researcher friend of mine shared this with me; comfree-sim. which is basically existence proof of my hunch: mujoco has terrible performance, and there are probably huge huge gains you can make in training, especially considering domain randomization
Positive users report speedups and better policies with ComFree-Sim over MuJoCo baselines while negative users argue video game physics sims are faster and speed alone is not what matters.
real robotics RL hasn't been tried cite: https://irislab.tech/comfree-sim/
let's do some math what is 2.5 times what seems to be worst case 2? this is the thing with compute multipliers they scale. 5x the speed if this works
let's do some math what is 2.5 times what seems to be worst case 2? this is the thing with compute multipliers they scale. 5x the speed if this works
I was already able to get up to ~250k sps from the ~100k sps baseline (2.5x speedups). My policies are now scoring better than the mujoco baseline with only twice the training time. This is *with* a numerics bug in the core training loop which i'm fixing!
@yacineMTB video game physics sims are even faster. But speed isn’t everything in physics sims…
@yacineMTB Just like Factorio IRL, or using a diamond pick in Minecraft at all times (finding enough extra diamonds to easily pay for the extra cost of tools)
@yacineMTB
@yacineMTB Why does five times faster seem 100x faster? I watched it fall 5 times compared to the other, yet still there's a disconnect 🤔
@yacineMTB Are we pivoting from doordash drone swarms?