Think it’s easy to write car driving AI? Watch this epic autonomous car pile up (in GTA5) until you’ve changed your mind.
It’s a damn hard problem.
I have been meaning to write about the results of the 2015 Simulated Car Racing contest since, well, 2015. But entering the event was so taxing that I’ve not really been able to do anything too special with it. I did win 100 Australian dollars as the 3rd place entry so that was nice. Woo hoo! But I’m somewhat burned out on it. It was a profound learning experience.
Even with polite environments in closed race track conditions, getting a car to not do very stupid things is quite tricky. Throw in real world street driving issues and it becomes orders of magnitude harder. Throw in trying to replicate how the human mind turns absorbed light into conceptual models of the world and the problem gets way more complicated yet. One of the best examples of that is this interesting video of the Hyundai autonomous car having trouble with rain.
I’ve been a big proponent of the idea of autonomous cars since I first realized they were completely feasible after hearing a talk by one of the DARPA entrants in 2006. What I mean is that we should have all stopped driving in 2006. The technology problems were all 100% solved. The computer technology was 100% ready. But of course we didn’t stop driving. I believed then and I still do that the DARPA challenge was far more challenging of a test than the minimum level of robot sophistication necessary for relieving humans of the soul-destroying chore of driving cars.
The discrepancy is a major difference of opinion I have with the entire field of autonomous vehicles. The conventional wisdom is that for a car to take over driving duties, it must be able to handle all of the situations that a human can handle as well or better than a human. I think this mentality is tragically wrong.
For some reason that I can not understand, people working on the problem of autonomous vehicles seem to believe that the ideal environment for autonomous vehicles is the exact same environment which is ideal for humans. If they don’t believe this truly stupid thing then they believe that it is completely infeasible to transition the driving environment from one that is idealized for humans to one that is favorable to robots (or both). They believe the road infrastructure is utterly immutable and any progress on autonomous vehicles must work within existing (i.e. 1950s) infrastructure and imitate human drivers in all ways. I think this is tragically wrong.
Humans have proven several times (roads, canals, railroads, freeways) that deploying infrastructure improvements on a massive scale is completely feasible if the pay off is sufficient. It is a failure of imagination to misunderstand the profound benefit to eradicating our modern civilization’s last real source of mass toil and horrific routine physical danger.
My experience with designing my own autonomous race cars and my close attention to the topic of autonomous vehicles have emboldened me to make some radical predictions. And they’re not good.
Autonomous cars are not coming soon.
I predict that I will not take a ride in a fully autonomous car that is generally available to the public until at least 2030. 2040 would not be the least bit shocking. Because the necessary technology to replace my manual labor in automobile transportation existed in 2006, it won’t go much longer than that, but there will be some surprises for groups like Google. The only reason we will ever escape the scourge of driving is that special niche applications will slowly move along a more sensible path. At first it will probably be simply limited driving scope. For example, long haul trucking on freeways is such an obvious application that it is eerie that it is not already saving someone money (here are some recent developments on the topic). The adaptive cruise control as found on Teslas and other cars is an example of severely limited driving scope with the appearance of progress toward a completely replaced driver.
Some of these projects and driving aids are nice and generally welcome but I think real progress will only come when people start to make the driving environment more hospitable to computer chauffeurs. Once people concede to make the relatively minor adjustments to the infrastructure necessary to impart surety to autonomous cars, we will finally see some progress towards a truly comprehensive solution. This is what we are waiting for. To wait for a Google style autonomous car is folly.
I predict that this vision of fully human-like autonomous cars is too hard and will be for at least 20 years. Perhaps 50. Perhaps we’ll never know because, like getting aircraft to have flapping wings, it’s a dead end strategy even if it seemed obvious to us at one time.
I feel like autonomous cars are where AI was in the late 1960s when extrapolated progress meant it was just a matter of a few more years before the whole of human intelligence was replicated with computers. Really, people believed that. But of course we all learned that while computers were damn good at many things, getting them to be more "human than human" was way harder than your mother made it seem when she built a human brain. I am more and more suspecting that the problem of driving like a human on human-centric roadways is just as vexing.
So that’s my official prognostication. I’m publicly declaring it in the hope that I can be humiliatingly wrong. I hope either some smart people at Stanford, CMU, Google, Bosch, Volvo, Uber, etc. make some kind of unimaginably major leap in the state of the art. Or that people in the field start to overcome their fear of changing the infrastructure to make it easy for cars to drive themselves. What exactly do such changes look like? Here’s someone (else) from San Diego having a go at the problem. But for the most part we have no idea what the future of autonomous cars will look like. Just like the future of computers weren’t HAL 9000, I believe the future of autonomous cars won’t look anything like Google’s misbegotten projects.
Again, I could be wrong. Please, for the love of god, make me a fool.
UPDATE: I just noticed I’m not the only one who is thinking like this.