I've been pretty busy, but I have a couple of hours. Going to start writing a friction physics engine from scratch in 2D
I got the deepmind baseline training in ~15 minutes, takes 30 minutes on my machine. ~470k SPS on my set up, still warp
But I need it to be faster
Users are excited about building custom 2D physics engines and rigid body simulations with particle collisions because the technical details and parallelization potential feel inspiring and approachable.
No Digg Deeper questions have been answered for this story yet.
what if you simulated every single particle of your robot would it be parallelizable then?
I'm trying to accelerate RL for robotics. I think that everyone is doing it wrong. My goal is to use my influence to help change the trajectory that things are going by creating end to end reproducible loops that help unclog people's brains from the memes
I've been pretty busy, but I have a couple of hours. Going to start writing a friction physics engine from scratch in 2D
I got the deepmind baseline training in ~15 minutes, takes 30 minutes on my machine. ~470k SPS on my set up, still warp
But I need it to be faster
cite @YuanmingH thanks for sharing https://yuanming.taichi.graphics/publication/2018-mlsmpm/
what if you simulated every single particle of your robot would it be parallelizable then?
After looking into common contact solvers for 5 seconds I realize that this is the worst way possible to solve this problem
This is the sort of programming that LLMs don't really help you with. Just C neovim and some music.
I'm going to make a leg that kicks a soccer ball in 2d, an env that can be learned by a single GPU. I'm going to train it, end to end, and use my learning from that to do 3D
Unfortunately all the stuff out there is way too stuck on hifi sim accuracy. People trying to solve the sim2real problem by spending more compute on the sim. So I need to do it myself. I'm going to do in in 2d to learn, then take it to 3d. It's not going to be easy
this is so stupid. just voxelize literally everything and solve it in parallel
god damn it i love game engine people man
@yacineMTB @jeankaddour @karpathy @yacineMTB how would you set this RL env up to scale to millions of sps? This is discrete event and not physics timesteps
https://github.com/yuanming-hu/taichi_mpm
One reason I'm guessing they are slow is circular dependencies requiring iterative solvers. My foot hits the ball, but also, the ball hits my foot, and there is friction, and we need to resolve each of these as a function of the other.. so why is it fast in real life?
The world god gave us is (probably) some sort of sick cellular automata, each local region is responsible for computing itself. No circular dependencies
The core exploit I'm attempting is that all of the tricks that people use for sim2real basically shake out to what happens when you write a low fi simulator. Mujoco playground's G1 baseline has obs noise, for example. So why spend all this time computing accuracy?
One reason I'm guessing they are slow is circular dependencies requiring iterative solvers. My foot hits the ball, but also, the ball hits my foot, and there is friction, and we need to resolve each of these as a function of the other.. so why is it fast in real life?
cite https://developer.nvidia.com/gpugems/gpugems3/part-v-physics-simulation/chapter-29-real-time-rigid-body-simulation-gpus
Unfortunately all the stuff out there is way too stuck on hifi sim accuracy. People trying to solve the sim2real problem by spending more compute on the sim. So I need to do it myself. I'm going to do in in 2d to learn, then take it to 3d. It's not going to be easy
gyatt damn this shit parallelizable as heck
cite https://developer.nvidia.com/gpugems/gpugems3/part-v-physics-simulation/chapter-29-real-time-rigid-body-simulation-gpus
You might say that, because they're using an asymmetric critic / actor, the critic needs the pure accurate state.. but I've beaten their wallclock training time by 2x without the asymmetric critic! Mine uses 93 obs for both, not 93/207. Is it actually necessary?
imo no
@yacineMTB You might need a significantly large quantum computer for that.
@yacineMTB This might be worth listening to on 5x:
https://youtu.be/Kvsvd67XUKw?is=dIXAivBdYKOvEC7D
this is literally fine
https://github.com/yuanming-hu/taichi_mpm
