Joseph Suarez's PufferLib powers nanoG1 to train a Unitree G1 humanoid robot to walk in under 60 seconds

@JulianSaks great work! can you test this on the mujoco playground baseline env to see if handles perturbations & additional joints?

that's cool. puffer wins again!

@JulianSaks 🦾

ultra fast RL achieved! g1 policy trained for < 60s at 1.3M SPS on a single rtx pro 6000. all open source!

http://nanog1.com

code: http://github.com/kingjulio8238/nanoG1.git

HF checkpoint: http://huggingface.co/kingJulio/nanoG1

mini write up: http://juliansaks.com/feed/nanog1

@JulianSaks This is cool, would love to see it transfer on the real G1.

@JulianSaks Amazing work. This seems hyper-optimized for one robot's kinematic structure and may not transfer as cleanly to other morphologies. Nevertheless, curious to see this applied across various kinematic trees!

Also, the nano in nanoGPT meant simple, not fast 👀

@jsuarez thanks! looking forward to PufferLib 5.0

@vishcomestrue yes its highly optimized for g1 not a general purpose implementation but also if you have one embodiment that you care about it works. also who said making training fast can't be simple 😜

@JulianSaks fair point tho ✋😄🤚

@yacineMTB will do. I know for sure it doesn’t handle perturbations

@roblee_rl will be deploying it soon!

@NVIDIAAI @JulianSaks 💪🏿🔥🔥