Quantify the dependence of visual OPE accuracy on dataset coverage

Quantify how RoboFQE’s value-estimation accuracy depends on the offline dataset’s coverage of state-action regions visited by the target policy under visual observations.

Background

RoboFQE estimates policy returns from a fixed offline dataset, so its accuracy depends on whether that dataset contains sufficient coverage of the state-action regions encountered by the evaluated policy. The paper’s experiments demonstrate that limited coverage and out-of-support actions can produce extrapolation errors, but they do not provide a systematic characterization of this dependence for visual robot manipulation.

The unresolved problem is therefore to quantify the relationship between offline state-action coverage and value-estimation accuracy in visual observations, which would help establish reliability criteria for rollout-free evaluation and clarify when RoboFQE can be trusted.