-announced locations.

Separately, Smart Columbus is planning a second self-driving shuttle route in Linden, a low-income neighborhood with a dearth of public transportation options. Davis says the city will begin soliciting proposals from developers later this month. The goal is to connect daily commuters to bus stops and other transit hubs, closing what’s known as the “first-mile/last-mile” access gap.

Asked why Columbus is opting to test self-driving shuttles, rather than add more people-driven buses in Linden, Davis says the city’s ultimate objective is to help pioneer technologies that—through improved connectivity, awareness and sensitivity—can drastically reduce traffic-related fatalities and serious injuries. “Safety is a primary long-term hope for the technology,” she says.

Razzacki says there’s another potential benefit to replacing humans with software, though bus drivers won’t like the sound of it. “The human is actually one of the most expensive pieces of the equation,” he says. “Once we’re truly able to pull drivers out of these vehicles, it will be much more economical than other types of solutions that [Columbus] is using today.”