Oct 17 2023

Update on Self-Driving Cars

The story has become a classic of failed futurism – driverless or self-driving cars were supposed start taking over the roads as early as 2020. But that didn’t happen – it turned that the last 5% of capability was about as difficult to develop as the first 95%. Around 2015 I visited Google and they were excited about the progress they were making with their self-driving car. They told us, clearly proud of their progress, that they used to measure their technology’s performance in terms of interventions per mile – how many times does the human driver have to grab the wheel to keep on the road. But now their metric was miles per intervention. It seemed plausible at the time that we would get to full self-driving capability by 2020. This is common when trying to predict future technology. We tend to overestimate short term progress, mostly because we extrapolate linearly into the future. But problems are often not linear to solve. Initial rapid progress in self-driving technology turned out to be misleading, and it is taking longer to make it over the final hurdle than originally thought (or at least hyped).

Despite not meeting early hype, self-driving technology has continued to progress – so where are we now? The Society for Automotive Engineers (SAE) has developed a system for noting the level of autonomous driving, from L0 to L5. Levels 0-2 require a human driver to be behind the wheel, ready to take control when necessary. These levels are more accurately considered “driver assist” technology, than autonomous or self-driving. Level three can control the steering wheel and brake, can provide lane centering and adaptive cruise control. This level is currently available in Tesla’s and other vehicles.

Level 3 is the first level where the human driver can fully surrender control to the autonomous vehicle, and is not required to pay attention. Level 3 and 4 can fully drive the car but only in limited conditions, whereas level 5 can fully drive the car in all conditions. Right now we are at level 2 and cautiously transitioning to level 3. However, the difference between level 2 and 3 is often a legal rather than a technical one. Even when vehicles might theoretically be capable of level 3, the car manufacturers may not get approval and market them as level 3, because when the vehicle is in full control the manufacturer is legally responsible at that point for whatever happens, not the driver.¬† Many self-driving cars, therefore, will remain paused at level 2, even as the technology improves, until the manufacturer is confident enough to get level 3 approval. Some market themselves as “level 2+” to reflect this.

But we are seeing level 3 cars start to hit the road, specifically in the US there are two vehicles approved to operate actually at level 4 (fully driverless), Waymo and Cruise. Cruise, produced by GM, is marketed as a truly “driverless” car – there is literally no driver behind the wheel. Its use is limited as a taxi service and only within city bounds. Initially it was approved only for night-time driving, but now can operate during the day. Right now Cruise operates only in San Francisco, but is gathering data to prepare for expansion to Seattle and Washington, DC. So basically, it’s happening – fully driverless taxi services, like Johnny Cab from Total Recall, but without the annoying and completely unnecessary robot.

What has been the experience with these services so far? Overall, very good. Cruise keeps its own data, and claims that their vehicles are safer than human drivers in comparable situations. They had to gather their own human data as well, rather than relying on statistics average out over many different driving conditions. They essentially conducted a study in San Francisco, the city in which they would be operating, to determine a baseline human performance. After millions of driving miles, they claim Cruise does better than human drivers. They claim 54% fewer collisions, 92% fewer collisions as the primary cause, and 72% fewer collision with risk of serious injury.

The National Highway Traffic Safety Administration (NHTSA) is now investigating two collisions with pedestrians involving Cruise vehicles. Both involved pedestrians crossing the street after the light had turned green. In one incident the pedestrian was hit at 1.4 miles per hour. In the second a pedestrian was hit by a car driven by a person and thrown in front of the Cruise vehicle which was unable to stop in time. These incidents gain attention, but they hardly seem the fault of the Cruises system.

Still – this is the challenge of self-driving cars. They will have the share the roads with humans, who are inherently unpredictable, and do things like cross the road when they aren’t supposed to. They may also need to drive in poor conditions, like extreme weather (although that is level 5).¬†Even we assume that Cruise’s own data is fair and accurate, we need to put it in perspective. This is data from a single city. I presume their software is trained on San Francisco. Plus, city driving is limited in speed. The cars are not merging onto highways, or navigating back country roads.

Still, it’s happening. We have crossed the threshold into full driverless cars, even if in a limited capacity. From here we can reasonably extrapolate incremental advances – approval in more cities, then in other environments, and in more and more conditions, with better and better safety performance. With city-bound taxi services, companies like Cruise are picking the low-hanging fruit, but then will expand from there.

The flip side of tending to overestimate short term advance is underestimating long term advance. It seems we will get a world with driverless cars, but perhaps about a decade later than initially promised. It’s also not going to be everywhere all at once, but a slow roll out. But nothing but incremental advances are now necessary to have a full self-driving future, or at least as much as we want.




No responses yet