Mismatch of state feature.

[zhengsipeng]

Dear author, thanks for your great work!

I've noticing in your Libero2Lerobot processing, the state feature is "x,y,z,roll,pitch,gripper,gripper" with the last two dim as the state for gripper. However, I saw many other datasets use "pad, gripper" instead of "gripper, gripper". What's the difference between these two feature? By the way, I also notice you posted an issue on OpenVLA repo before. How shall I use LiberoState when trying to combine different robot control datasets together?

Thanks for your time.

[Tavish9]

Hi, thanks for digging into details.

Actually, it's "x, y, z, axis_angle1, axis_angle2, axis_angle3, gripper, gripper", I will update this notation soon, both in libero2lerobot and openx2lerobot.

We follow the default transform of OpenVLA, which uses 2D gripper as the last 2 dims.

For pad, gripper, please check issue you mentioned carefully. Most use 2D gripper and notate it as pad, gripper due to the misleading config.

• Libero state encoding openvla/openvla#209

For generalist robot training, It's good to use pad, gripper. However, for libero benchmark finetuning, I recommend using gripper, gripper.