Seven years ago, Google started a project to pursue a futuristic idea: Develop cars with software so advanced that it could take over all the driving for humans. Today, not only are Google's fleet of self-driving cars navigating complex conditions in four U.S. cities, an entire industry has exploded with startups and automakers clamoring to develop the same technology.
"In 2009 we were just this crazy research project working on a science fiction idea," Dmitri Dolgov, head of Google's self-driving technology, told Fortune. "And look at where we are today. "
Google self-driving cars have logged 2 million fully-autonomous miles on public roads, 90% of which were on city streets, the company announced Wednesday. Considering the hours spent on the road, Google's cars now have the equivalent of 300 years of human driving experience.
And that's not counting the more than 3 million miles that self-driving car software "drives" every day in Google's advanced simulator.
Google's self-driving software is tested on modified Lexus RX450h SUVs as well as a prototype that it introduced in 2015 that is without pedals or a steering wheel, and instead relies on sensors and software. The prototype is equipped with a backup steering wheel and brakes when on public roads. Google hopes to commercialize its technology by 2020, giving the company a little more than three years to perfect its software, navigate complex and ever-evolving regulations, and launch its product.
What It Feels Like to Race in a Self-Driving Car:
Google solved the first 90% of the self-driving car problem within the first two years, Dolgov explained. It's that last 10% that Google has been spending most of its time on.
"It's where the core technical challenge lies," Dolgov said. To solve that last 10%, engineers have focused on advanced driving skills like understanding how to recognize and navigate construction zones, interpret hand signals from cyclists, and identify and react to emergency vehicles. The team is on a never-ending task of making its cars more human too.
Finally, the team is working on giving the cars extremely rare experiences—either through driving on public roads or in the simulator—so that the software is better prepared when odd situations pop up. For instance, Google self-driving cars have encountered ducks in a road, a woman in a wheelchair chasing a turkey, and more recently, a horse.
Get Data Sheet, Fortune’s technology newsletter
Other automakers and startups are also in a race to develop self-driving cars.
Tesla (tsla) and Uber as well as traditional automakers like Ford (f) , General Motors (gm), and Toyota (tm) are all develop self-driving technology. And then there's the growing collection of startups, notably Nutonomy, an autonomous vehicle software startup that has managed to beat them all to market with a self-driving taxi service in Singapore.
Each company's approach is different, and Dolgov is quick to note that what Google is doing is distinctly different from Tesla and other automakers. Tesla cars had logged more than 130 million miles on Autopilot, its semi-autonomous highway driving technology, by May 2016, the automaker has said in a blog post.
"There are miles and there are miles, there are hours and there are hours," Dolgov said. "Most of ours have been on city streets and this is a very rich environment compared to driving on freeways."
"It's kind of like playing a piano," he added. "You can play scales all day or you can performing the Flight of the Bumblebee. It's the same notes and mechanics, same actions, but you put them together in very different ways and one experience is more interesting than the other."