Categorize Multinet v0 OpenX datasets based on Action Spaces for OpenVLA

[pranavguru]

Categorize the OpenX datasets used for Multinet v0 for OpenVLA profiling (https://docs.google.com/spreadsheets/d/1pwYeYQLkcBWVHwgt0OtL0KlL7VlfZ7q6Y3xfobbOZ7U/edit?usp=sharing) into:

• Datasets with identical action spaces to the datasets OpenVLA was fine-tuned on
• Datasets with action spaces that are subsets of the action spaces of the datasets OpenVLA was fine-tuned on
• Datasets with action spaces that are completely different, but smaller than the action spaces of datasets OpenVLA was fine-tuned on
• Datasets with action spaces larger than the action spaces of datasets OpenVLA was fine-tuned on

[Locke0]

Tracking OpenVLA eval datasets here

Thoughts on this? @pranavguru Next Steps:

• [ ] Eval OpenVLA 8-bit on small datasets like
  • [ ] NYU ROT 4D
  • [ ] Tokyo PR2 Fridge Opening 7D