This feature would be useful for those who want to perform training using a collection of datasets.

Currently, you need to load each dataset and check the .keys() of a single item to access this information.

For example, the info file currently only contains:

{
    "codebase_version": "v1.5",
    "fps": 30,
    "video": true
}

Reference: Koch Pick Place 1 Lego Dataset Info

In this file, it would be helpful to also include the names and shapes of the available inputs.