Autonomous Hype Needs to Cool Off


I'll start by stating that I'm no autonomous or AI expert by any measure. But what I'm going to say should be painfully obvious to all and that's why I'm confused with all the hype. Specifically the hype from companies and their followers that truly believe we will have autonomous vehicles on the road in 5 years. I also want to be clear that I am excited and ready for this tech to mature just like most readers on this site. But we have a long wait for so many reasons.

I'm not going to be all negative, we are actually nearly there, but the 80/20 rule applies here perfectly. We have the sensors and data to pull off autonomous driving on most major roads and highways in a lot of conditions. This is the easy(ish) part. The last 20% is will be extremely difficult. It is easy to look at the current rate of progress and assume Here are the hurdles holding us back that keep me deeply skeptical of a near term roll out.

  • Inconsistent road design, sign design, markers, etc: There are major differences in the US that most humans can deduce but machines must be able to adapt and learn. Think of all the times you've been confused while navigating a new area. Outside the US this problem becomes an even bigger hurdle.
  • Constantly changing road conditions: construction is a big one, near me my route changes daily and you often can't see road markings at all since they've used the shoulder. Mapping and reviewing even highly detailed 3D models won't work outside of major metro areas. Cars must be able to adapt on their own, instantly.
  • The real rules of the road are not binary: Even extremely well thought out if/then statements can't handle all situations. That double white line? Don't cross it (except when necessary) but only if conditions are safe (based on what?). For example, I recently hit a huge hawk. It took off from the median right in front of me and I had cars on all sides, and close behind. I made a choice to keep going rather then slam the breaks, I killed the hawk and damaged my car, but I didn't cause a pile up. I made a value judgement, the risk of fender benders was too high for the hawks life. If it was a child I would have done differently. A computer can go through these options faster then me, perhaps even anticipate the hawks actions faster, but who gets to make this value judgment; a programmer, regulator, the car? There is no law that determines what level of risk each animal deserves and there are uncountable grey areas on our laws/rules like this. With autonomous vehicles there can be 0 grey area.
  • Weather is a big one. Rain, snow, fog, ice, hell even hi heat could and do effect sensors. When conditions go to hell so do your normal ways of driving. You get creative, something computers aren't good at yet. Leading to my main technical point.

No one is out there trumpeting the arrival of real AI in the short term. We agree it is a huge technical and even ethical problem. But true driverless cars (ones without steering wheels that can take us anywhere our current cars will take us) require it. In fact they require AI that is smarter than us, faster than us and less fallible than us. Because the arrival of Autonomous vehicles is supposed to reduce risk and give us more freedom, right? Anything short of that is reliant on short sighted algorithms that can't consider every situation on the road. Yes, I believe we will have highway and city autonomy in <10 years by some luxury manufacturers but for the rest of us and those who travel the road less travelled it will be decades at least. I agree with John Sircusa who said on the Accidental Tech Podcast recently that he hoped he would see it in his lifetime, but is doubtful.

We haven't even covered the biggest hurdle of all. Our infrastructure and regulatory issues have a long way to go and that won't happen easily. We need standards to be developed, tested and enforced for car-to-car communication, road markings, autonomous road testing, liability concerns, and on and on. This won't work when Tesla can afford and attract the best software engineering talent but Chevy gets the leftovers and puts unsafe cars on the road. In order to get cars that are autonomous and affordable this must be done all over the world at the same time, by everyone. It requires more effort and collaboration then the Interstate Highway System except on a global scale. This would be a hard task even if the current political climate wasn't a shit show.

Basically it comes down to this. AI is a long way off and infrastructure will take a long time to happen. Even if the most rosy outlooks were true, car development itself is slow. New models come out every 6 years typically and autonomous cars would require even more R&D. So why do I keep hearing estimates of 5-10 years from so many outlets? If I'm missing something, please show me the demonstrations, research articles and data. I want it just as bad as everyone else.