Datasets:

---

openvla

/

modified_libero_rlds

like 30

Follow

OpenVLA Collaboration 199

Could a description of the features please be provided in detail? The paper/website/hf dataset/github are also somewhat sparse in this regard.

To quote the dataset metadata and question it:

1. 'observation.state': Tensor(shape=(8,), dtype=float32, description=Robot EEF state (6D pose, 2D gripper).),
   What exactly is 6D pose? cartesian 3D position + euler angle? If so, which of the 27 euler angle representations? What are the units? position=m/cm/mm? Orientation=degrees/radian?
   What are the units of the 2D gripper? m/cm/mm?

2. 'action': Tensor(shape=(7,), dtype=float32, description=Robot EEF action.),
   What are the 7 values? 6D pose (i.e. xyz rpy) + 1D gripper? 7D pose (xyz qx qy qz qw) but no gripper? 7 joints? degrees/radians?

3. Are the actions absolute poses or delta to previous state?

It is of course possible to make some assumptions, but one wrong minor assumption can completely nuke performance. Therefore, I consider it prudent to ask.

I appreciate all help. Please and thank you!