I just got back from visiting JPL today. That long car trip got me thinking about the best general strategy for pursuing the goal of autonomous vehicles.
There are two competing approaches. One approach is the refinement of overly ambitious AI until the car can drive on all roads and situations just like a human. I’ve already mentioned that I strongly believe that this way is not happening any time soon.
The other approach, my approach, is to figure out what we could have been doing ten years ago, and just get on with it. Today! To better explain the general concept I’m trying to outline, I’ve thought of a specific example of my strategy. What follows is a strategic outline that could, from a technical and practical standpoint, begin to be executed immediately.
Let’s start with the company that has done the most to advance the state of the art of autonomous cars. No, not them; I’m talking about Tesla. Tesla is in a unique position to really make a big difference. They’re not a perfect fit for my strategy, but for illustration purposes they’ll do.
Tesla already makes cars. They don’t exactly make the right kind of cars, but in some important ways they do. The fact that their cars are computer controlled drive-by-wire is important. The fact that they have cameras and enough sense to be able to do their lane keeping mode is helpful. But, as we know, the best hands free driving in a Tesla still demands that the driver pay scrupulous attention. This is the thing that I think is critical to eliminate as quickly as possible where possible.
So we have a Tesla that darn near drives itself, but not really. What progress can be made? Here’s the strategy that Tesla should pursue.
Find popular long travel routes. A classic is Los Angeles to Las Vegas. Or Los Angeles to San Francisco.
Prioritize these routes by how monotonous they are to drive. Also prioritize routes that see no snow and little rain. The stretch of I-8 between Yuma and Gila Bend is a perfect example.
Start collecting Tesla Autopilot data on these routes. Note any quirks or problems that ever come up. Add site-specific software if need be. The goal at this stage should be to show that a significant number of cars made this trip with zero human input. Maybe more than would be expected to do so without a human mishap.
Build some special lots (think truck stops) right off the exits at each end of the long stretch. For example, one just east of Yuma and one just west of Gila Bend. For all intents and purposes, there is nothing in between but hours of very plain driving. These lots would be run by Tesla and they would have several functions. The primary function would be that a Tesla in this lot should know how to get on the freeway and then drive to the next lot’s exit, get off the freeway there, and park in that lot. That’s it!
Rent out Teslas. In this scenario, I should be able to rent a Tesla in San Diego or LA and drive to Yuma the normal unsafe irritating way. Maybe I could use attended autopilot where feasible, but that’s not too important. When I got to Yuma, I would pull off to the special lot, verify my plan and see if the road conditions are ok for such a thing, and turn over complete unsupervised control to the car for the next couple of hours. When I get to the Gila Bend lot, I wake up and resume driving.
When I get to my destination, say Tucson, I’ll have a car to do any flexible driving I want, but for this round trip, I’ll knock hours off of it where my attention and energy are required.
One tricky paradox with this is that Teslas, being electric, are not ideal cars for crossing the desert, an environment which can barely supply fuel to gasoline powered cars. The answer is that these special lots should also do battery swaps. This is especially true for the rented cars.
For some reason this vision of how autonomous cars could have been working for our safety and sanity a decade ago is repugnant to fancy AI specialists. They seem to feel that 1% of ideal is a glass 99% empty and would prefer just keeping that glass 100% empty until they can make it 100% full. Which should happen about the time that computers will make transportation obsolete for other, hard to imagine reasons.
I believe that by establishing a foothold of that 1%, it will inspire 2%. And then 4%. Soon almost any monotonous long drive should have an option to swap out your battery and have the car take you down the boring road. A lot more people would be buying Teslas. If Tesla was smart and used open standards for the batteries and certifying the road safe for simple autonomous use, then other companies could make compatible cars and people would demand the system be expanded with an exponential network effect. I believe that very soon we would be at the 95% we are at today with research cars trying to handle any road like a human. The difference will be that there will be a lot more real use of real autonomous vehicles. That final 5% will be easier to cure not just from that experience, but because it’s the final 5%; the will to just get rid of it and replace it with an easy environment will be much greater.
What I’ve described is just an example of working with the AI that is perfectly good today and not perennially wishing and hoping for some science fiction AI to get here in 5 to 10 years. I think the Otto truck transportation concept is closer to mine than many others. Perhaps with my kind of system, problems like rain and snow can be avoided by simply not allowing cars to be fully autonomous in such conditions. Sure, that’s not the best situation imaginable but it might be the best situation that is actually useful today. It would certainly highlight where more work truly should be done. Today the work seems focused on the impractical quirky situations that vex even human drivers, situations computers are probably never going to properly deal with without a lot of special help.
Not only am I proposing that this kind of plan can happen. I’ll go farther and make another prediction: that it will happen. In rough terms the way I’m describing the advent of autonomous cars will happen and current well-known efforts like Google’s will fail. Autonomous cars will first begin to be useful by the public only in some limited and incremental form and only on certified roads.
The question I have is really, when? Should we start today? Or should we wait for another 50 years of AI research to try to bring us closer to the human-level intelligence needed to drive like a human? A goal that may be unattainable, especially if the last 50 years of AI research is anything to go by.