LATIDIA · Robótica

Superando el atajo visión-acción: Entrenamiento de interfaz latente para modelos fundamentales de robótica generalizables

arXiv: 2609.12641v1Tipo de anuncio: nuevo Resumen: Los modelos de base de robot logran un sólido rendimiento en la distribución, pero a menudo se degradan bajo los cambios de distribución visual. Al aprender a generar acciones a partir de vis pre-entrenados

La noticia

arXiv:2609.12641v1 Announce Type: new Abstract: Robot foundation models achieve strong in-distribution performance but often degrade under visual distribution shifts. When learning to generate actions from pretrained visual representations, models may exploit task-irrelevant visual cues that correlate with demonstrated actions within the training distribution. Such vision-action shortcuts can undermine generalization when these correlations change under distribution shifts. Mitigating these shortcuts requires constraining how visual information is used for action generation while preserving task-relevant spatial information. We propose Latent Interface Training (LIT), a framework-agnostic two-stage strategy that first establishes a spatial-goal-conditioned action prior without images, then constrains visual conditioning through a pose-supervised latent interface. Stage 1 trains the action expert to generate