RxBrain: An Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination

The paper introduces Hy-Embodied-RxBrain, an embodied cognition foundation model that combines language reasoning and visual imagination in a single planning sequence. Language encodes task decomposition, planning primitives, constraints, temporal order, and decision logic, while visual imagination predicts intermediate and final physical states. The model uses a unified multimodal Mixture-of-Transformers architecture for language, image, and video understanding and generation. The authors also present an automatic pipeline that turns embodied videos into joint text-visual planning supervision, and introduce RxBrain-Bench. The abstract reports results in embodied understanding, generation, and continuous robot action generation, including promising real-robot performance without large-scale action-data pretraining.