[lolc]

All in all I'm quite impressed with the demonstration. It was way more thorough than previous videos I've seen. The main things the car is failing at from what I see are the hard things: Object permanence and ad-hoc reasoning. So no surprises.

Regarding object permanence: I was impressed overall with their detection. Still, you could see kids walking close to parents blink in and out of awareness of the car. Now I'm not saying humans are very good at tracking a multitude of actors. So at some point the machines will be "good enough". But that point seems way off when significant objects like kids can just disappear from awareness when they pass behind a stroller.

And about the ad-hoc reasoning: They have the whole city mapped out! Including traffic lights and turn restrictions. I'm not even clear whether they try to detect the signs at all. I'd assume that they have an operations center that hot-patches the map with everything cropping up during the day. So the cars would send in unexpected changes to the road and they would classify those changes and patch the map. Meaning the car is tethered to that feed and not autonomous in the strictest sense. Sure, such a center would be marginal cost given a large enough fleet. Still it's a subscription you'd need for your own robocar.

They mention a lot of things they are prepared for. And I can't help but think "oh they're really good" when they say "detect backed up lanes" or "creep into intersections". But that always leaves the question what happens when they're not prepared for something. When the rules don't fit. Can the car go over a curb if the situation warrants it? Does it back out of a blocked off section? Is it even able to weigh whether backing out is an option at this point?

so I'd like to see a "what we're currently stuck at" video. But I understand one can't very well attract investors with such a video.

[idavidrein]

I agree with a significant amount of your point, but with regard to object permanence, I would guess that they have prediction algorithms that don't only rely on the current-time perception, so if something blips out of sight for a second the system will still infer/predict it's existence (for a time - obviously if something is hidden for a long time it won't continue to not trust perception).

[lolc]

I'd be very interested to know how that works. But I don't think they have it.

The boxes they draw are very wobbly and dimensions expand and contract directly with sensor input. Maybe they only show fused output (in itself an achievement) and there is a later step they don't show. That would be weird though because if they want to brag about their model they would definitely show it if it was any good.

[idavidrein]

that's a fair point, but it seems reasonable to me that they would separate the sensory input and the predictive/higher-level aspects of their modeling. For example, we know for a fact that they must be doing tons of prediction for both cars and people, so I think it's likely that different models (not sensory) have the info that a person is probably still there.

[lolc]

Yes it's true they must have some form of persistence when they do predictions. But expected trajectory of other vehicles and pedestrians was missing from their video. A lot of other interesting feeds were missing too, so I don't know what to read into it. I tend to think that that stuff would look much worse. But maybe they just didn't want to clutter the video or show how advanced they are already.

[idavidrein]

yeah, it's possible that the stuff doesn't look very good, but my guess (maybe my hope?) is that it's too cluttering or through careful analysis could reveal IP about their predictive algos