arXiv:2605.29416v1 Announce Type: cross 
Abstract: Vision-Language-Action models have achieved remarkable progress in robotic manipulation, yet they suffer from a critical limitation: a lack of 3D scene understanding. This deficiency manifests as three intertwined challenges: weak extraction of 3D spatial positions without enforcing multi-view consistency, inadequate 3D instance understanding, and fragile reasoning under occlusion. Although mature 3D perception methods exist, their direct integration into VLA pipelines is hindered by architectural incompatibility and by heavy reliance on costly instance-level annotations. To address the above challenges, we propose 3DVLA, a plug-and-play framework that injects robust 3D reasoning into pretrained VLAs without requiring extra manual labels or discarding VLM priors. Specifically, 3DVLA tackles the three challenges through: (1) pervasive 3D feature encoding with explicit multi-view consistency constraints across all modalities and a Spatially-Conditioned Geometry Aggregation method, (2) an instance estimation module with high-level instance tokens for 3D instance awareness, and (3) a masked self-supervised 3D encoding branch that retains its predictor for visual token completion to handle occlusions. We integrate 3DVLA with multiple VLA baselines and evaluate on LIBERO-Plus and RoboTwin 2.0. Results show consistent and significant gains in manipulation performance, validating both the effectiveness and plug-and-play compatibility of our approach.

يمثل إدخال 3DVLA تقدمًا كبيرًا في نماذج اللغة-الرؤية-العمل (VLA) من خلال تعزيز الفهم المكاني والنسخي ثلاثي الأبعاد، مما يعالج القيود الحرجة في المناورة الروبوتية. يدمج هذا الإطار تفكيرًا ثلاثي الأبعاد قويًا في نماذج VLA المدربة مسبقًا دون الحاجة إلى تسميات يدوية إضافية، مما يحسن استخراج المواقع الثلاثية الأبعاد والقدرة على التفكير تحت التداخل.

La introducción de 3DVLA marca un avance significativo en los modelos de Acción-Lenguaje-Visión (VLA) al mejorar la comprensión espacial e instancial en 3D, abordando limitaciones críticas en la manipulación robótica. Este marco integra un razonamiento 3D robusto en modelos VLA preentrenados sin necesidad de anotaciones manuales adicionales, mejorando así la extracción de posiciones espaciales 3D y el razonamiento en condiciones de oclusión.

L'introduction de 3DVLA représente une avancée significative dans les modèles Vision-Language-Action (VLA) en améliorant la compréhension spatiale et d'instance en 3D, répondant à des limitations critiques dans la manipulation robotique. Ce cadre intègre un raisonnement 3D robuste dans des modèles VLA préentraînés sans nécessiter d'annotations manuelles supplémentaires, améliorant ainsi l'extraction des positions spatiales 3D et le raisonnement en cas d'occlusion.

The introduction of 3DVLA marks a significant advancement in Vision-Language-Action (VLA) models by enhancing 3D spatial and instance understanding, addressing critical limitations in robotic manipulation. This framework integrates robust 3D reasoning into pretrained VLA models without the need for additional manual annotations, thereby improving the extraction of 3D spatial positions and reasoning under occlusion.