It's tricky to test self-driving cars. Even if you have hundreds of thousands of miles under your belt, it's still difficult to account for every possible real-world peril. Researchers think they can fast-track that experience, however. They've developed a sped-up testing process that should accomplish a lot in just a small amount of time. Instead of a holistic approach that gauges everything at once (and often goes for miles without a meaningful event), the new method breaks things down into individual components you can test frequently and repeatedly in simulations. If you want to gauge the car's reaction to someone cutting you off, for instance, you just focus on that -- you use stat analysis to determine how the car would behave in "boring" moments.
The team's own experiments also limited their metrics to the likelihoods of crashes, injuries (including severity) and "conflict events."
Based on estimates, the improvements would be dramatic... to put it mildly. Researchers believe that 1,000 miles under their method could be equivalent to between 300,000 and 100 million miles of real-life driving. You could match Waymo's yearly driving experience (635,000 miles in 2016) in the space of a day.
Just don't get too excited. The researchers are aware that they need to account for many, many more situations before this testing method is ready for practical use. How does a self-driving car handle jaywalkers, overloaded trucks or snow-covered streets? And of course, there's no guarantee that you'll see such a massive improvement with every car or every test. If the real results come even vaguely close to this, though, the automotive industry could be in for a shakeup. You may still end up waiting several years or more for viable driverless cars (there's still issues like regulation to consider), but they could be genuinely road-ready when they arrive. You wouldn't have to worry quite so much about your ride going awry due to limited testing, and it could be much easier to iron out whatever glitches are left.