Inventory the original streams
Cameras, robot state, actions, and other sensors can have different recording rates and timing behavior. Ask for each stream’s timestamps and any available information about dropped records. A delivered table with one row per step may already represent a resampled combination. Its apparent regularity does not establish that every underlying signal was measured at that interval.
Explain the resampling rule
Nearest-neighbor selection, interpolation, and repeated values have different meanings. Record which operation applies to each field and whether it uses future information. A method suitable for a slowly changing state may be inappropriate for a discrete gripper command or event flag. Preserve enough source timing to inspect how the derived step sequence was constructed.
Make missingness explicit
Distinguish no command, unavailable measurement, invalid sensor reading, padding, and real zero. Ask for masks or status fields where appropriate. Repeated images can represent a stationary scene or a missing camera update; the pixels alone may not explain which. Report gaps in relation to task events so reviewers can determine whether essential behavior remains observable.
Review the loader as well as files
A valid export can become misleading if the loading process fills gaps or drops steps silently. Inspect representative episodes before and after loading, including examples with known missing records. Agree on exclusion or retention rules according to the model task. Capture rate labels and file validity are useful inputs to review, but they do not replace a clear account of stream association.
Review checklist
- Identify original stream timestamps
- Document field-specific resampling
- Use explicit missing-data indicators
- Inspect known gaps through the loader