Those assumptions make you understand the problem, it is the scientific method after all.
If you consider ALL the variables, you would end up with something like the Standard Model Lagrangian, and even that makes some assumptions (explicitly or implicitly) in order to work.
Have we uncovered how crowds swarm digital content? It feels like these hidden algorithms trapped away inside Meta probably have the answer as well. Those would be meaningful research papers from Meta if they were ever to do it.
I’m sure how they swarm digital content is well-understood (internal to these companies, at least). They’re interested in the why, though. Being able to make viral content on demand would be incredibly valuable.
I don't know about straight up not caring if you walk into people along the way but the average separation distance we like to keep will certainly vary by place and that may well vary when the switchovers will occur. Maybe 15 degrees in one place with larger spacing vs 10 degrees in a place where people tend to walk closer together.
It's curious how Man does these things naturally, but needs many man-years of studies and experimentation to arrive at gross approximations of how we function.
All the logic of how we move in crowds is in our brains, we apply it every day. Yet we don't know it explicitly. Something that we do consciously, even.
But, in this case, it's rather emergent behaviour that's pointed out. We know how we behave how the person in front of us steps out of line: we move to the side ourselves or slow down. What's less obvious is when and how this cascades onto the whole crowd.
I don't know exactly what their physical analogies with fluid dynamics are, but this sounds like turbulence.
I often wonder how we impart this kind of knowledge to AI.
Imagine someone steps into a subway carriage with only one other passenger in it. They're not likely to sit directly beside or opposite the other person. Would a robot follow this kind of etiquette?
I suppose if you emphasize personal space in the training data, but would the developers realize they need to do that?
I mean, that actually is A good point about training data. There is more publicly available data of people acting weird than acting normal. We don't usually upload videos of people just existing.
Maybe one of the reasons it took so long is that MIT orientation included the directive that you're supposed to walk on the right-hand side in hallways. Otherwise, a mathematician could've just perched on Infinite Corridor's overhead vantages for a week, to come to recognize the natural patterns. :)
A "lane" requires people to simply walk behind someone else. Even if everyone is going in the same direction (or else the opposite one), staying in a lane requires patience. When some people are in great hurry compared to others, they start to look for ways of getting around their lane predecessor, disrupting the lanes. People not in a hurry will prefer to follow someone going the same way, because that person shields them from having to awkwardly dodge oncomers left or right. Either that person handles it directly, or else their own predecessor, etc.
What was really weird about the press release was there's absolutely no discussion of whether lanes are desirable. There's sort of a vague implication that they must be, but I tend to suspect they aren't.
21 comments
[ 7.9 ms ] story [ 68.6 ms ] threadThose assumptions might be good enough for modeling students walking to & from classes at MIT. But more generally? Nope.
If you consider ALL the variables, you would end up with something like the Standard Model Lagrangian, and even that makes some assumptions (explicitly or implicitly) in order to work.
See for example https://m.koreaherald.com/article/10447649
All the logic of how we move in crowds is in our brains, we apply it every day. Yet we don't know it explicitly. Something that we do consciously, even.
But, in this case, it's rather emergent behaviour that's pointed out. We know how we behave how the person in front of us steps out of line: we move to the side ourselves or slow down. What's less obvious is when and how this cascades onto the whole crowd.
I don't know exactly what their physical analogies with fluid dynamics are, but this sounds like turbulence.
Imagine someone steps into a subway carriage with only one other passenger in it. They're not likely to sit directly beside or opposite the other person. Would a robot follow this kind of etiquette?
I suppose if you emphasize personal space in the training data, but would the developers realize they need to do that?
If you train it on CCTV footage from inside of subway carriages: yes, I suppose.
Will robots be super awkward and/or aggressive?