Comment by michaelt | Hacker News Reader

Comment by michaelt | Hacker News Reader

Parent Full thread

Imagine you're driving down a street when a kid playing on the sidewalk runs behind a truck and disappears from your view.

If the kid appears in the road from behind the truck, the computer can handle slamming on the brakes very easily - probably faster than a human can - without understanding the kid any better than a group of LIDAR points or a rectangle of pixels labelled 'obstacle' by a neural network.

But if you want to brake before the kid appears from behind the truck? For that, a fully attentive human driver will be making a bunch of estimates about what the kid is doing, whether they seemed to have noticed the car, how old they were, and so on. In other words, applying a theory of mind.

Needless to say, if the latter is a must-have feature, that's a pretty hairy problem.

Of course, it's possible the decrease in deaths from being fast on the brakes in simple situations will outweigh the increase in deaths from lacking a theory of mind in complex situations. If that's the case, the self-driving car programmer's job would be a good deal simpler!

There's no need for a theory of mind - the self driving car can identify the kid as a pedestrian, recognize that it started moving in a possible collision course, then disappeared, thus prompting either slowing down to a non-fatal speed until the truck has been passed.

Children playing in the street are a common occurrence in residential areas, I see no reason why you would not develop a set of rules and heuristics to handle them. Identifying a pedestrian as child, and knowing whether it is running or playing, is well within the capabilities of modern computer vision.

It's worse than that. That small thing heading toward the road... was that a leaf blown by the wind, or was it a ball? If it's a ball, you'd better be already braking, because a kid is likely coming right behind it, and paying attention to the ball rather than the road. If it's a leaf, though... you don't want to hit the brakes for every blowing leaf.

even worse: sometimes kids actually do chase leaves on the wind, or butterflies.

> There's no need for a theory of mind - the self driving car can identify the kid as a pedestrian, recognize that it started moving in a possible collision course, then disappeared, thus prompting either slowing down to a non-fatal speed until the truck has been passed.

So ... a theory of mind.

Shouldn't it be possible to train for these scenarios using imitation learning and expert demonstrations? For example, Tesla seems to save replays of when its vehicles would have behaved significantly differently compared to how the driver actually behaved, and this data is supposed to be quite useful.

Is this type of crowdsourced driving data a crucial part of achieving Level 4+ self-driving? If so, it seems that driving around the same six city blocks of SF or cruising down Central Expressway in MV is going to produce diminishing returns in terms of producing measurable progress.

> For example, Tesla seems to save replays of when its vehicles would have behaved significantly differently compared to how the driver actually behaved, and this data is supposed to be quite useful.

I don't think such a system would catch a false negative like the above, where the human would slow down cautiously but the self-driving system would do nothing. That situation is indistinguishable from a human slowing down to read house numbers.

To realize the problem, the system would need a full model of "what would the car be doing if not for the human input" in order to find a later point of alarming divergence.

I thought waymo already displayed this ability? That's object identification and prediction, it's not impossible to do.

Reply on news.ycombinator.com