arXiv:2605.29577v1 Announce Type: new 
Abstract: Vision-Language-Action (VLA) models have emerged as a promising framework that unifies perception, reasoning, and control for robot manipulation by adapting pretrained vision-language models (VLMs) to action prediction. However, VLM-derived representations are often insensitive to subtle visual distinctions required for low-level control, causing state aliasing between visually similar states that require substantially different actions. Prior VLA studies improve visual understanding by generating visual or reasoning outputs, such as future frames, 2D grounding points or traces, or intermediate spatial reasoning steps, but these objectives typically shape the vision encoder only indirectly through end-to-end prediction and do not explicitly analyze state aliasing in the learned visual feature space. To mitigate state aliasing, we introduce inverse dynamics learning as an auxiliary objective that directly supervises the VLA vision encoder. By predicting the action between current and future observations, our objective encourages the encoder to capture fine-grained visual distinctions that determine low-level actions. We further use pseudo-reversed supervision to expose the encoder to a broader range of action directions and improve generalization under limited robot demonstrations. Our method applies to diverse VLA baselines, uses only standard observation-action pairs without additional annotations, and preserves the original inference pipeline at test time. Experiments on CALVIN ABC-D and SimplerEnv show consistent gains across diverse VLA baselines. Frozen-encoder probing and state-feature alignment analyses further show that our method learns state-discriminative visual representations that reduce state aliasing and better align with robot state changes.

أدت التطورات الأخيرة في نماذج العمل-اللغة-الرؤية (VLA) إلى إدخال التعلم الديناميكي العكسي كطريقة للتخفيف من مشكلة تداخل الحالة، والتي تحدث عندما تتطلب الحالات المرئية المتشابهة إجراءات مختلفة. يهدف هذا النهج إلى تحسين أداء نماذج VLA من خلال الإشراف المباشر على مشفر الرؤية، مما يعالج قيدًا حرجًا في النماذج الحالية التي غالبًا ما تسيء تفسير الفروق البصرية.

Los avances recientes en los modelos de Acción-Lenguaje-Visión (VLA) han introducido el aprendizaje de dinámicas inversas como un método para mitigar el aliasing de estado, que ocurre cuando estados visualmente similares requieren acciones diferentes. Este enfoque tiene como objetivo mejorar el rendimiento de los modelos VLA al supervisar directamente el codificador visual, abordando una limitación crítica en los modelos actuales que a menudo malinterpretan las distinciones visuales.

Des avancées récentes dans les modèles d'action langage-vision (VLA) ont introduit l'apprentissage des dynamiques inverses comme méthode pour atténuer l'aliasing d'état, qui se produit lorsque des états visuellement similaires nécessitent des actions différentes. Cette approche vise à améliorer la performance des modèles VLA en supervisant directement l'encodeur visuel, abordant une limitation critique des modèles actuels qui interprètent souvent mal les distinctions visuelles.

Recent advancements in Vision-Language-Action (VLA) models have introduced inverse dynamics learning as a method to mitigate state aliasing, which occurs when visually similar states require different actions. This approach aims to enhance the performance of VLA models by directly supervising the vision encoder, addressing a critical limitation in current models that often misinterpret visual distinctions.