Autonomous cars may be safer drivers than their human counterparts. But when it comes to who’s smarter on the streets, humans still have the edge—for now.
The past 100 years of automotive experience has given humans quite a head start in understanding the nuances of the road, from how to react to pedestrians and cyclists to how to navigate tricky four-way stops.
As the robo-vehicles continue logging real-world miles, they’ll eventually be able to make more of these decisions for themselves. Deep learning technology allows the cars to recall “memories” of similar situations in the past to inform their choices in the present.
But in order to reach that level of competence, they’ll need lots of experience. So, for now, they’re borrowing some of ours.
Self-driving companies are collecting data from a number of sources to help bridge the knowledge gap for their machines.
Google’s self-driving subsidiary Waymo and electric carmaker Tesla are leading the pack of information gatherers. But each company has its own school of thought on how to best teach our computers to drive. The type of data the machines receive may impact how they eventually maneuver the roads.
Way Mo’ Data than Anyone Else
Back when Waymo was part of Google, it got the jump on autonomous research years before the concept was trendy. So it’s no wonder the company has acquired more self-driving data than nearly anyone else.
Earlier this year, Waymo reported that its autonomous fleet had logged more than 5 million real-world miles. That topped the combined on-the-road experience of every other company testing in California in 2017.
And a spate of recent deals penned by the self-driving magnate will likely only further that total—in a hurry. Waymo’s fleet of autonomous Chrysler Pacifica minivans currently numbers around 600. But this January, the company announced it would expand its autonomous armada by “thousands” of vehicles. An additional boost will come from the 20,000 electric vehicles Waymo ordered from Jaguar late last month. Engineers plan to fit the fleet with their autonomous technology.
All that automotive power adds up to a ton of driving data. Each vehicle is fitted with eight cameras, five radar devices, and three different LiDAR sensors. All of those capture a vast amount of information that gets relayed back to Waymo.
Still, the Google sister company’s specialty is virtual driving, with computer simulations guiding its vehicles through more than 5 billion virtual miles since 2016. It amassed more than 2.7 billion miles last year alone.
The simulations are informed heavily by Waymo’s real-world choices. Data collected by the company’s physical fleet is used to help build full-scale computer models of the cities it’s testing in. The team then sends 25,000 “virtual cars” to cruise those digital streets every day.
Taking advantage of the virtual world’s infinite possibilities, engineers create all manner of strange situations to put the simulated rides through. The results are then uploaded directly back into the real cars navigating the roads. This completes the feedback loop and allows Waymo’s machines to generate millions of miles of “memories” for every inch of actual driving.
Still, there’s one company that may be able to give Waymo a run for its money, in both quantity and quality of data.
Waymo wasn’t the first to reach 5 billion miles—that honor went to competitor Tesla, who announced the milestone back in July of 2017. But unlike its computer-savvy competition, the electric carmaker has stuck to the physical world for its data collection.
The herculean effort stems from Tesla’s business model. It harvests information directly from the real-world rides of its customers. A car may belong to an individual, but the company still owns the data generated by its travels. It can retrieve the statistics remotely or when a car is taken in for a service.
Yet even when Autopilot is not engaged, Tesla can keep traveling along the learning curve. A so-called “shadow mode” feature of the system allows the company to track situations when Autopilot would have reacted differently from its human driver, and log what the computer would have done instead.
With more than 200,000 vehicles on the road in America alone as of this April, the company’s lucrative data returns will only continue to grow.
Still, there’s one area where Tesla is lacking. The company doesn’t believe in LiDAR, the laser-guided system that helps autonomous cars “see” where they’re going. In the past, CEO Elon Musk has called the technology “a crutch” and noted that it’s too bulky and expensive to incorporate on any mass-produced autonomous model.
Several deadly and dangerous accidents taking place in Autopilot-guided cars have also failed to change Musk’s mind on the issue. But the CEO is far in the minority in his belief, with many in the self-driving world believing LiDAR is essential for autonomous driving.
Whether or not the gamble will pay off for Tesla is ultimately a matter of time—a concept that even all the data crunching in the world could fail to accurately predict.