Eighteen months ago, Tesla Motors embarked on an initiative that would sow the seeds for its self-driving car project.
The Elon Musk-led automobile company wanted to give its vehicles the ability to drive themselves, but it didn’t just want to introduce the technology without thorough testing, Sterling Anderson, Tesla’s director of autopilot programs, explained on Tuesday at an MIT Tech Review technology conference in San Francisco.
So Tesla began to “quietly” install equipment into its existing Model S cars that would let Tesla engineers slowly make software updates over time, Anderson said. These software updates would take advantage of the newly installed 12 ultrasonic sensors embedded in the vehicle, forward-facing radar system, and GPS hardware.
This hardware and software combination would also give Tesla the the ability to pull “high resolution” data from cars, Anderson said. The company could then analyze the data and learn how driver use their cars and how they react to the features Tesla steadily rolled out over the next 18 months like its cruise control, collision warning update, and its auto parallel parking feature.
“We were behind the curve,” Anderson acknowledged compared with other companies like Google
that were leading the charge in autonomous driving technology.
Get Data Sheet, Fortune’s technology newsletter.
With the upgrades, Anderson now believes Tesla is at the forefront, although that is open to debate.
“We have found that there is no substitute for empirical real world data,” Anderson said, saying his company gets data from real-life situations rather than experiments—a subtle dig at Google, which is testing its own self-driving cars.
Anderson said that Tesla vehicles with installed self-driving hardware have driven roughly 780 million miles since October 2014. Those vehicles drive on average, 2.26 million miles per day, he said. Over the last seven months, Tesla cars with the autopilot feature turned on have driven around 100 million miles, he continued.
Tesla has also learned from its driver data that its vehicles, when on autopilot, can drive better than their human operators in some situations. For example, Tesla cars in autopilot mode tend to better stay in the center of road lanes than humans, who tend to veer too far to the right or left.
That information will be helpful to Tesla as it continues to “refine the features” of autopilot, Anderson said. Still, Tesla has a lot more to learn about driver habits as it continues to improve its software. Musk told Fortune in December that it may be two years before Tesla has fully autonomous cars. Each software update essentially represents a step to a complete, autonomous car.
For more about Tesla, watch:
One of the biggest challenges facing the company is the having to introduce its software updates and new features worldwide. However, driving laws and road markings may be different from country to country, which makes it difficult for a one-size-fits all autopilot feature.
“It is easy to do a demo on a ten mile stretch of road that will wow spectators,” said Anderson. “It is enormously more difficult to robustly capture all the use cases in the world.”
Story updated to clarify that the Tesla vehicles drive on average 2.6 million miles per day.