I try to keep up on autonomous vehicle (“AV”) development as it interests me and because son Ed is quite involved (Communications Director of PAVE (Partnership for Autonomous Vehicle Education)). This video came up on his Twitter feed this morning, and I decided to check it out.
JJRicks, who has documented a number of Waymo rides on YouTube, had his driverless taxi go rogue, and it’s all documented in this video, the relevant part starting at 12:00. It’s pretty embarrassing, both for the technology as well as the poor rider support, which was not only unable to get the taxi back on course, but actually put it into a much more exposed and potentially dangerous situation, blocking both lanes of a parkway because of some construction zone cones. Why?
There are two very distinct approaches to driver assist and autonomous vehicle systems: Everyone other than Tesla, and Tesla’s. Waymo’s (and others) approach is based on a “stack” of rather rigidly applied functions that are programmed, and are heavily dependent on high definition maps and geo-fencing (limiting to certain geographic areas that are well-mapped and documented). Tesla’s very ambitious approach is based on the assumption that AI (artificial intelligence) will steadily improve, and its system is designed to work in as many situations as possible, which is rapidly expanding on a regular basis.
What happened here with this Waymo taxi is that it encountered a blocked lane (cones) to the right, and it refused to make a right turn using the available left lane. The reason this happened is because it’s still utterly unable to cope with an unexpected event like this. Waymo, which operates in Chandler, Arizona, has to constantly update its system for construction zones like this, but apparently this one was not input for one reason or another (there is a specific reference to that situation by a Waymo person in the video).
Eventually the Waymo is coaxed to make the turn, but then stops at the first cone it encounters. Frankly, this is pretty embarrassing, as dealing with cones (moving into the other lane of open traffic) is considered pretty elementary. I did not realize just how dumb Waymo Driver still is, and that every construction zone has to be entered into its software so that it knows what to do when it encounters it.
And all this is happening in Chandler, Arizona, a suburb of Phoenix noted for its very easy grid of streets and highways. How far would a Waymo taxi get in New York or New Jersey? And Waymo Driver has consistently been held up as the most advanced AV systems.
It all points out that AV development has quite a ways to go, and that the “trough of disappointment” is a lot deeper and wider than may have been imagined.
I accept that Tesla’s approach has involved over-promising, as it’s essentially a moon-shot. And the whole AV industry (and some regulators) strongly disagrees with it, and denounces at every opportunity, as in every crash with a Tesla, whether Auto Pilot or FSD (“Full Self Driving”) was actually a factor or not even on. It’s consistent with Silicon Valley’s “break things and fix them later” approach, which does inevitably involve collateral damage of one sort or another. But in the long, long run, it does make me think that Musk’s vision that AV systems can and need to learn to cope with essentially any situation rather than depend on being told in advance that there’s going to be a closed lane because of construction.
So if you’re worried that they’re going to come and take the keys away from your car and force you to ride in an autonomous pod, relax. It might be quite a while.