Question about the learned gait frequence

[cyoahs]

Hello, this implementation about amp for quadruped robots based on isaacgymenvs is really exciting!

However, I have a little question about the result. I watched the video here frame by frame and found that the gait frequency from the learned policy is about twice of the mocap data. Same phenomenon is also found in the implementation of AMP_for_hardware (both video and paper).

Perhaps it is part of the feature of the AMP method, such large deviation on the frequency of the motion is not found in the original AMP paper. Have you noticed this problem or have you got any idea about this?