Distinguish names from physical variation
A list of object categories does not describe size, shape, material, or how an object is used. Identify the properties relevant to the task without requesting unnecessary detail. Two differently named objects may create similar manipulation conditions, while two instances of the same category may require different grasps. Describe the interaction your model must handle.
Record object identity and setup
Use instance identifiers where repeated objects need to be tracked across episodes. Document the relevant initial placement and target arrangement. If objects are replaced, damaged, or altered during collection, record those changes when they affect interpretation. A stable category label should not conceal a changed physical setup that matters to the control or evaluation question.
Separate object and scene changes
Changing objects, layout, camera view, and control mode at the same time can make variation difficult to interpret. Decide whether the project needs controlled comparisons or naturalistic combinations, and request metadata accordingly. Neither design should be presented as the other. Review task coverage across object instances rather than assuming every object appears in every required action.
Define held-out objects carefully
Evaluation may hold out instances, categories, properties, or combinations. State which boundary is intended and retain grouping information needed to maintain it. Related recordings and transformed copies should follow the same rule. Include required object families, permitted substitutes, and evaluation conditions in the inquiry so available source coverage can be assessed without promising universal object generalization.
Review checklist
- Describe relevant physical properties
- Track instances and setup changes
- Review object-by-task coverage
- Define the held-out object boundary