Most vision-language-action (VLA) models for robotic manipulation operate reactively, directly regressing actions from current observations without explicitly modeling future states. A new arXiv preprint, ELAN4D, identifies this as a limitation and introduces a plug-and-play adaptation that adds embodiment-centric 4D supervision to VLA training.
The title and abstract describe the approach as 'embodiment-centric' and 'plug-and-play,' implying that it can be layered onto existing VLA models rather than requiring a new architecture. The goal is to push policies beyond reactive action regression, presumably by incorporating spatial-temporal structure related to the robot's own embodiment.
The abstract is truncated, so the paper's exact supervision mechanism and experimental outcomes are not detailed in the source. The preprint is listed as arXiv:2605.30484v2 with an updated announcement type.