Hacker News Reader

Comment by paxys | Hacker News Reader

I think people vastly overestimate the challenges of weather conditions for self driving. With modern car tech (traction monitoring, ability to redirect torque to a specific tire, ABS, radar) an automated car is going to have an easier time navigating snow/ice/rain than a human driver.

The real challenges when navigating city streets are the human ones – delivery vehicles blocking lanes, municipal worker fixing a manhole with a single cone to redirect traffic, pedestrians/bicyclists appearing out of nowhere, no one following traffic signs. This is the kind of stuff that tests "intelligence".

View on HN

> I think people vastly overestimate the challenges of weather conditions for self driving.

This remark makes me wonder if you've ever lived in an area that actually experiences winter.

Around here, dead of winter, there are no lines visible on the streets. Heck, after a good snow storm the lanes are basically a function of group consensus.

This is a situation where automation has the advantage. With detailed position information and detailed maps, the fact that the lines on the street can't be seen is irrelevant. The car didn't need them anyway.

(noting that Waymo requires full detail maps to be able to drive an area, including all signage).

Unless you're tesla and you're using computer vision to determine where you are on the road.

Tesla is using maps too. Source: the recent Tesla AI day https://m.youtube.com/watch?v=j0z4FweCy4M

So, couple things.

First, it's optimistic to assume that data is accurate and up-to-date.

Second, I can't emphasize my "group consensus" point enough.

Anyone who's driven in real world winter conditions has seen a day where three lane roads turn into two. Or lanes form in the shoulder. It'd be actively dangerous to insist on driving according to the underlying lane markings during those types of road conditions.

Maybe in a world where all cars are autonomous and using map data that could work. In reality it really doesn't.

> With modern car tech (traction monitoring, ability to redirect torque to a specific tire, ABS, radar) an automated car is going to have an easier time navigating snow/ice/rain than a human driver.

Huh, what? Human drivers can already take advantage of all of those, and they still find snowstorms and torrential rain challenging.

The challenge is understanding what you see (and hear), and dealing with very noisy and limited--sometimes actively misleading--inputs.

Usually because they are driving far too fast for the road conditions.

How well do sensors and vision systems handle winter conditions like snow and lack of lane markers?

Easy, my dumb level-0 car can tell me when it's icy. And finding lane markers is one of the easiest tasks in self driving (the hard part is knowing when to ignore them).

You're being downvoted for the flippant and dismissive tone of your comment, but I do wonder how computer-driven cars will determine when it is acceptable to violate lane markings and road signs. Boston in winter is more than just traction control. There are snow piles that might be icy, ridges left from a plow, shifting conditions, and bad visibility. I suspect it IS a hard problem.

> And finding lane markers is one of the easiest tasks in self driving

It's not a matter of "finding" lane markers. There are no lane markers visible after it snows.

Lane markings are a fraction of the triangulation.

We ourselves identify and confirm other urban waymarks via captcha which feeds the nav data -- bridges, signs, hills, hydrants, chimneys, lights. There is mass live verification from android auto in vehicles. There are many yearly layers of street view images and scans.

Right, and what about rural waymarks? A highway in the middle of nowhere at night during a snowstorm?

I don't think we'll see a system that can handle that in my lifetime.

I suspect it will know where the lane markings are better than human drivers. They are mapped ahead of time and the car can likely localize itself via other landmarks to determine where they are without being able to see them.

The harder part is driving like a human and detecting that a path has been made in the middle of two lanes in heavy snow and not obeying the lines at all.

The first idea seems like it would require a lot a lot of data stored in the car. Is it feasible? And even so, to be that dependent on matching up with existing pre-mapped data suggests a system that would be quite slow to roll out across a country.

If a pedestrian slips in deep snow while crossing a street and is no longer visible because the snow obstructs them, does the car see a clear path and kill someone or not?

Does a human see a clear path and kill someone or not?

So I’ve encountered this in real life.

The human driver detects the pedestrian, laughs at the fall, then get worried and wait for them to get up, because a human knows someone fell and didn’t magically disappear.

Doesn't seem like it would require AGI for a self driving system to "know" that someone fell and didn't magically disappear.

Challenging weather conditions mean human drivers become even more unpredictable.

Waymo et al will have to install snow tires or else no matter of traction control or even all wheel drive are going to help when your tires cannot find grip.

Source: grew up watching subarus do 360s on the freeway.

Reply on news.ycombinator.com