I see, thanks. Yes, although Tesla doesn’t disclose what their algorithm is, there’s no reason to believe it’s any more sophisticated than “there was some input”, although I believe it has to be purposeful input, to thwart the old “hang a weight on one side of the steering wheel” ploy. Agreed that it would be desirable for the algorithm to be fancier than that, as long as they can do it without throwing too many false positives, which can have negative consequences that are not immediately obvious.
Since I brought up the consequences of false positives, consider: what do you do about it when you decide the driver is too drowsy?
You probably go through a series of alerts with the penultimate step being sirens and flashy lights to wake them up. (The final step is, if they don’t respond to sirens and flashy lights, you declare an emergency and do the best you can to bring the vehicle safely to a stop.)
Once they’ve responded to your alert, what do you do then? At that point maybe you can turn off the automation. Now you have a driver you believe is drowsy, who you’re forcing to manually guide the vehicle, instead of having your not-drowsy automation do it. Maybe the driver has the opportunity and good sense to pull off to the side of the road and catch a nap. Maybe they can’t or don’t.
If your drowsiness detector is reliable, that’s about the best you can do, so you live with it. If it’s not reliable, and you’re making non-impaired drivers put up with sirens, flashy lights, and being put into “autopilot jail” because of a deficiency in your system, then you have a bunch of annoyed customers who are driving their cars manually instead of with benefit of assistance from your automation. Which, let us recall, is presumed to make driving safer when used correctly, so if you’re disabling it unnecessarily you’re making them less, not more, safe.
Almost all the foregoing, as it turns out, describes the existing Tesla system, with the only caveat being the algorithm they use to determine attentiveness.
Edited to add: come to think of it, you can even argue that Tesla does provide some measure of drowsiness detection, in that if it has to escalate to the siren-and-lights level several times in a row, it does put the driver in “autopilot jail”. If you accept that “driver didn’t notice the lower-intensity alerts” is a proxy for “driver isn’t alert” then it follows that this qualifies as the kind of system we’re talking about. Of course, nothing is so good that it can’t be made better.
You would imagine that, yes. As I’ve mentioned upthread, my experience in four+ years of using such a system is exactly contrary to what intuition suggests, however. I can only speculate as to the reasons, but my guess is that because the car is handling most of the lowest-level fine-motor work – the most monotonous part of driving, in fact – I get less fatigued and am freed to spend more of my attention on overall supervision. It certainly seems to pan out that way. This is why I’ve asked several people “have you used it?” It’s easy to form an opinion about a system one hasn’t used, based on simply reasoning about it. Often, reality matches our reasoning. Sometimes it doesn’t, though. That’s why we need data before drawing conclusions.
