retargeting performance of g1 and t1

[gapbridger]

dear authors

thank you for open-sourcing such an incredible work including omni-retarget and a complete motion tracking pipeline
i was trying to play around with the retargeting pipeline and discovered that for lafan retargeting, g1 and t1 has moderate performance in terms of the retargeted foot angle with respect to the ground. the t1 robot has a large angle with respect to the ground. and the g1 robot also has a moderate angle. but in gmr, it seems the angles could be tuned to a small value.

i am curious is there any configurations that we could tune to improve the performance especially for the foot contacting angles. in gmr, it seems we could improve the performance by tunning the ikconfig jsons.

thank you for your work and looking forward to your suggestions.

[lijiaman]

Hi, our current retargeting pipeline is position-based optimization. Could you share more details about the expectation? Maybe share some visualizations? Thank you.

[gapbridger]

hi Jiaman, thanks for the response, here are some visualizations regarding foot angle of g1 on lafan dataset walk1_subject1

these two figures are when i perform retargeting but default_scale_factor set at
DATA_FORMAT_CONSTANTS: dict[str, FormatConstants] = { "lafan": { "default_scale_factor": 1.40 / 1.7, }, "mocap": { "default_human_height": 1.78, }, }

it seems the ankle is off the ground. when i adjust the default_scale_factor to 1.23 / 1.7, the result is

it seems ankle pitch is much better, but there's still no constraint on the final ankle roll angle. i expect the foot to be flat on the ground when it touches the ground. one thing i tried that is effective is to shrink the ankle roll joint range. in this case it seems the retargeted ankle roll value is relatively smaller. but i think this is not a systematic way to handle this issue.

in summary, here is my two issues

1. it seems we need a systematic instruction on how to tune the scale factor or maybe it's just a tuning parameter that need to be highlighted?
2. is it possible to add an optional foot orientation constraint in the optimization problem? if you could provide some hint on how to add this constraint, i could implement one and test out the result and maybe propose an PR?

thank you.

[gapbridger]

also, i would like to point out that the logic of estimate_human_orientation in uitls.py seems odd and not stable for custom data sources (although i still need to get to the bottom of this). but could we simply use the orientation of of the base link (pelvis or hips) of the captured data as the estimated orientation to properly align robot?