[Question] Possibility of replacing Unitree G1 USD in IsaacLab with Unitree's official USD

[jnskkmhr]

Question

Hello, I recently started training a PPO locomotion policy using the Unitree G1 humanoid robot.
I noticed that the USD files provided by IsaacLab (g1.usd and g1_minimal.usd) have different configurations, especially the waist position, compared to the official USD files released by Unitree.
Since the policy cannot be deployed on the actual hardware in its current form, I think it would be better to replace it with the official USD model (either the 23-DoF or 29-DoF version).

Also, unfortunately, the G1 velocity-tracking task does not produce a functional policy when using Unitree’s official G1 asset.
Specifically, when I replaced the ArticulationCfg in IsaacLab with the configuration provided by Unitree (https://github.com/unitreerobotics/unitree_rl_lab/blob/61bfba15d35f1a93e3bacab85fe06b31643c83b7/source/unitree_rl_lab/unitree_rl_lab/assets/robots/unitree.py#L298), the policy failed to converge within 1500 iterations, which is the default maximum for the G1 velocity-tracking task.

Below, I have attached a one-to-one comparison between Unitree’s official 23-DoF G1 asset and the version provided in IsaacLab.
It would be greatly appreciated if the NVIDIA team could retune the rewards, observations, and other task settings to accommodate the updated asset.
Thank you for your understanding.

Build Info

Describe the versions that you are currently using:

• Isaac Lab Version: 2.3.0
• Isaac Sim Version: 5.1

[RandomOakForest]

Thank you for posting this. The team will review it. I'll move this post to our Discussions for follow up. Thank you for your interest in Isaac Lab.

[jnskkmhr]

Hi, actually the policy I trained looks good.
I needed to change a lot in

• observation space (need history data, assymetric observation space)
• reward design

These changes are inspired by https://github.com/Hellod035/LeggedLab
Also, training takes way longer as I needed 20K iteration on RTX4090 which is approximately 10 hours training.
Should I make PR?
And you might want to tune further as gait at the beginning of walking still looks weird (takes 1-2s before making step)

g1_locomotion.mp4