Evaluation of alternative conditioning mechanisms for action inpainting

Investigate whether classifier-free guidance and ControlNet-style conditioning improve the masked-action inpainting implementation relative to the conditioning mechanism used in the Inpainting Policy.

Background

The Inpainting Policy conditions action denoising on visual observations, proprioceptive state, and a local MAP window, while a RePaint-style procedure provides an alternative inference-time mechanism for enforcing specified action tokens.

The paper states that the current implementation may not be optimal and identifies classifier-free guidance and ControlNet-style conditioning as alternatives that remain unexamined. Their effectiveness for improving control adherence or task completion is therefore unresolved.