H1_2 Walking issues

[NirajPudasaini]

We trained the policy for 10k steps. When the policy is played, it works fine in the mujoco, but there is an issue in the real deployment. Please find the attached video. Any suggestions?

mj_lab_walk.mov

[NirajPudasaini]

The policy from Unitree RL Gym - https://github.com/unitreerobotics/unitree_rl_gym/blob/main/deploy/deploy_real/deploy_real.py

works with the H12 robot.

[klavier1127]

It may be helpful to check whether there are any differences between the training parameters and the deployment parameters. Based on the video, it is possible that some deployment parameters are not consistent with those used during training. I would recommend carefully comparing the training configuration and the deployment configuration to see if there are any mismatches.

[NirajPudasaini]

@klavier1127 we didn't change anything. We just ran the given code in the repo for Unitree H12. What deployment params do you think should we touch? Kp/Kds? or armatures? Did you guys changed the param in the deployment demonstration that you have in the repo?