Questions about OmniAction timing/FPS and jaco_play_events audio paths

[StevenFryto]

Hi, thanks for releasing RoboOmni and OmniAction.

I have two questions while inspecting the dataset and trying to understand the training setup.

1. Timing / FPS consistency with the source OXE datasets

For subsets such as bridge_*, fractal_*, jaco_play_*, and ucsd_kitchen_*, does OmniAction keep the same frame rate / control frequency as the corresponding source OXE datasets, or were the trajectories temporally re-sampled during preprocessing?

I would especially like to confirm:

• whether OmniAction preserves the original timing of the source datasets
• whether any temporal downsampling or upsampling was applied
• whether the alignment between audio/context signals and robot trajectories assumes the original timing or a modified timing
• whether different subsets use different effective frame rates

This would be very helpful for interpreting action timing correctly and for reproducing training faithfully.

2. jaco_play_events audio path issue

While inspecting jaco_play_events, I found that the speech_conv field contains paths such as ./events/jaco_play_247_1.wav and ./events/jaco_play_247_2.wav.

However, I could not find the corresponding jaco_play_* audio files in the public dataset repository under speech/events, or elsewhere under speech.

At the same time, some metadata JSON files under speech/events appear to contain internal or original-generation speech paths, so I am not sure whether:

• the jaco_play_* audio files are missing from the public release
• the files were renamed or reorganized
• or there is an additional preprocessing step needed to map speech_conv to the actual audio files used in training

Could you clarify the intended audio directory structure for training on OmniAction, especially for jaco_play_events?

Thanks a lot.