Why only lower half of snu_humanoid is used?

[AleximusOrloff]

Is there some issues with convergence? Has someone tried to train full snu_humanoid? At first glance it seems that unnatural running style could be caused by different mass distribution and other consequences of absence of upper torso, head and hands Its not issue actually, I just very curious :-)

Anyway thanks for really great job done, its actually amazing!

[eanswer]

Hi, thanks for mentioning this more complex example. We didn't try the full snu_humanoid yet. It would be definitely interesting to see whether the running gait can look better with the upper body there.

[AleximusOrloff]

I found line: self.filter = { "Pelvis", "FemurR", "TibiaR", "TalusR", "FootThumbR", "FootPinkyR", "FemurL", "TibiaL", "TalusL", "FootThumbL", "FootPinkyL"} @https://github.com/NVlabs/DiffRL/blob/main/envs/snu_humanoid.py#L34, So one who did it, did it definitely on purpose. I just curious what that purpose was, were there any issues with full body simulation or something else? I'm away from my workstation, so cannot check it on my own :-(

[eanswer]

We added this filter in our very first experiment with muscle-actuated humanoid to start from a fewer-DOFs version of the problem. And we forgot to change it back to the full-body version after we made it work on the lower-body version. You are more than welcome to try shac with the full body version.

[ViktorM]

The simulator has some limitations, like the lack of support for self-collisions. The full humanoid model could be simulated and trained, but much more effort and time would be required for additional reward tuning to make the gait look biologically plausible, especially taking into account the absence of self-collision.