One robot moves blocks quickly while another carefully lifts a delicate object. Which is better? A single success number cannot answer that when the tasks and testing conditions differ. Researchers proposed a framework that asks readers to look more closely at what each experiment actually measured.

The approach considers five dimensions of evaluation for models that control robot actions. Important questions include how convincingly a result was tested, what it costs to adapt the system to another task, and whether another team could repeat the experiment. Reproducibility means that others can obtain comparable results when the procedure and conditions are clear.

The authors point out that published studies do not always provide enough information for a fair comparison. If cost or reproducibility evidence is missing, a reviewer should not invent a number and combine everything into one score. Their framework can therefore decline to produce an overall ranking instead of declaring a winner without enough support.

This matters when watching robot demonstration videos too. How many attempts^[1] were made? Which ones failed? How much preparation was needed before filming? The framework does not make any robot smarter; it helps people judge claims about robots more carefully. The researchers call the evaluation approach RoboSeed.