A new preprint on arXiv takes aim at how multi-task robot manipulation policies are compared. The authors note that published systems differ in architecture, scale, and pretrained priors all at once, which means no single component can be credited for a given result. They argue that the visual representation is the dominant factor, and that fine-grained visual features are what separate strong policies from weak ones.
The paper does not offer a full empirical demonstration in the abstract, but it frames the argument as a corrective to the field's habit of comparing whole systems rather than isolating variables. Because this is a single preprint, there is no independent source to agree or disagree with its claims yet.
If the argument holds, it would shift attention toward visual encoder design and representation learning rather than policy architecture or model scale. The authors suggest that researchers should treat visual representations as a first-class object of study when building and evaluating robot manipulation policies.