Dr. Chris Urmson is driven by an urgent motivation: his pre-teen son and the dangers of driving. The director of Google's Self-Driving Car program has four years to make the tech company's autonomous car viable before his eldest child can apply for a license. And his team of engineers has made incredible progress since Dr. Urmson joined Google nearly seven years ago.
To demonstrate, a junior member of his engineering team guides the pair of us to The Prototype, a small pod-shaped self-driving car, and beckons us into the only seats.
There's no steering wheel, no pedals, no instrument panel or gauges. What would normally be the driver seat doesn't even scoot fore and aft because, well, it doesn't have to. The Prototype isn't about form, it's about function. It's a platform for Google's software, not a design statement.
Remain Seated, Please
"Please fasten your safety belts and keep your hands and arms inside the vehicle at all times," says the engineer. She closes the door and waves to us from outside. After a few moments, this curious electric-powered car begins moving on its own.
If this sounds like a ride at Disneyland, that's because The Prototype comes across as a trackless version of the old PeopleMover. We're even riding around on the roof of a building, the unoccupied top floor of one of Google's many parking structures in Mountain View, California.
As it follows its predetermined course between unoccupied parking places, the pod car negotiates corners, accelerates (up to a point) and slows down, all the while giving the impression that it knows exactly where it is in the world. Along the way it encounters "unexpected obstacle" scenarios in the form of a pedestrian walking across the road, a car pulling into its path and a meandering bicyclist.
None of them results in abrupt braking or a swerving maneuver because the car's many sensors detect the objects and predict their likely paths well in advance. It also helps that we're not moving faster than 15 mph.
This is all well and good, but the cutesy styling and the contrived setting makes it feel like a carnival ride, not transportation. For that we need to wade out into real traffic, in which case current regulations require a human driver to be present and ready to take over in case something goes haywire.
Taking It to the Streets
The pod cars have lights and mirrors, and a steering wheel and pedals can be bolted in rather quickly, but Google also has a fleet of self-driving Lexus RX 450h SUVs. They chose the Lexus because the body's sloping rear hatch gives the rooftop sensor array a good all-around view.
Google's self-driving car development center abuts a suburban neighborhood. Our driver pulls his hands and feet away from the SUV's controls and presses the "Go" button. Within 50 yards we're all passengers as the car moves along between single-family tract homes, with cars parked parallel along both sides of the typical residential street.
Our vehicle makes the next left and we're confronted with an elementary school that has just let out, the typical end-of-day chaos in full swing. Minivans are lined up with doors slid open, parents and kids are walking the sidewalks and a crossing guard is on duty.
This had better work.
How It Does What It Does
The car knows where it is and how it's moving through the world based on GPS satellite data and an inertial measurement system that consists of gyroscopes and accelerometers. But GPS isn't accurate enough to pinpoint a location to fractions of an inch, so the car's software brain compares live camera data with an onboard map to zero in on its precise location.
This map is nothing like an ordinary Google Maps or Street View dataset. The self-driving car needs a far more detailed map based on rigorous surveys that accurately locate every curb, every lane line, every crosswalk, along with road elevation and banking. It must also know the locations of fixed objects like signals, road signs, traffic islands, driveways and bus stops.
Transient objects like cars, people, bicycles, animals — even traffic cones — are tracked and classified by a system of cameras and an array of laser and radar detectors, and they can follow hundreds of them at once. Once tracked and classified, the car must then predict what will happen in the future in order to identify and react to specific objects that may intersect with its intended path.
The end result boils down to steering, throttle and brake commands. But like all driving, this is a continuous process in which scanning, identification, planning, decision-making and execution are ongoing continuously and simultaneously.
Traffic ahead is inching forward, bottled up behind a line of vehicles waiting for someone to turn left into the crowded school parking lot. Our car stops just shy of a small intersection. The car has determined that only 80 percent of itself can fit on the other side of the intersection. Not enough to continue.
Other drivers wishing to turn right from the side think we're letting them in. But once they proceed there's even less room for us to fully clear the intersection. So we wait. The scenario repeats. And repeats. We don't move for a full 90 seconds while four cars turn right and slide in ahead of us. We can only imagine what the drivers behind us are thinking.
Farther on, the crossing guard raises his stop sign and kids step off the curb. Our car has been programmed to identify handheld signs and halt gestures, so it stops. An audible announcement confirms it has seen the temporary stop sign. The car moves on after the intersection clears and the guard drops the still-visible but now upside-down sign slack by his knee.
There's a badly parked trailer around the next corner, and a kid in the middle of the street on a skateboard farther on. The Google Lexus pauses and moves forward haltingly as it reacts to these unexpected encounters. The braking isn't always as smooth as it would be under human control, but it is appropriate.
We finally come to a larger arterial street and a red light. The car wants to turn left, but the cars going straight have all but covered the turn-lane entrance. A car the size of ours could squeeze through, but it'd be a close shave and the left rear tire might nick the painted line. The Google car waits instead, deferring to the safer course of action.
This is always its default course of action. After all, Google is attempting to erase car accidents altogether so there's no room for error.
Are We There Yet?
It's an impressive display, but there's still much work to do. For one, the mapping required is painstaking and complex. The Lexus can only go off-leash on a certain tiny subset of roads that have been carefully measured and surveyed.
And then there's the weather. Some of the sensors can see through fog and snow, but others can't. And the system may need to respond differently in places where snowplows don't strictly follow the lane lines. Very little testing has been carried out in such environments.
Google prototypes operate independently, but vehicle-to-vehicle and vehicle-to-roadway communication protocols (which do exist) are necessary to leverage the full potential of the technology. A "hive-mind" system could enable shorter following distances that safely allow more cars to occupy a given mile, and at higher speeds.
Google's fleet has accumulated 1.2 million miles of self-driving experience, much of which has been fed back into new software releases. That may sound like a lot, but that's 40 years at 30,000 miles. One of our more experienced staffers figures he's driven more than that by himself. It's not really that much.
Still off in the Future
Many experts feel the most likely application will be a sometimes-autonomous system that can be switched on during dreary commute traffic or long interstate drives. Tesla is working on such an autopilot-style system. After all, some people like to drive and it's hard to imagine every single mile of road and driveway in this vast country ever being mapped to a sufficient degree for full autonomy everywhere.
Google's engineering team will counter by saying there's uncertainty involved in getting the driver back in the loop — woken up and fully alert — when the need arises. They have chosen the fully autonomous route because they feel it's the only way around this. Also, they feel that going all-in is the best way to move the development process forward quickly.
Will Dr. Urmson reach his goals in four years? It doesn't seem likely. There's still much to be done, and Google hasn't yet determined how it will bring the self-driver to market when it is ready. It has not yet signed any automotive partners, but it did hire John Krafcik, former CEO at Hyundai and a veteran of Ford engineering, to lead the group.
Not to mention there's a long list of competitors also pursuing some form of this revolutionary technology, including Apple, Tesla and essentially every traditional automaker. It's an arms race of sorts. And Google, from what we've seen and sampled, is winning.