Comment by BugsJustFindMe | Hacker News Reader

Comment by BugsJustFindMe | Hacker News Reader

Parent Full thread

> As I've been saying for years now, the first job is to determine that the road ahead is flat enough to drive on. Then decide where you want to drive. I did the DARPA Grand Challenge 16 years ago, which was off-road, so that was the first problem to solve. Tesla has lane-following and smart cruise control, like other automakers, to which they've added some hacks to create the illusion of self-driving. But they just don't have the "verify road ahead is flat" technology.

This matches my perception as well and it continues to blow my mind. Like it just seems like it would require full-on incompetence to use only specific pedestrian/car/sign/lane classifiers rather than splitting the world first into ground/obstacle by reconstructing geometry and then assigning relative velocity, acceleration, and curvature to coherent obstacle segments irrespective of their identification. But the videos always make it look like this is exactly what's happening. And worse it always appears to be happening on an instantaneous frame-by-frame basis with things flickering in and out of existence as if "thing there" and "nothing there" are somehow equally safe guesses.

The flickering is concerning. Kind of suggests their NN hasn’t learned basic object persistence yet.

That's the thing. The software shouldn't have to "learn" fundamental basic real world shit like that. We KNOW that things don't just flicker in and out of existence. The system should be built on that premise from the start.

Oh but that's what basically every gps software is doing - one moment you're slowly trailing on the highway, the other you're placed on the bridge above.

Most GPS software is smart enough to not flicker between the highway and the bridge repeatedly though. It's not like this category of problems is a wholly new thing. The car needing to switch lanes twice a second like in one of those videos is fairly unlikely, so if the ML is spitting out that it should, there probably ought to be a layer on top of that that's interpreting the ML output that can take the decision to pick one. Picking the wrong lane is less wrong than being indecisive about it.

I don't know about "most gps software" but I have a Garmin doing that and Google Maps does the same. So in my experience it's 100%. Now if such well established software products didn't figure out such basic stuff, either we're calling it wrongly "basic" (and it's a difficult problem) or the responsible factors just don't bother. The major difference is though, while gps navigation can do with frequent recalculations, realtime obstacle avoidance less so.

> there probably ought to be a layer on top of that that's interpreting the ML output that can take the decision to pick one

It needs to be an input, not just a filter on output. A heuristic that doesn't account for the previously observed state is insufficient because prior existence information changes what gets detected. And it needs to lean heavily toward "if we saw something there before, the safest course of action is to assume that it's still there until strongly indicated otherwise".

It can be a filter on the output and account for previously observed state. E.g. one trivial (and probably pretty bad) solution in that direction would be to boost the probability of small state changes as opposed to large changes.

> if we saw something there before, the safest course of action is to assume that it's still there until strongly indicated otherwise

Normally I'd agree with you, but driving a car is a bit of a special case since in may cases stomping on the brakes is way worse than doing nothing. Take the video of the tesla seeing spurious traffic lights[0] for instance, should it really do an emergency stop in the middle of traffic in that case?

All the machine learning can do is give you probabilities on what the state of the world around the car might be. The software needs to pick from those probabilities some state to act on in such a way to be the least likely to endanger anyone. There isn't a simple correct answer. all the variables need to be weighed. currently teslas seem to ignore some pretty important variables while driving.

[0] https://twitter.com/sascha_p/status/1400173874285744129

First, jumping and flickering aren't the same thing. In my experience, google maps is somewhat unwilling to rapidly reposition you if two paths are equally probable.

Second, the cost of deciding that you're in the wrong lane is very different from the cost of deciding that there isn't an obstacle in front of you. The worst thing that happens when a navigator gets your position wrong is you miss a turn and spend a few extra minutes in the car. The worst thing that happens in the other case is someone dies instantly in a cloud of blood.

IIRC this is the stated reason they're dropping radar, the camera network maintains object persistence but the radar disagrees very frequently with the camera network causing the flickering because the system doesn't know which input to trust. Apparently by throwing out radar signals they get smoother object recognition.

> Apparently by throwing out radar signals they get smoother object recognition...

...and driving straight into barriers.

> the system doesn't know which input to trust

Well, clearly their camera network is untrustworthy, so...

I don't think radar can sufficiently differentiate between stationary obstacles and background - hence the long habit of smashing into stationary objects.

the snipped linked are on 9 beta, which is only vision. object flickering and other issue remains unabated.

Reply on news.ycombinator.com