Users praised the MuJoCo G1 baseline training achievement in 15 minutes 28 seconds because it unlocks rapid iteration loops for prototyping custom rewards.
No Digg Deeper questions have been answered for this story yet.
Most Activity
For SPS, wins were removing python overhead entirely (no python running during training) and reducing computation used for physics. For training hypers, biggest win was fixing a numerics bug on pufferlib to sweep better
here's trasnfer to mujoco cpu on wasm
Just barely missing my 15 minute goal for mujoco playground's g1 baseline. This is walking + perturbations, nothing changed in the env set up
Going to publish all of this up on github soon. Updated dingbotics dot com
@yacineMTB @_Stocko_ based + thanks for sps optimization breakdown
@yacineMTB looks like me trying to dance
@yacineMTB Niiice, almost there robot gains be like that
@yacineMTB Trying to train faster? (I’m not an expert in this shit don’t roast me)
@yacineMTB my man is breakin it DOWN
@yacineMTB 15 min is the iteration-loop unlock. coffee-break RL. every grad student can prototype a custom reward in the same shape as a compile-check loop. field-level effect is more weird papers, faster.