Assign a purpose to each stream
A wrist camera may observe local contact while an external view shows the broader workspace. State which views the target model will receive and what evidence each should provide. Additional views can create redundancy without resolving the important occlusion. Review complete task sequences rather than assuming a camera count describes useful coverage.
Keep camera identity stable
Record stream names, physical placement, image properties, and applicable setup identifiers. Ask how moved or replaced cameras are represented across sessions. A field named camera_1 may not mean the same physical viewpoint in every source. If a common naming scheme is created, retain the original identifiers and document the mapping so unusual examples remain traceable.
Document timing and gaps
Streams may differ in frame rate or contain missing segments. Explain how images are paired with each other and with robot records. Distinguish a deliberately absent view from a dropped or invalid frame. If missing views are filled during export, retain a mask or another explicit indication so repeated images are not silently treated as fresh observations.
Review transformations and deployment fit
Resizing, cropping, and view selection should be documented alongside any geometry requirements. Confirm that evaluation inputs match the intended deployment setting; a policy tested with every camera may answer a different question from a policy using only the wrist view. Send source requirements with named camera roles and timing expectations, allowing Fiund to assess particular collections against them.
Review checklist
- Define the role of each camera
- Preserve identity and setup versions
- Document pairing and missing frames
- Match evaluation views to deployment