Hacker News Reader

Comment by swalsh | Hacker News Reader

This might be a stupid question, because I don't really know what I'm talking about.. But is there a way this could be used to improve physics engines? My understanding is that a physics engine generally uses continuous collision detection. I don't know if noise plays a roll in casting for collisions, but if you could cast less rays, and fill in the holes in the same way... that seems like it could improve performance there too?

View on HN

Aren't the collision models already pretty simple as is? Simple shapes like boxes, spheres and lines, for which the collisions are fast to calculate. Don't know how it works if you have full blown meshes colliding though...

I wouldn't be surprised to see something like that.

Remember that Krishevsky's software implementation for Alexnet of making neural networks work with GPUs, wasn't initially intended by Nvidia either. But, look how that worked out.

The problem is that you might get a lot of instability.

The most obvious case: stack 20 boxes one slightly above other and let them fall.

If you will have random noise in there the simulation can become unstable/explode.

Yes. Here's a video that describes an approach using neural networks to estimate physics instead of using a traditional simulation: https://www.youtube.com/watch?v=atcKO15YVD8

Yes. I think I’ve seen some preliminary papers with related ideas (because that seems like where ML is at) but I don’t remember anything substantial.

It could be an interesting area of research. Start branching out from ray tracing to other integration problems.

I dont know, it depends on implementation. Collision detection doesn't involve any graphics, it might not paralellise well. so it might still be easier on the CPU.

Reply on news.ycombinator.com