arXiv:2605.28548v1 Announce Type: new 
Abstract: Embodied Vision-Language Models (VLMs) have demonstrated impressive performance and generalization in robotics, particularly within Vision-Language-Action frameworks. However, a significant gap remains between the high-level semantic focus of standard text-guided pre-training paradigms and the low-level spatial and physical knowledge critical for execution in embodied environments. In this paper, we introduce GEM, a Generative-supervised Embodied vision-language Model designed to bridge this divide. We propose integrating a depth map generation task directly into the VLM pre-training phase. By training this generative objective jointly with the main model, we observe substantial improvements in embodied intelligence, significantly enhancing both semantic understanding and physical operation capabilities. To support this paradigm, we curate and release GEM-4M, a comprehensive large-scale dataset featuring a mixture of grounding, reasoning, and planning data paired with high-quality depth supervision. Extensive experiments demonstrate that GEM achieves state-of-the-art results across diverse embodied benchmarks. Furthermore, our deployed action model, GEM-VLA, exhibits vastly superior task execution abilities in both simulation environments and real-world evaluations. Code, models, and datasets are available at https://zhaorw02.github.io/GEM/

تهدف مقدمة نموذج GEM، وهو نموذج لغة-رؤية متجسد تحت إشراف توليدي، إلى تحسين الأداء والتعميم في الروبوتات من خلال دمج مهمة توليد خريطة العمق في مرحلة ما قبل تدريب نموذج اللغة-الرؤية. تتناول هذه الطريقة الفجوة القائمة بين الفهم الدلالي على مستوى عالٍ والمعرفة المكانية والفيزيائية اللازمة للتنفيذ الفعال في البيئات المتجسدة.

La introducción de GEM, un modelo de lenguaje-visión encarnado supervisado por generación, tiene como objetivo mejorar el rendimiento y la generalización de la robótica al integrar la generación de mapas de profundidad en la fase de preentrenamiento del VLM. Este enfoque aborda la brecha existente entre la comprensión semántica de alto nivel y el conocimiento espacial de bajo nivel necesario para una operación efectiva en entornos encarnados.

L'introduction de GEM, un modèle de langage-vision incarné supervisé par génération, vise à améliorer la performance et la généralisation de la robotique en intégrant la génération de cartes de profondeur dans la phase de pré-entraînement du VLM. Cette approche répond à l'écart existant entre la compréhension sémantique de haut niveau et les connaissances spatiales de bas niveau nécessaires pour une opération efficace dans des environnements incarnés.

The introduction of GEM, a Generative-supervised Embodied Vision-Language Model (VLM), aims to enhance the performance and generalization of robotics by integrating depth map generation into the VLM pre-training phase. This approach addresses the existing gap between high-level semantic understanding and low-level spatial knowledge necessary for effective operation in embodied environments.