A white Lexus cruised along a road near the Google campus, braking for pedestrians and scooting over in its lane to give bicyclists ample space.
The car eased into a turn lane, waited for a green light and a break in traffic, then continued on its way in the Silicon Valley city of Mountain View.
It even avoided stopping on train tracks.
But there was nobody holding the wheel. What looked like the work of a conscientious driver was a Google car making all the moves -- with an AFP reporter in the back seat.
Google used machine learning to teach cars how people drive and, from there, to anticipate what motorists in surrounding traffic are likely to do.
"Computers have really good reaction times. They don't get distracted, drowsy, fall asleep, and they don't drive drunk," Google self-driving car software team lead Dmitri Dolgov told reporters getting an intimate look at prototypes at the Computer History Museum.
"They don't need to stop messing with the radio to see what is happening, or even take time to move a foot from the gas pedal to the brake."
The bustling street crowd paid little heed to the self-driving car, which sported a whirling gadget on top about the size and shape of a large coffee can.
The roof-top device used radar and lasers to track everything around it.
A camera peeking out from the Lexus front grill watched what was ahead.
Data is processed by onboard computers programmed to simulate what a careful driver would do, but at super-human speeds. And, naturally, the Google autonomous car was connected to the Internet.
A "Googler" from the technology titan's test driving team had a laptop computer that showed what the car "saw" -- everything from cyclists and traffic signals to orange cones and painted lines in the street.
Another Googler was in the driver's seat, ready to take over in the unlikely chance a human was needed to make a driving decision.
A red button could be hit to grab control from the computer. A tap of the brake would do the same.