Guess what computers excel at? Driving consistently
  on consistent highways.

Watching the video, would you say the computer was excelling, or that the road was radically unusual? I wouldn't.

The location is the end of some tidal flow lanes rejoining the main freeway lanes with an unusual configuration of lane markings and barriers. I drive past the spot all the time and can totally see a computer messing up here.

In other words, "works great if it works, might actively try to kill you at random, there's no clear way of telling when it might try the latter, it's your fault in any case." Yeah, that's very much the safe driving paradise I'm being promised "in two years" for a decade now, by the technooptimists. It's almost as if the technology is at 50%-ready instead of the 90% that marketing types seem to peddle, who would have thunk?

pretty much

The computer decided that staying in the middle of a wide lane was more important than avoiding a large obstacle where there was sufficient room on either side. That's bog-standard target-fixation fallacy. Is that excelling?

It might not have been radically unusual, but that doesn't make it consistent either.. (looked like lanes where closed to me)

> "Guess what computers excel at? Driving consistently on consistent highways."

Possibly on paper. In reality, as of right now, computers are clearly far from excelling at this specific task.

Assuming perfectly spherical cars on perfectly consistent highways in a vacuum.

OK, but there are clearly cases where the autopilot gets into trouble, and the human needs to take control again quickly. The human driver absolutely cannot doze off.

Now, I realise people in old-fashioned non-autopilot cars can and do doze off, and that's very dangerous. But it's not clear to me how the autopilot improves that situation. Relying on the autopilot actually encourages you to doze off.

We already have simple remedies like "pull over if you feel tired" and "never ever pick up your phone while driving (or you'll lose your licence)"

Expecting "drivers" to be able to instantly switch from "somewhat zoned out" to having the situational awareness to resolve the problem that AP fails at is unrealistic.

The are two very real issues/threats here.

the first being people are being put in harms way by either false sense of trust invoked by the name or the mixed messages from Tesla

the second is that if the first is left unchecked Tesla could single handily set back Autonomous driving for all by souring public and government opinion.

it needs a new name that aligns better with what it can do. it could be a safety system which gently corrects a driver and takes over in an obvious emergency internal or external. as it stands now it is just dangerous

The problem is autopilot actively engages you even less than highway driving. I argue this exacerbates some of the problems that cause humans to checkout.

I don't know what the answer is but it feels like GM's super cruise does a more adequate job of acknowledging the realistic limits of the technology and explicitly white lists roads where the technology is available for use.

I personally think that without some sort of sensors or beacons in the road, autonomous driving via camera and LIDAR sesnors is ever going to be good enough to achieve level 5 autonomous operation.

I think the world modeling is the part that autonomous cars can obviously be better at. They have all sorts of advantages over humans; access to much more detailed maps than the average person can recall, multiple types of sensors pointed in multiple directions, a system designed to integrate all that information in a way that is advantageous for driving (we have a system evolved for whatever it evolved for) and so on.

It's the sophisticated behaviors necessary to safely drive through that world model that are the issue.

The success of emergency breaking systems (that aren't advertised as "driving" assist) are pretty good evidence that the sensors can serve well as input to safe behaviors.

If a human can navigate current roads, why shouldn't a computer be able to? We may need to develop a new types of sensors for vehicles, but that seems like a better/easier plan than installing beacons on every road, everywhere.

And what about beacon maintenance? Seems like most cities have a hard enough time keeping up with pot holes, lines, etc. as is.

If a human can write a novel, why shouldn’t a computer be able to?

I have no reason to believe a computer cannot write a novel.

They probably need to be able to reproduce before they can write a compelling novel.

Beacons is a great idea -- have some standardized beacon system, and a public map of autopilot-approved roads.

Following the beacons safely would be a vastly easier problem than trying to completely replace a human driver in all situations, but it would still give you about 90% of the benefits.

Lines fade away because there is little money for maintenance but beacons that cost multiples are the solution?

The first thing I thought when I read beacons: Hackers are going to have a field day with them. Add malicious beacons to streets and cars will drive off road at high speeds.

Well, look at cat’s eyes (sometimes called Botts’ dots in the US, I think?)

I assume they’re more expensive to install than just painting a few lines, but they’re very robust and long-lasting, and they’re fantastic for human drivers. It’s not a stretch to imagine something similarly useful for computerized cars, that links to a standard road database.

Hacking is a risk, sure. I envisage you’d lock it down by having a cryptographically signed master map; if the observed beacons diverge from the map, the autopilot system would refuse to proceed. (OK, I guess that allows a DoS attack at least.)