[jacquesm]

Perception is where it begins. Then you can start to reason. Case in point: just the other day I drove past a line of cars. Right hand side passenger door opens, so I slow down by lifting a bit and sure enough, less than a second later the driver side door gets flung open. If not for reasoning then I would not have lifted and likely would have hit either the door or the former driver of the car when they got out. Those little hints that we humans are good at processing (passenger disembarks: driver likely to follow) would be very hard to teach a computer in such a way that it would not lead to false positives all the time.

[inventtheday]

couldn't that specific correlation be learned with enough data?

[bumby]

I think maybe you could, but I think it would take a lot of data to correlate the edge cases.

Think about seeing a playground ball rolling into the street. Most human drivers would anticipate a child jumping into the street after the ball even though they’ve never experienced that situation before because they can correlate the perception with non-driving scenarios (I.e. playground + ball is associated with children). A self-driving car may not have the non-driving context and I’m not sure society is willing to accept the outcome while the cars gather enough data to learn.

And that’s to say nothing of all the off-nominal maintenance related conditions to account for.

Maybe there can be enough data sharing to update models effectively. It will be interesting to see this industry evolve

[DayDollar]

Comp vis often reclassifys objects, loosing history.. You see them get in- you know they are closing the doors. Comp vis sees a butterfly, and sees a car with doors half open.

If you do not allow for frequent doubt, aka re-classification, you get hallucinating systems, imagining a bike-rishka as a slow car with people hanging out

[bumby]

There’s been some cases that highlight the danger of this as well as bad examples of how mitigations were poorly implemented.

I believe a case of a pedestrian being killed by a self driving car was related to the system reclassifying enough to initiate a deliberate delay timer. In a system traveling 60mph this delay in decision was enough to cause a mishap

[Animats]

Cite? The Uber killing was simply that the self-driving system was not allowed to do an emergency stop. All it could do was alert the driver. Presumably because they had a high false alarm rate.

The serious accidents involving self-driving cars have all been huge failures at the basic "don't run into things" task. Google/Waymo, which seems to be past that, has subtle problems, like "projected that bus couldn't fit through space to left of car in wide lane and so turned slightly into path of bus and was scraped." The standard Waymo accident is "advanced into intersection, detected cross traffic, stopped, was rear-ended by human driver at low speed".[1]

[1] https://www.dmv.ca.gov/portal/vehicle-industry-services/auto...

[bumby]

“The self-driving car was fully autonomous at the time at the accident” [1]

It appears the mitigation for a high false positive (I.e., nuisance alarm) was to program in a delay (I.e., “action suppression”) which inadvertently introduced a worse hazard, the timetable from it is in the link below

[1]https://www.theregister.co.uk/2019/11/06/uber_self_driving_c...

https://dms.ntsb.gov/public/62500-62999/62978/629713.pdf