DEARBORN, MI — Riding along in the back seat of Ford’s self-driving car is an uneventful experience. The car steers itself along straight and curved roads, heeds stops signs and traffic lights, yields to oncoming traffic when turning left, and waits for pedestrians to cross the road. It waits longer for the road to clear than virtually any driver you know, which may be part of its ability to drive autonomously in 2016: Ford’s self-driving car is like being driven by your overly cautious grandfather, but with better reflexes.
Even this brief demonstration suggests Ford and other automakers are not blowing smoke when they say they’ll have truly self-driving cars within five years. One open area for improvement by 2021 may be the ability to deal with the outlier exceptions, such as a child chasing a ball into the street. In my brief test drive, a car coming out of a side street and turning more or less into our path made Ford’s test driver choose to take over control briefly.
I rode in a Ford Fusion heavily modified, to say the least. Multiple Velodyne lidar devices (above) sat on a roof bar. Long-range radars faced front and rear, while short-range radars faced left and right (image below). The car also had a long-range monochrome camera facing forward and a shorter-range stereo color camera to track traffic signals and also provide short-range distance information to supplement the radar. The car retains the driver controls, and it’s crewed by a test driver behind the wheel, ready to take over; a second engineer rides shotgun and verifies all systems are working.
The loop we drove was 1.8 miles on and around Ford’s suburban Dearborn campus. It took about 10 minutes. It has been mapped in advance by 3D cameras to provide an accurate guide to the roads, structures by the roadside, stop signs, crosswalks, concrete medians, side streets, and driveways. Without this, confident self driving wouldn’t be possible. The roads are so well-mapped that lane and road-edge markings aren’t really necessary. This is the exact opposite of what’s required for lane departure warning and lane centering assist to work on today’s cars that are modestly self-driving — they can pace the car in front via adaptive cruise control, stay centered in the lane, and via blind spot detection warn of cars nearby when you’re about to change lanes.
Our car did a great job maintaining its position on the road despite the lack of lane markings. I was surprised how quickly it accelerated away from stop signs and traffic lights: no burned rubber, but certainly a brisk acceleration. If the speed limit said 25 mph, the car drove at exactly 25 mph. That suggests that on the interstate, if you don’t have option “go with the traffic flow,” you’ll be passed by most everyone on the road and occasionally flipped off by drivers with places to be.
There were two other surprises: Our test car was painfully slow in starting up again. Once a pedestrian clears the crosswalk, the car waited for about 4 seconds before starting up. Ditto for a stop sign where there’s no one around. That’s one algorithm I’d tweak in a heartbeat. The car also anticipated a pedestrian who’s about to step into the crosswalk and waited. It’s possible a lot of drivers would figure they could be through the intersection long before the pedestrian made his or her first step into the crosswalk.
The most complex maneuver was a left turn against oncoming traffic. We got to the traffic light, a couple oncoming cars turned left, and the road was clear for — it seemed to me — a couple hundred yards ahead before the next oncoming car would reach the intersection. Our car was programmed to wait, and wait, even though the oncoming traffic took, by my thousand-one count, 8-9 seconds to reach the intersection. That traffic cleared, and then we made the turn. Obviously, Ford and others working on self-driving have a couple years to fine-tune the car’s sense of adventure.
In making the left turn, we carved a slow, precise arc. We stayed in lane, and didn’t come close to clipping the lane markings in an effort to shave a few yards off the arc. On that, the car is better than most human drivers. Once the wheels were straight, the car accelerated briskly, just enough to make up for the lost time waiting for oncoming traffic.
Urban and suburban driving involves dealing with others who misjudge timing, or cut in front of you, or do the thousand other quirky things that are part of the driver’s experience. In our case, we were cruising at 25 mph and a car was about to pull out of a side street on our left. The driver should have waited for us to go past. Instead, the driver pulled out and started turning left into the lane to our left. (It was a multi-lanes-each-way road.) The car chirped several times and the driver in our car took over to ease us away from the possible incident. Our car probably would have dealt with the problem autonomously, our driver said, but better safe than sorry.
Based on this brief drive, it’s possible to believe Ford and other companies have a shot at making cars be self-driving in five years. It’s going to require smaller cheaper sensors and ever-more processing power to deal with the 3 million bits of data generated each second as the car rolls along. All the roads must be precisely mapped. One reason Uber is starting a self-driving project in Pittsburgh is that it has, or will have, those streets mapped. Uber may have to decline trips that go outside the mapped geofences, or have the driver — there will be one — take over once the car reaches the mapped limits.
Five years to a self-driving car may actually mean 3-4 years until designs are locked, and then another 1-2 years of adaptation and fitting the pieces into the vehicles actually put on sale. It’s a short time. It may be doable.