California joining Nevada with frameworks for allowing driverless cars on the state’s roads got me thinking about how this all evolves. My assumption is that driverless car safety increases as the density of driverless cars increase. That is to say, when we’re 100% driverless cars there should rarely be any accidents because the cars can coordinate with one another to prevent such mishaps. However up until that point, the driverless car OS needs to reactively deal with human drivers in other cars which leaves a lot more room for the unknown.
The onboard computer needs to account for other cars it believes could cause an accident. The most basic part of this is reactive to an immediate threat — a car cuts you off or breaks suddenly and your car needs to stop suddenly less it hit the other vehicle. The next wave of intelligence emerges from analyzing available realtime information from the car’s proximity to anticipate potential threats. A speeding car coming up on the left could cause your car to push to the right hand side of its lane, giving a few extra inches to the passing auto. We probably do things like this unconsciously while driving and the computer needs this sort of AI too. But what if we could go beyond what a driver was doing right now and understand the history of how he drives generally or reacts in certain situations. Then the computer could create a predictive model of what was likely to happen given the specific drivers (well, vehicles) on the road around them at any time. A database of intention for cars instead of people.
Here’s how it could work: the driverless cars basically pass data back and forth and into a centralized db. Imagine that a driverless car is at all times gathering information about the cars around itself. Additionally it’s using OCR to identify cars by license plate and record information about those vehicles – speed, distance it keeps from other cars and so on. You can essentially build a compendium of driving habits by car using license plat as the unique identifier. For example, your driverless car “sees” that plate number CSR4408 just pulled 20 feet behind you. It does a lookup into the car database and find that this driver is very aggressive, which increases the chance of an accident. So your car moves one lane over and lets it pass rather than risk a tailgater.
Would there be privacy implications here? Certainly you are allowed to observe other drivers and make judgments about their safety. Our car’s software today is able to gather enough information about your historical speeds to already fill in some of this information. Of course some cars are driven by more than once person but unless you start doing facial recognition too, one would assume the fingerprint would match the car and not the driver. When a car is sold, the record would be cleared & restarted since there’s a different driver (this wouldn’t be a problem since car sales are recorded with the DMV as change of ownership). And I guess opting out means obscuring your license plate :)
[I should note that although I work for Google, I have no knowledge of our driverless car project outside of what has been discussed publicly and the speculation above is merely forward looking]