Some questions regarding the retargeting pipeline

[CavalloneChen]

Congratulations on the V2 update! I have some questions regarding the motion retargeting (from AMASS to Unitree G1) pipeline:

1. Is the coordinate system of the retargted motions different from that of the PHC retargeted motions? I tried to run PHC training on the motions retargeted from ProtoMotions (V1 pipeline to be specific), but the result seems to be very weired.
2. Is there any instructions on how to utilize the brand new Mink pipline to retarget from AMASS to Unitree H1/G1 robots? Will the output contain the same set of keys as ProtoMotions V1 pipeline? And will the coordinate system be the same as V1 ?

Thank you for your time.

[tesslerc]

I haven't looked into the G1 yet, so I can't provide a definite answer.
What I can say is that the current implementation of MotionLib expects the data to adhere to the kinematic joint ordering via DFS traversal (isaacgym standard).

The Mink pipeline is here.
You'll see retarget_motion which is the main loop.
You need to define which joints to track and the root offsets from AMASS to G1. You can weight the joints based on importance, and add parameters such as velocity limits to get smoother results.

I believe we should have the G1 officially integrated within ~2 weeks.

[CavalloneChen]

Thank you for the great news. I noticed that CMU's LeCAR-Lab has recently unveiled two interesting repos https://github.com/LeCAR-Lab/HumanoidVerse and https://github.com/LeCAR-Lab/ASAP . Maybe you can find some G1 related resources there.