A potential error in humanoid_amp

[Charrrrrlie]

Hello,

When calling _reset_ref_state_init function, the rigid_body_state in _set_env_state will not be updated. And the base observations in humanoid_amp_base will be calculated with a default/old rigid body state.

That may cause a potential error when we leverage the base observations in the simulation. e.g. call self._rigid_body_state in pre_physics_step. (The current AMP algorithm uses the observations in humanoid_amp.py, thus it won't effect on the results)

Moreover, I cannot add functions like set_rigid_body_state_tensor in _set_env_state since the simulation is under Physx. So…In my case, I have to manually change the value of self._rigid_body_state and force it not to refresh in reset_actors.

Is there any solution if it is indeed a potential error?

Thanks for any help!

[heyuanYao-pku]

Same question, any solution now?

[Charrrrrlie]

Same question, any solution now?

Not yet :-(

And there are so many differences compared with the implementation of ASE. e.g. this problem you may also need to pay attention to it.