Jev controls every traffic light in this city when I turn it off the average wait time rises by more than 600% we need this rn

jev causes collisions even in your vid, and jev off is the most unoptimized traffic grid I have ever seen.

Try to better model the time it takes cars to accelerate, especially when they're not expecting a green light and suddenly it switches. Also take into account car braking time and light in-between time. I bet it will still be an improvement, but nowhere near as drastic as that.

Dude, just make some logic if n cars > xx then change colors. Or find the best rules with RL, which is an early problem for post-grad students studying RL

This would be a very good use case But then how would you be able to check all the cars? I guess that’s my only question

you can try it here:

You're skipping Yellow. That'll make a HUGE difference. Frequent flopping from green to red and you WILL have massive accidents without in the hour. Yellow is 3-6 seconds all by itself, because that's how long it takes to prepare to stop and clear the cars already in the crossin

so you want every car to be driven autonomously with perfect stopping and moving, this is a very good idea, but not with humans behind the wheel

try convincing a local government to give you data on this...i hate policy

I think jev is probably one of the first models in a long time that feels both easy and confidently applicable to the real world. LLMs are really intelligent, yet there’s still this lack of confidence in certain situations about whether you can actually trust the output enough to act on it. I think jev is gonna be a very big part of a pretty major paradigm shift. And it’s basically basal ganglia.

The light turns green and even in a 7+ car line the last one starts moving almost instantly. Try this in real life and you will see not everything that looks good on paper actually works because of us, humans.

we rlly don’t need this though, jev isn’t free and adding it is solving a problem we don’t have, is it cool? yes, would I want it? yes, do we need it? no

I agree this can be much better than just timer based traffic light Potential issues I can think of: 1. Real life movements are unpredictable. No car is taking left, right or u-turns here. 2. Car might break down, there could be obstacles. Those can increase lead time ..

In America they could have done that since the 70s. It is a simple problem. But they want Americans to burn more fuel. 😇

Oh yes lets put an AI classifier in charge of something that can be solved and verified by a Cs bachelord in a afternoon. We really need this...

First problem is data collection of traffic condition, simple traffic AI control could be easily programmed by hand, if there is perfect data.

SCATS does this globally pre AI

I hope you didn’t burn too much time on this -

For this type of use case, a dedicated ML model is better, cheaper, and faster.

@_chenglou Sick! China probably already has something like this btw

Does this take the physics of a semi trailer into account?

That's again a very good use case. Traffic is a big problem in all Indian metro cities like Bangalore, delhi. If this is implemented it will help reduce Traffic in real time

What exactly does "off" mean: a fixed interval, green until traffic is detected, or a random timer? 🤡

This sandbox needs a lot of work before you can claim any success here

mfers make glorified pacman and get high off of their own fumes

experimenting with the game 2048; i'm struggling to get jev to perform much better than random. would love to see a quick write up of your strategy

Peak idea to Make something with yes and no

is there a git repos?

600% is not a stat you bury in a caption. That's the whole pitch.

No, lol.

Let’s not do that.

Cool

not enaugh time for pedestrians to cross the street

does it know the position of each car ?