turn based speed control too slow with CSLC

[stiff osprey]

Decides that a turn needs slower speed too late, or decides to send new speed values too late, as a result cruise can't slow the car fast enough for the new desired speed to be useful. Instead it usually slows down after the apex of the turn which is worse than not slowing at all in some situations especially with following traffic (police, you look drunk)

[sleek iris]

@stiff osprey limitation of the model. cslc works by looking at the model curvature predictions, if the model doesn't see the curve soon enough cslc won't either. Usually only good for curves that are pretty visible and only need a 10 maybe 15 mph slow down tops

map based turn speed control is meant to help in situations that turn based doesn't see soon enough, but it also has its own issues and limitations. Ultimately the approach that comma is trying to do (NOO) using map data in the driving model will be the best solution, but it can't really handle a lot of curves yet

[ocean surge]

Thanks @sleek iris , that explanation was way better than I could come up with!. But yep, if the curve is not seen there's no adjustments than can be done, and on those cases MTSC would be probably the best solution.

Another thing is if you use SLC, and you have map priority or map only set, and the OSM data downloaded, it will suggest you to set your speed to the upcoming speed limit. At least here in EU, most of the exits or curves that would be hard to take fast have a different speed limit marked, so SLC tends to actually help predict the curve a little, at least in many of the cases prevents a stronger slowdown

[stiff osprey]

I mean I can see the curve on the screen with a path showing the curve long before it starts editing the cc

My thoughts were that it was just a timer or wait that was waiting a bit too long, or waiting for a confidence percentage or something

[ocean surge]

I can't really say for certain, we are using @sleek iris's implementation, which @wind spruce ported over to SP. I tried but like I said, with tuning and lat I am not the best 😄 so I leave it to the pros hahaha. But from my empirical experiments, I can say it works 90% of the time that I need it, specially with very clear curves ahead

[sleek iris]

The start of the curve doesn’t really count, the start of the curve isn’t aggressive enough to need to slow down for. It needs to see the apex of the curve before it knows how much it actually needs to slow down. It literally just calculates a speed for the sharpest curvature the model is currently reporting and uses that as the target

No timer, no state whatsoever

Doesn’t even know about confidence. It’s actually a remarkably simple solution with next to no logic other than finding the sharpest curvature that is being output and calculating a target speed

[stiff osprey]

In short; fuck.

[sleek iris]

In fact part of what makes it better than previous solutions I’ve tried is that it doesn’t even try to figure out the distance to the curve. The model doesn’t see far enough for trying to calculate the distance to actually make sense. Basically if it sees curvature it might as well just immediately start slowing down

[stiff osprey]

Yeah your explanation made it clear it's slowing as soon as it has a solution to do so

Which is not enough for these back roads around here unfortunately, but it's still better than without

[sleek iris]

Yeah, I live in a rural area, hence my work on mtsc lol

[stiff osprey]

I will agree if it's a very large turn and it sees the entire curve it does seem to work pretty well, I just didn't understand how it was making the decision really