A task instruction expresses a requested goal, while a robot episode records what happened. Keep those two pieces connected without treating the instruction as proof of success. For sourcing inquiries, include example language, expected action data, and the observable evidence needed to tell correct execution from a plausible but wrong interaction.
Define instruction scope
Explain whether text describes an entire episode, an intermediate subtask, or a command issued during execution. Identify object references and destination terms that must be grounded in the scene. If an instruction can refer to several similar objects, provide a disambiguation policy and examples of requests that should remain uncertain.
Preserve execution context
Link instructions to the relevant episode and time interval. Record changes, corrections, or operator interventions rather than keeping only the final rewritten description. A later summary may describe the outcome accurately but omit what the robot was originally asked to do. Keep retrospective descriptions distinguishable from model input instructions.
Specify robot observations and actions
Language alone does not supply a robot policy target. Name the embodiment, action representation, controller, and required sensor streams. If human task video is being considered as supplementary context, label it separately so a buyer can distinguish semantic demonstrations from robot-executable trajectory records.
Evaluate grounding and completion
Review wrong-object, wrong-destination, partial-completion, and unobservable-outcome cases. Decide what evidence determines success and how ambiguous instructions are handled. Separate language variation from physical task variation when planning evaluation, and keep related episodes grouped according to the intended generalization test. Preserve the original text alongside any normalized task label so later reviews can revisit the interpretation.
Include in your inquiry
- Instruction examples and temporal scope
- Object references and ambiguity rules
- Robot signals and action representation
- Completion evidence and evaluation design