The policy trained by DreamWaQ exhibits a self-rotation behavior

[zxzdsdx]

Hello, while reproducing the go2_dreamwaq task, I found that the trained policy makes the quadruped spin in place. Have you encountered this issue?

[lupinjia]

Don't know whether you mean in training or in validation. Paste the code and the video

[zxzdsdx]

After completing training with "python legged_gym/scripts/train.py --task go2_dreamwaq", when I applied the policy to MuJoCo, I found that in the first three seconds of the video—the robot dog exhibits spinning behavior. However, behaviors like climbing stairs and walking are fine.

DreamwaQ_walk.mp4

[lupinjia]

try to add a dof_close_to_default reward, weight specified to -0.05

[zxzdsdx]

Thank you. I'd also like to ask: do you have any suggestions regarding sim2real? Will motor delay have a significant impact?

[lupinjia]

Add more domain randomization terms. Motor delay does influence the deployment performance, but it depends on the communication quality of your hardware