Skip to content

Question about action-video temporal alignment in RoboFactory dataset #6

Description

@CarlaFinocchiaro

Hi, thank you for releasing MultiWorld and the RoboFactory dataset!

I am currently working with the TwoRobotsStackCube data and I have a question about the temporal alignment between the RGB videos and the action arrays.

For example, for one episode I found:

-episodeXXXXXX_action.npy: shape (2, T, 8)
-episodeXXXXXX_head_camera_global_video.mp4: T frames

Could you please clarify the transition convention used when exporting the dataset?

Specifically, does:

action[:, t, :]

represent the action applied from video_frame[t] to obtain video_frame[t+1], i.e.

observation[t] + action[t] → observation[t+1]?

Or is the action aligned with the observation resulting from that action, meaning that a one-frame shift is required for action-conditioned prediction?

Also, were the RGB frames or actions temporally resampled/interpolated during the RoboFactory → MultiWorld dataset export?

Thank you very much!

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions