Leg posture issues and regularization reward convergence problem in g1_ground training

[vain85]

Thank you for the great work on this project!

I tried training the g1_ground task, and here are the results after 10,000 iterations. I noticed that the leg posture does not look very reasonable, and the knee joints seem to be close to the lower limit (around -0.087267).

In addition, within the regularization reward group, some of the rewards did not converge and instead showed a downward trend. I am not sure if this behavior is expected or if there might be an issue in my setup.

I’ve attached the reward plots below for reference.Could you please take a look and let me know what might be going wrong?

Thanks again for your efforts and contributions!

[TaoHuang13]

Hi. I think it is normal to observe some training curves decrease, as not all rewards are consistent with each other.

Regarding the lower-body postures, many users obtain different behaviors using their own machines. This may be caused by the randomness of different machines.

[vain85]

Hi. I think it is normal to observe some training curves decrease, as not all rewards are consistent with each other.

Regarding the lower-body postures, many users obtain different behaviors using their own machines. This may be caused by the randomness of different machines.

Thank you for your reply!

The 23-DoF robot performs quite well. However, when I tried training the 29-DoF version, the posture looked very strange, and the robot could not stand properly.

Since the waist function currently only constrains the yaw direction, I did not add the two new DoFs of the 29-DoF model to the waist configuration in the .cfg file. I thought the dof_pos_limits settings might handle these constraints, but the results were not satisfactory.
def _reward_waist_deviation(self):
wrist_dof = self.dof_pos[:, self.waist_joint_indices]
reward = (torch.abs(wrist_dof) > 1.4).float()
return reward.squeeze(1)
I’ve attached the reward plots below. Could you please take a look and let me know if there’s anything wrong, and whether the waist function needs to be modified to account for the additional DoFs?

Thanks again for your time and support!

[TaoHuang13]

Hi, this is a good observation! I think you should additionally add two reward functions about waist pitch and roll, which should be more constrained than the dof_pos_limit reward. vain85 ***@***.***> 于2025年10月10日周五 21:57写道：

…

*vain85* left a comment (InternRobotics/HoST#31) <#31 (comment)> Hi. I think it is normal to observe some training curves decrease, as not all rewards are consistent with each other. Regarding the lower-body postures, many users obtain different behaviors using their own machines. This may be caused by the randomness of different machines. Thank you for your reply! The 23-DoF robot performs quite well. However, when I tried training the 29-DoF version, the posture looked very strange, and the robot could not stand properly. Since the waist function currently only constrains the yaw direction, I did not add the two new DoFs of the 29-DoF model to the waist configuration in the .cfg file. I thought the dof_pos_limits settings might handle these constraints, but the results were not satisfactory. def _reward_waist_deviation(self): wrist_dof = self.dof_pos[:, self.waist_joint_indices] reward = (torch.abs(wrist_dof) > 1.4).float() return reward.squeeze(1) I’ve attached the reward plots below. Could you please take a look and let me know if there’s anything wrong, and whether the waist function needs to be modified to account for the additional DoFs? Thanks again for your time and support! 1.png (view on web) <https://github.com/user-attachments/assets/3257c9d4-2b6b-4a4a-b1f2-9f26e5a5075b> 2025-10-10.21.52.44.png (view on web) <https://github.com/user-attachments/assets/6fb8bd58-e66c-4382-87f1-8aabb4465a9d> 2025-10-10.21.53.04.png (view on web) <https://github.com/user-attachments/assets/49d761d9-7776-4fb6-add1-cded92f2b387> 2025-10-10.21.53.14.png (view on web) <https://github.com/user-attachments/assets/5f6d9810-6877-4397-88df-8fb1c78b141f> 2025-10-10.21.53.50.png (view on web) <https://github.com/user-attachments/assets/15fb5927-7ca2-490c-a2e4-ed22e9063b6c> — Reply to this email directly, view it on GitHub <#31 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AOIKLIRNDEZIWJTWJ27I2W33W63MHAVCNFSM6AAAAACHW3PNECVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTGOJQGI4TMNRYGY> . You are receiving this because you commented.Message ID: ***@***.***>