Embodied intelligence systems must integrate diverse sensory inputs—vision, language, audio, touch, 3D signals, and action trajectories—to perceive and act in the physical world. This special issue focuses on multimodal learning approaches that enable embodied agents to build effective world models from these heterogeneous inputs. The work addresses unique challenges in embodied settings, including spatiotemporal understanding, cross-modal fusion, causal reasoning, and handling real-world conditions like noise, missing data, and partial observability.
The special issue invites research on multimodal representation learning, information fusion, and reasoning methods for embodied systems. Studies should explore how embodied agents leverage foundation models and integrate multiple modalities for tasks ranging from perception and interaction to planning, decision-making, and world model generation.