Porting the Perpetual Humanoid Controller (PHC) to MuJoCo for motion imitation, 70% done.
I replicated the state space of PHC in isaac to MuJoCo, right now it has reached ~98% success rate for single network (no PNN yet).
Code (work in progress) :
Biggest issue right now is I can't recreate the black-box PD controller from Isaac yet. Still investigating.
Training speed for motion imitator is about 5~10x times slower than in Isaac 🥲
If anyone wants to add MJX support to this repo, please DM/contribute!
Personally, in 2021 I was stuck at ~90% success rate for motion imitation (without using external forces) on AMASS for a year. Tried some crazy ideas like and almost gave up.
Thanks to Isaac Gym and its speed improvement, I was able to learn a ton...
...and design a better state space and training procedure. I was a little worried whether it will transfer to MuJoCo, but it did! A milestone to me!
Will continue to develop this repo to bring the full capabilities of PHC into MuJoCo. End-goal would be in-browser PHC wecam demo
@zhengyiluo
Great work! Have you ever tried outputting motor current (convert to torque for input into torque actuators) from the policy, as opposed to position?
@stevenuecke
Actually, I am trying that now (in Isaac Gym)! I find it performing worse than position control, though that could change with tuning.
In the SMPLSim: , We do support different controllers (including force control). Haven't tune them much though...