Identify what was controlled
Ask how the operator interacted with the system and what the recording stores. Human input device values, transformed robot commands, and measured robot state are distinct records. A teleoperation label alone does not reveal which of these is available. Document mappings between them if the buyer needs to understand how a demonstration became an executable robot action.
Describe controller context
Record relevant control modes, constraints, and changes during the session. Determine whether assistance, smoothing, or other transformations affected the issued commands. Ask whether a field represents the operator’s request or the value actually sent to the robot. Preserve these distinctions in the dictionary so model targets are chosen with an understanding of the demonstration pipeline.
Review timing and pauses
Operators can pause, correct, or wait while the robot responds. Determine how those intervals appear in the data and whether idle steps were removed. Removing them can alter the relationship between observations and actions. Check that command and feedback records retain interpretable timing, especially around mode switches or periods when the operator is not actively controlling motion.
Keep operator variation interpretable
Working style, task familiarity, and collection instructions can influence demonstrations. Request only the operator metadata needed for the project, with appropriate handling. Group related sessions when designing evaluation. If the goal is to compare control conditions, make sure each condition is documented rather than treating all teleoperated episodes as one uniform source of behavior.
Review checklist
- Distinguish human input and robot commands
- Document control modes and transformations
- Retain pause and timing semantics
- Record relevant session grouping