robfiras / loco-mujoco

Imitation learning benchmark focusing on complex locomotion tasks using MuJoCo.
MIT License
475 stars 38 forks source link

How do you get the ground truth expert data for humanoid robots? #7

Closed wenshuaizhao closed 7 months ago

wenshuaizhao commented 7 months ago

Hi,

Interesting work! But I didn't get the information from the paper, how do you get the ground truth data with actions for humanoid robots, such as Atlas?

Thanks

robfiras commented 7 months ago

Hi! Thank you for your interest in this work! We are currently preparing the datasets with ground truth data. They will be uploaded in the upcoming days. I will notify you! Best, Firas

wenshuaizhao commented 7 months ago

Hi,

Thanks for your prompt reply! This dataset would be awesome. But I would also like to know how you will obtain the ground truth data. Specifically,

Thanks!

robfiras commented 7 months ago

We optimize the motion capture towards a base humanoid model using Qualsys' optimization software (see paper for reference). Then we manually adapt this data to the different humanoid embodiment using affine transformations. To get the actions, we use the baseline algorithms provided to learn a policy that imitates the motion capture data as good as possible. We will provide the experimental setup, such that you can train an agent yourself if you want.

wenshuaizhao commented 7 months ago

Thanks for your detailed explanation!