RxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination
RxBrain ties text planning to predicted visual states inside one embodied-model sequence.
The paper presents Hy-Embodied-RxBrain, a foundation model meant to represent plans through both language reasoning and visual imagination. Its planning sequence uses language for decomposition, constraints, order, and decisions, while generated visual states ground each step. The authors train it with an automatic pipeline that turns embodied videos into aligned text-visual planning supervision. They also introduce RxBrain-Bench and report promising extension to continuous robot action generation without large-scale action-data pretraining. HF Daily Papers' note
The paper presents Hy-Embodied-RxBrain, a foundation model meant to represent plans through both language reasoning and visual imagination. Its planning sequence uses language for decomposition, constraints, order, and decisions, while generated visual states ground each step. The authors train it with an automatic pipeline that turns embodied videos into aligned text-visual planning supervision. They also introduce RxBrain-Bench and report promising extension to continuous robot action generation without large-scale action-data pretraining. HF Daily Papers' note
score 4