Hacker News Reader

Comment by bronco21016 | Hacker News Reader

Waymo’s Michigan testing facility is very close to my home. I’m really curious to see how they do on snowy pothole ridden roads in Michigan. I’ve seen very few of them on the road this summer but if there’s supposed to be heavy testing and validation going on perhaps we’ll see an army of them on the roads here soon.

I’m curious how self driving focused groups are solving lane detection and managing surrounding traffic on snow covered roads. It’s not uncommon for a multi lane highway to be completely covered with snow and multiple tracks for different ‘lanes’ that overlap. It’s also not uncommon to see cars just making their own lanes. This occurs even after the snow has stopped flying.

View on HN

"I’m curious how self driving focused groups are solving lane detection and managing surrounding traffic on snow covered roads. It’s not uncommon for a multi lane highway to be completely covered with snow and multiple tracks for different ‘lanes’ that overlap. It’s also not uncommon to see cars just making their own lanes. This occurs even after the snow has stopped flying."

Yep. As a Michigander, I totally agree. You see 6 to 12 lane roads reduced to "random" 2 lane roads when there is a good pileup of snow. And the snow drifts make it even worse!

Humans are good at improvising. Not using all the lanes when the lane markings are covered by snow, and the road is slippery, sounds like a reasonable strategy. And nobody ever told us that's what we should do.

With autonomous vehicles, you need to plan every situation in advance. Our current AI technologies can't improvise. It seems to me that current AI is basically just a decision tree with some neural networks sprinkled on top.

We don't want machines to improvise because we want to be able to test their whole gamut of behavior as thoroughly as possible. We can build AI that can improvise, but that's not the kind of software you want to be liable for as a manufacturer.

I am not at all convinced that improvisation is an easy add-on. While AI has demonstrated some creative ways to satisfy goals that have left options open (sometimes inadvertently), real-world improvisation has to be appropriate for the situation and the unusual difficulties it poses, and, in the case of driving, it must not create additional problems or (of particular relevance here) present significant dangers.

In general, I think, improvisation requires causal reasoning. It also often needs a broader knowledge than that needed for the nominal task (for example, when driving on a potholed road, one needs to have some understanding of how it affects the car's response, and sometimes of issues such as ground clearance.) Constraining the vehicle to avoid such issues would constrain the option to improvise.

>We can build AI that can improvise

Do you have any sources demonstrating examples of AI with true improvisational capabilities?

More like decision tree/rules engine with pattern recognizing neural networks underneath. You detect features and then take largely pre-programmed actions based on those features.

The larger the "largely pre-programmed", the less it is like improvisation, until, at some point, it isn't. As I suggested above, effective improvisation probably requires causal reasoning.

Reply on news.ycombinator.com