[manyxcxi]

I'm no Luddite but I basically yelled at my business partner the other day when he took his hands off the wheel, eyes off the road, and was punching out an email as if we were on the train.

I've done plenty of computer vision projects and pitched in on some autonomous driving/flying hobby projects- so I'm definitely willing to accept our robot driving overlords, but maybe I know just enough about it to know we're not all the way there yet.

I feel like Tesla is in a bit of a Catch-22 in that to get the system better you want as many people driving with it and feeding you data. At the same time people are people and they will get lulled into a complacent state and begin not paying attention pretty quickly.

That being said, if there was some data that said right now, today, Tesla autopilot is better than the average (or maybe up to the 75th percentile) driver, I'd be okay with people just letting it take the wheel and not screwing things up.

EDIT: Just to be clear, we were in a Tesla w/ autopilot. We weren't just coasting down the highway at rush hour in a regular car.

[foofoo55]

Tesla could gets lots of quality data with the autopilot disengaged and running in a passive state such that the autopilot's calculations are compared to the human driver's actions. Any large variances (such as the autopilot wanting to suddenly veer left when the driver didn't) would be very valuable. I've done this with experimental machine control algorithms a number of times.

[majewsky]

What makes you think they're not doing it?

[6nf]

There's not enough data yet to say for sure, but currently Tesla's autopilot record is one fatality per 70 million miles (assuming the recent Chinese death involved autopilot)

You can compare this to your local road stats. I don't think autopilot is there yet in terms of safety.

[arekkas]

computing an avarage over one or two data points is questionable statistics