This is why self-driving cars suck at making unprotected left turns

Human or robot driver, left-hand turns in a car are tough. Unprotected turns? Even harder.

Cruise, the self-driving car company backed by General Motors, put out a video Thursday showing its self-driving electric Bolts making left turns all over San Francisco’s busy streets. The company says it makes 1,400 unprotected left turns every 24 hours.

For an autonomous car company it seems like this shouldn’t be hailed as such an achievement, but it is. Waymo, the Google self-driving car spinoff company, notoriously struggled in Phoenix to turn left during testing on public roads. Its Chrysler Pacifica minivans were just too timid to make it through the oncoming traffic. People have noted that the robo-taxis have improved in those situations.

In an emailed statement, Cruise president and CTO Kyle Vogt said, “In an unpredictable driving environment like SF, no two unprotected left-turns are alike. By safely executing 1,400 regularly, we generate enough data for our engineers to analyze and incorporate learnings into code they develop for other difficult maneuvers.” 

So while this is an important step for Cruise in its goal of launching a fleet of robo-taxis in San Francisco by the end of the year, it’s also a good reminder about the limits of autonomous vehicles.

Bob Leigh, senior market development director of autonomous systems at RTI, works with 40 different companies building autonomous vehicles of some sort: passenger vehicles, flying cars, hyperloops, and more. In a phone call Thursday he started off explaining that left-hand turns are just hard. Period.

Human drivers in the U.S. crash 10 times more making left-hand turns than right-hand turns. But the machines should be able to handle judging the risk, details, and surroundings required to successfully turn left. You can plug in a risk algorithm and have the car system scan for pedestrians,  the timing of oncoming traffic, and more. And yet, the machines still struggle with the maneuver. 

Ultimately it comes down to this: turning left on American roads is a very human, social move. It’s almost a negotiation, Leigh says. Drivers edge out into the lane, trying to assert themselves through. 

“A right-hand turn is a consistent maneuver,” Leigh explained. “A left-hand turn is a lot of variability.”

There’s pedestrians, oncoming traffic, and small gestures and nudges drivers make that are hard for a machine to emulate. It takes a lot of training, data, simulations, and iteration to teach an autonomous vehicle to balance a reckless, dangerous move barreling across oncoming traffic and staying timid and conservative and waiting for 10 minutes just to turn left. 

“That’s all communication,” Leigh said. And we know social interaction is not self-driving cars’ strongest suit – it’s a computer, after all. Cruise’s video shows it’s mastering what’s considered one of the hardest parts of autonomous driving in the U.S. “You can show a level of maturity,” Leigh said.

After all, we can’t all be like UPS, making only right-hand turns.