Great work! Have you considered releasing the G1 training code?

[sjtuyinjie]

The pretrained checkpoint on G1 performs amazingly. I tried to train this policy on G1, but even with retargeting, the G1 robot behaves very oddly, though still succeed to finish tasks sometimes. Would you consider releasing the G1 training code? I'm very eager to know you handle the embodiment gap between SMPLX and G1.

[Sirui-Xu]

Hi @sjtuyinjie,

Thanks for raising this, and for your contributions to the project. Yes, we’re planning to release the G1 training code in the next few weeks, after we get through a couple of upcoming deadlines.

At a high level, the key for bridging the SMPL-X → G1 embodiment gap is to avoid dense full-body positional tracking. Instead, we use a sparse positional reward on a few key effectors (e.g., feet and hands), and pair it with full-body rotational constraints to keep the overall motion consistent. The remaining alignment is largely corrected by contact & interaction rewards, which help the policy self-correct when pure tracking would lock it into unnatural poses due to the embodiment gap.

In the meantime, if you’re comfortable sharing a bit more detail about your setup, e.g., retargeting method, reward design, or failure modes, I’m happy to take a look and help troubleshoot.

Hope this helps, and thanks again for your interest!

[sjtuyinjie]

Thank you very much for your assistance!
To provide more context: I am retargeting SMPLX reference motions to the G1 robot using an approach similar to GMR[https://github.com/YanjieZe/GMR], and I employ full-body keypoints for reward computation. I have organized the relevant environment code (including the retargeting function), configuration files (with reward weights), and a video which show failure cases, all of which I sent to your email.

Please take your time to review only when convenient. I would greatly appreciate it if you could share the G1 training code directly—I will handle the setup and execution on my own end.

That said, if you’re too busy, no worries at all! I’ll go ahead and try the methods you mentioned above on my own.
Thanks again for your help!