Question about training pipeline sharing and improving fast turning during quadruped demo retraining

[darkwuta]

Hi, thank you for sharing this excellent project. The overall work is very impressive, and I learned a lot from it.

I am currently retraining the quadruped demo in this project, and I have run into some difficulties with fast turning behavior. I would like to ask two questions:

1. Do you have any plan to share the training pipeline or more detailed training procedure for this demo?
2. Do you have any suggestions for improving fast turning during running?

What I have tried:

• I used the Yaw difference between the future first frame and the future 16th frame to simulate the controller signal.
• This seems to help turning while walking.
• But for fast turning during running, the result is still not good, and drifting remains obvious.

Here is my current situation:

• Slow or moderate turning during walking can be improved to some extent.
• However, during running, fast turning is still difficult to learn.
• The policy tends to drift during fast turns.

Any advice or experience would be greatly appreciated.
Thanks again for building such a great project.