[Klaus23]

Sounds like they were relying solely on their neural network path prediction, which failed when the truck was dragged at an odd angle.

A simple lidar moving object segmentation, which doesn't even know what it's looking at but can always spit out reasonable path predictions, would probably have saved them.

I think Mobileye is doing something like this, but they release so little data, which is always full of marketing bullshit, that it is hard to know what exactly they are working on.

[digitallis42]

It's unlikely to be neutral network based. This sounds like a model prediction failure. You take a mathematical model of car motion: the rear wheels generally don't steer. The front steered wheels can cause the car to drive along an arc. If you want to predict the arc that will be driven, you take the initial starting heading of the vehicle and project forward in time based on your understanding of the vehicles steering angle. For most "driving in lane at velocity" cases, you generally would assume that the vehicle has very little steering angle input.

We're now getting to see where autonomy needs to develop "spider sense": the scene in front of me feels wrong because some element isn't following the expected behavior maybe in ways that can't really be rationalized about, so we'll become much more conservative/defensive when dealing with it.

[im3w1l]

I'm thinking it might make sense to have a sort of hierarchy of models. The stupidest model predicts that everything will be stationary. The second model predicts that everything will travel in a straight line. The third model tries to predict a circular arc based on same fusion of path history and observed steering input. Fourth model uses a notion of action to predict whats going to happen, like "the car is changing lanes". Fifth model uses body language and common sense to predict intention "the pedestrian wants to cross".

Each model can potentially predict longer into the future but also has more complexity and things that can go wrong. So you keep track of how well each model is doing (on an object basis) and if one level is failing then you fall back on a stupider one. You might also want to increase caution if your models are not doing well (lower speed and increased safety distance).

[roxgib]

These cars can and do slow down or even stop and wait for human assistance in response to unexpected situations. I'm actually quite surprised this didn't trigger here, although we don't really know much about the specifics of the situation.

[Klaus23]

It is definitely based on neural networks: https://www.youtube.com/watch?v=RpiN3LyMLB8

[rasz]

Its NN black box alright, for a hand crafted model you would have something like:

    if prediction != reality
     {mismatch+= whatever the time tick is}

    if mismatch > second (or two)
     {abort driver assist}