The necessity of action signal delay for sim2real

[zsheddie]

Dear mujoco_playground team,

Firstly, thank you for your excellent work! I'm trying to use this framework on our own legged robot and it works well in the simulation environment so far, but we failed in the sim2real part. I've noticed that most open-source frameworks have added motion delays and sequential observations during training. Therefore, I'd like to ask a few questions:

1. Did you complete the Sim2Real work on G1 robot without using motion delays during training? If so, how did you solve the delay problem on real robot motors?
2. I found a task in mujoco_playground called "manipulation/franka_emika_panda_robotiq/push_cube" that includes motion delays, but when I tried to add motion delays to our bipedal robot using the same method, it failed, and the reward curve didn't rise properly (it works fine without lagged actions). Do you plan to add code related to motion delays in future updates? Or could you give any sim2real advice about that?

Thank you very much in advance! :)