GPU training: how to “lock” an end-effector (Surface-Gripper-like) for crawling?

[dadihrannar1]

Hi all,

I’m working on a research project where a robot crawls using 2–3 Franka Panda arms as legs. To make this work, each “foot” (the end-effector) needs to stick to the ground during stance and then release during swing, very similar to the Surface Gripper behavior.
In a previous project I used the Surface Gripper successfully, but I had to train on CPU. For this project I’d like to train on GPU. From what I can tell, the Surface Gripper is still CPU-only in Isaac Sim 5.0.

Questions:
Is there a way to get a Surface-Gripper-style attach/detach working when training on GPU?
If not, do you recommend any simple alternatives on GPU for temporary “lock/release” at the end effector (e.g., a fixed joint toggle, a constraint trick, or a material/friction approach that’s reliable for locomotion)?
If the current implementation can’t run on GPU, would it be feasible for me to build a GPU-compatible variant (an extension) that mimics the same attach/detach behavior? Any pointers on the right APIs or examples to follow would be super helpful.

Environment
Isaac Sim 5.0 (Ubuntu 22.04)
Isaac Lab (RL training)
Goal: run many parallel environments on GPU

My focus is the locomotion policy rather than detailed mechanical grasping, so a practical, GPU-friendly way to “lock” the end effector would help a lot. Thanks in advance for any guidance or examples!