Hacker News Reader

Comment by jjav | Hacker News Reader

There aren't infinitely many scenarios to consider, but even if that's a figure of speech, there aren't thousands or even hundreds.

If there's ten kids nearby, that's basically ten path scenarios, and that might be reduced if you have great visibility into some of them.

> My brain might be able to quickly assess a handful, but certainly not all.

What would you do if you can't assess all of them? Just keep driving same speed?

If the situation is too overwhelming you'll almost certainly back off, I know I would. If I'm approaching that school block and there's like 50 small kids running around in all directions, I have no idea what's going on and who is going where, so I'm going to just stop entirely until I can make some sense of it.

View on HN

> here aren't infinitely many scenarios to consider, but even if that's a figure of speech, there aren't thousands or even hundreds.

There are a very, very large number of scenarios. Every single possible different state the robot can perceive, and every possible near future they can be projected to.

Ten kids is not 10 path scenarios. Every kid could do a vast number of different things, and each additional kid raises the number of joint states to another power.

This is trivially true. The game that makes driving possible for humans and robots is that all these scenarios are not equally likely.

But even with that insight, it’s not easy. Consider a simple case of three cars about to arrive at an all-way stop. Tiny differences in their acceleration - potentially smaller differences than the robot can measure - will result in a different ordering of cars taking turns through the intersection.

It’s a really interesting problem.

> Ten kids is not 10 path scenarios. Every kid could do a vast number of different things, and each additional kid raises the number of joint states to another power.

This is the difference between computing and humans. The car will attempt to compute all possible path scenarios because it has no instict, and it might not be possible to compute everything in real time so it might fail.

But the human will easily deal with the situation.

Try running through a sports field in an elementary school during lunch, full of unpredictable kids running around. Can you make it from one side to the other without crashing into a whole bunch of kids? Of course you can. You didn't need to try to compute an exponential number of scenarios, you just do it easily. The human brain is pretty amazing.

In fact no computer approach attempts to compute all possible path scenarios since we know that’s not tractable.

And current practical approaches are mostly end to end (or nearly) ML systems that do not compute a lot of alternative paths, and they work in approximately constant time independent of the scenario.

You strongly imply that computers can’t drive, but you could have written that in a Waymo.

It should be trivial for Waymo to implement a "drive carefully near schools" feature, and if really spicy "drive REALLY carefully near schools at these times" feature.

Safe driving starts with speed, lowering speed and informing the passengers seems like a no-brainer.

Feels like bitter lesson fodder to special case things like this

It was a figure of speech, but I think you're undercounting. When you consider interactions between all the things, even with just a handful of variables (and I think there are many more than a handful) you get a huge number of scenarios.

But you can group them into categories that can all be handled the same. For example, a child darting from behind a vehicle is a line-of-sight issue.

To fix that, you program the car to handle situations with obstructed vision, which will handle not just this specific scenario, but all relating to obstructed line-of-sight — basically slow down enough to be able to stop in time in case something jumps out from behind the obstacle.

Really though, this is less of an engineering problem and more of a social cost-benefit analysis one.

On one hand, I'd say hitting a kid at 6mph in the worst case scenario once in a blue moon probably isn't that big of a deal.

On another, someone here calculated that "even 1MPH slower would likely have resulted in no contact in this scenario".

So really, it's not possible to say whether this was handled properly or not without access to data only Waymo has and establishing some standard of how much injury we're okay with vs the impact on travel times. Remember, we're seemingly ok with ~40 000 americans dying every year due to car transportation.

Reply on news.ycombinator.com