So far, the ordinary-looking SUV had traversed the United States without incident. It had gone through tunnels and under overpasses. It circled roundabouts and stopped for traffic lights. Now, on the last day of a scheduled nine-day journey, it was poised to become the first autonomous car ever to complete a coast-to-coast road trip. First, it needed to contend with the Girard Point Bridge.
Riding in a rear seat, "I saw that bridge coming, and I thought, 'Oh my gosh, this is going to be a grab-the-wheel moment," said Kathy Winter, vice president of software at Delphi Automotive.
The car, an unassuming Audi SQ5 nicknamed Roadrunner, had been well-tested. Back in January, a few inebriated pedestrians fell flat in front of the car during a demonstration in Las Vegas. It was the quintessential worst-case scenario, and the car admirably hit the brakes.
More than drunken louts, bridges present a sophisticated challenge for the six radar sensors that feed data to the car's internal processors. Instead of sensing solid objects, radar sensors can read the alternating bursts of steel beams and empty space as conflicting information. "They're a radar engineer's worst nightmare," said Jeff Owens, Delphi's chief technology officer.
Girard Point Bridge, a blue skeleton of girded steel that spans the Schuylkill River, might be a bigger challenge than most. Traveling across the lower level of its double decks, the autonomous car's radar sensors had to discern between two full sets of trusses. Cross the Schuylkill, and Delphi's engineers felt confident they'd reach their destination: the New York Auto Show. For now, the sternest test of the trip lay directly in front of them.
A Data-Mining Adventure
Until that point, the toughest part of the journey had been finding an open gas station in El Paso, TX.
Trust in the technology had already been established. The main reason Delphi set out on the cross-country venture with a team of six certified drivers and two support vehicles was to capture reams of data. What better way to do that than dusting off the classic American road trip and dragging it into the 21st century?
They did exactly that, capturing three terabytes worth of data across 3,400 miles and 15 states. Autonomous systems controlled the driving for more than 98 percent of the trip, with engineers manning the controls for on-and-off ramps and as a precautionary measure in construction zones. Every time they touched the controls, others in the car logged notations.
Delphi has more than a half-century's worth of experience with some underpinnings of the technology. The company was the first to install cruise control on a production vehicle – the 1959 Cadillac Eldorado – and Delphi has worked on radar sensors since a partnership with Hughes Aircraft materialized in 1990.
What's new for Delphi are the software algorithms that cull information from 20 sensors aboard the autonomous vehicle and, within milliseconds, synthesize them into driving decisions. Each of the four cars in Delphi's global autonomous fleet contain six long-range radars, four short-range radars, three vision-based cameras, six lidar sensors and one localization systems. They're brought together and processed in a multi-domain controller that Owens calls "the greatest change in automobile architecture in 20 years."
Car manufacturers have added features controlled by software over the past two decades; they've added the processing power behind them on a piecemeal basis. The multi-domain controller essentially gives automakers a fresh start, centralizing processing power in one place. This is not only important for the autonomous cars of tomorrow, but also beneficial for the advanced safety systems already making real-time decisions on cars today.
"You can add things later and ask it to do more things, because it has that capability," Owens said. "Today, you really can't do that, because each of the individual components are designed for just that job. ... If your long view is that you're going to add more and more automation, this is a way to set your architecture for the next ten years."
The purpose of the cross-country trip was to collect reams of data on how all these systems converged. Delphi knew the car could perform on well-worn autonomous routes already mapped in Silicon Valley and Las Vegas. After successful demonstrations at the Consumer Electronics Show in January, top officials decided it was time to venture into unknown territory.
Technology Outpaces Legal Infrastructure
If the car was ready for the road, it was uncertain whether the road was ready for the car.
California and Nevada both have motor-vehicle laws that govern the use of autonomous cars on their roadways. Only two others state, Florida and Michigan, have similar laws, but neither appeared on Delphi's cross-country itinerary. Before the car entered its swing across the South, the company needed to get permission to pass through each state. That wasn't easy. While autonomous technology might be understood throughout the auto industry, wide swaths of the American public have yet to get a glimpse at its potential.
"A few of the states said, 'We're fine with you driving through but could you keep a hand on the wheel,'" said Winter, noting such interference would, of course, defeat the trip's purpose. With a deeper explanation, they earned permission to pass through via autonomous driving.
Over the next two decades, autonomous cars are expected to gain traction with mainstream car buyers. Ernst & Young's global automotive practice projects driverless cars will comprise four percent of global vehicle sales by 2025. That share will skyrocket to 75 percent by 2035.
Legal infrastructure to support autonomous driving currently lags behind the technology, but a transition is underway. Four states already have autonomous-driving laws on the books, and Owens expects the National Highway Traffic Safety Administration will soon grade semi-autonomous features as part of its five-star safety ratings.
The car's transition is also a gradual one. Owens does not believe there will be a watershed moment when autonomous cars suddenly take over the road. Rather, that a subtle transition that's already occurring with the influx of features like adaptive cruise control and active lane assist will continue.
Toward that end, Delphi wants its autonomous car to look like an production-ready vehicle. No lidar sensors are affixed to its roof; they're cleverly hidden within the SQ5's existing design. So are radar sensors, installed behind plastic body panels. The mess of computers and wires that fill the trunks of many experimental vehicles are carefully stowed in the space otherwise reserved for the spare tire.
Passing motorists would be hard-pressed to detect anything unusual about the vehicle. Of course, there's one big exception. "They're looking at us," Winter said, "and nobody's got their hands on the wheel."
Reaching The Final Destination
Crossing the Girard Point Bridge was anti-climactic. What appeared as a potential "grab-the-wheel moment" was not one at all, Winter said. The car drive down the middle part of the middle northbound lane, and the final roadblock had passed.
From there, Roadrunner cruised through New Jersey and into Manhattan, where it sat on display on the third floor of the Waldorf Astoria, still sporting dust from the Texas prairie and a week's worth of splattered bugs across its grille.
As competitors from Google to Mercedes-Benz all develop autonomous cars, this was a crowning moment for Delphi and engineering teams that came from four company offices located across the country.
In reality, the trip is in the rear-view mirror and the hard work now begins. The teams will spend week and month sifting through three terabytes worth of data and searching for anomalies. One area already up for study is occasional quirks in the camera system that occur when the sun-angle is low. Another that came a surprise during the road trip was the way the car behaved around trucks. "The car was scared around of tractor trailers," Owens said. "The car edged to the left just a little bit when it would pass trucks, and that was an interesting observation."
In that way, the autonomous car acts a lot like the human drivers they'll someday supplant.