However, it is an open question whether or not it will require additional / different sensors (like lidar).
You are driving your car and see that the pickup truck in front of you has an empty cardboard box in it (you can tell its empty by the way its floating around in bed of truck). Suddenly a gust of win picks up the box and it lands directly in front of your car. Its not very big, about 12 inches tall. Do you: (a) Slam on breaks? (b) Roll over it?
Now what will the "AI" in your car do?
Self-driving cars are not coming anytime soon. They are a whole slew of problems that you can intuit solutions to fairly easily, that are still beyond the ability of current tech.
Manned vehicles are not coming anytime soon. They are a whole slew of problems etc. -- you got my point.
I think focusing too much on pesky details is very much a fallacy in this case - you do not want an "AI" to react like a human in all situations, you only want it to drive in a way that is conservative enough not to endanger people too much. And we clearly aren't that far from this goal right now.
Humans who drive cautiously may go for a million miles without an accident. The best self-driving cars (i.e. Waymo) disengage on average every 11,000 miles.[1] It seems to me that a disengagement is equivalent to becoming unconscious without warning, and presumably we both agree that a given human does not have a heart attack while driving every year.
Humans, even with all the people who drive drunk, or texting, or falling asleep, average about 80 million miles between fatalities. Going 11,000 miles between events of total loss of control is nearly four orders of magnitude worse.
[1]https://www.forbes.com/sites/alanohnsman/2019/02/13/waymo-to...
However, I think this metric could be irrelevant in the case of a home/work commute. The 11,000 miles average appears to have been obtained basically by randomly driving Waymo cars on Californian roads. But a usual commute is much less than 11,000 miles, and if your self-driving can do it by itself once then probably it can do it twice. As the article puts it:
"The value of the data is limited, however, as the figures don’t factor in the complexity of environments in which vehicles are tested–dense urban settings, versus low-speed suburbs or less complex highway driving–nor do they show conditions including weather, light or speed."
Nevertheless, you seem to have missed my point : I was arguing that coming up with a specific use-case example that may (or may not, actually) go wrong is an argument that goes both ways.
If we pick a hypothetical where a human fails and an AI doesn't, does that mean self-driving cars are ready?
Do you think human drivers are faultless drivers?
Is their a fair cost-benefit analysis for not wasting enormous amounts of collective time driving? (Driving is a very unproductive activity).
They still have massive, on-going investments in anti-aging and self-driving vehicles. I don't know of any moonshots they've killed yet.
Google fiber is mostly dead.
Project Ara?
I'm envisioning the Mr. Fusion in Back To The Future as a Google product. Surely it would fail the second it left that narrow 5-10 year window of support. Error: 403, Forbidden; the user is not authorized for trans-time operation.
So, if Mr. Fusion was a Google product, as soon as McFly transported the DeLorean to the future (outside Google's window of support) and the onboard computer synced the time over-the-air and the Mr. Fusion updated its clock, McFly would be stuck in the future, though presumably in that time he should be able to find some competitor to Mr. Fusion which he could replace it with (with Doc's help; McFly never seemed to be very good with technical things).