Make the dictionary explicit
Record units for positions, angles, velocities, timing, forces, and other requested measurements. Include component order and sign conventions. A plausible numerical range does not establish a unit. Ask whether fields are normalized, clipped, or stored in source units and whether values have already been transformed before delivery. Keep those meanings accessible without requiring a specific training codebase. Provide example values with explanations for fields whose conventions are easy to misread.
Describe orientation and gripper conventions
Orientation fields need a stated representation and component convention; gripper values need an explanation of what opening, closing, and intermediate values mean. Some sources record targets while others record measured state. Treat each field according to its documented semantics rather than assuming a familiar name implies a common convention across collections.
Define unavailable and invalid values
Explain whether zero is a real observation, a default, or a missing-value substitute. Request masks or status fields when records can be absent or invalid. Padding introduced during batching should be distinguishable from actual recorded steps. Without those distinctions, a loader can turn an incomplete sensor record into a misleading example of stationary behavior.
Review the full processing path
Conversions can occur during export, loading, and model preprocessing. Track where each one happens so values are not converted twice or interpreted in the wrong scale. Inspect representative fields before and after the proposed pipeline. Share the required output conventions in the sourcing brief and retain source definitions to make disagreements traceable.
Review checklist
- State units and component ordering
- Define orientation and gripper meaning
- Separate invalid data from real zeros
- Track conversions across the pipeline