Hacker News Reader

> We sort of form our own "lanes" based on where we think the lines SHOULD be.

And that is exactly what FSD does. You can search YouTube, there's plenty of examples if you want to see how it performs in different weather conditions. Snow is challenging but not impossible.

View on HN

That is not my experience at all. It either says “Full self driving unavailable. Poor weather detected,” or it freaks the fuck out and tries to kill me.

FSD can’t adequately handle rain, let alone snow where the road lines are not visible.

Based on many hours of footage, rain doesn't seem to be a big problem unless it's very heavy, but the experience can vary I suppose. Snow for sure is still a problem, but there is a path to improving it. The current state of perception shows that it can be solved by deep learning (more training data for these conditions), in the sense that in many situations it can guess where the lines and lanes are based on other context, like tire tracks, positioning of other cars, width of the road and lane meta data, snow protrusions etc. There's also the problem of when to treat snow as an obstruction to go around and avoid vs when to just go over it. There will always be more uncertainty in these conditions though, just like there is for us humans, we are often unsure where the lanes are exactly, and one of the challenges if having the system deal with this uncertainty better. And then control has clearly not been worked on for snow and ice yet.

Reply on news.ycombinator.com