[Question] Best approach for H1_2 custom motion: Isaac Gym vs Isaac Lab

[dscmbbltd]

Question: Best approach for H1_2 custom motion - Isaac Gym or Isaac Lab?

Hi all,

Thanks for the great work on this repo.

We're looking to create a custom walking gait that includes coordinated arm movement on the H1_2. Given that this repo is built for Isaac Gym, and that Isaac Lab is now the officially supported successor, I had a couple of related questions:

---

1. Feasibility of migrating H1_2 support to Isaac Lab

The IsaacLab repo currently supports Unitree H1, but does not include the H1_2. We're exploring whether it's possible to adapt this repo's existing H1_2 setup to Isaac Lab, which seems like it would involve:

• Rewriting the task configs using Python @configclass format (DirectRLEnvCfg)
• Converting the H1_2 asset as a USD (which we have done)
• Adapting rewards, observations, and resets to match Lab's simulation step flow
• Replacing any Gym-specific logic with Lab equivalents (e.g. VecTask → DirectRLEnv, use of Cloner, etc.)

Is this something you've looked into internally? Or are there known challenges that would make this migration difficult?

---

2. Would you recommend continuing with Isaac Gym for this use case?

Given that our aim is to build new gaits for the H1_2 with arm movement, is Isaac Gym still suitable for this kind of task, or would you advise making the jump to Isaac Lab now - despite the extra porting effort?

Open to either path, but want to align with the direction you're taking and what’s most maintainable and performant long term.

Thanks again. We'd appreciate any guidance you can share.

[daihuiao]

Hello, I'm also conducting the migration work of H1_2. I modified the H1rough environment in the original environment, re-matched the action space, and re-matched the joint names corresponding to the rewards, but some problems occurred: One is that after converting from an urdf file to usd, the model turned pure white. Is this the incorrect conversion method Furthermore, there was a problem with the collision sensor. isaaclab used the joint of the head and trunk to determine collisions, but after the model was replaced, the sensor of torso link didn't work Another issue is that there is a problem with the way the models work. It seems that they are colliding with each other Have you ever encountered similar problems? We would be extremely grateful for any help.