Empirical compatibility of MAP-based inpainting with broader foundation action models

Investigate whether the masked-action inpainting formulation is empirically compatible with WAM and flow-matching-based vision-language-action foundation models.

Background

The paper introduces the Masked Action Space (MAP) interface and an Inpainting Policy for reconstructing unspecified components of robot action trajectories from heterogeneous spatiotemporal references. Although the framework is described as model-agnostic, the experiments instantiate it with Diffusion Transformer and U-Net-based diffusion policies.

The authors explicitly leave empirical validation on WAM and flow-matching-based vision-language-action models unresolved. Establishing compatibility would test whether the proposed masked-action editing formulation transfers beyond the diffusion-policy backbones evaluated in the paper.