We outline the core research pillars that constitute the structural backbone of this handbook. They traverse the full pipeline from spatial representation and perception to reasoning and generative capabilities, and ultimately extend into concrete real-world applications. Each pillar aligns closely with a corresponding Part of the handbook, serving as a thematic axis for both foundational methods and emerging challenges in spatial intelligence.
-
•
Representation (Part II: Spatial-aware Representation)
This pillar investigates how spatial information is encoded and represented across modalities and dimensions. We begin with single-modality representations, including 2D (e.g., , visual frames), 3D (e.g., , point clouds, meshes), and 4D (e.g., , dynamic scenes, spatio-temporal fields), as well as non-visual modalities such as language and audio. Particular attention is given to cross-modal alignment and fusion mechanisms that unify heterogeneous data sources. These representational strategies provide the necessary structural foundation for downstream reasoning, manipulation, and generation. -
•
Perception (Part III, Section 6: Spatial Perception)
This pillar focuses on how intelligent systems extract structured knowledge from spatial environments. Topics include object detection, semantic and instance segmentation, depth and surface normal estimation, visual odometry (VO), and ego-motion prediction. These perception modules translate raw sensory inputs into geometric or semantic abstractions, enabling agents to comprehend and navigate the physical world. -
•
Reasoning (Part III, Section 7: Spatial Reasoning)
Reasoning in spatial intelligence spans multiple paradigms: symbolic reasoning over geometric primitives or logical propositions; geometric reasoning over shapes, topologies, and configurations; neural reasoning using deep architectures; and LLM-driven spatial inference based on large-scale pretrained priors. These methods enable agents to interpret relations, deduce spatial constraints, and plan in dynamic settings. -
•
Generation (Part III, Section 8: Spatial Generation)
This pillar examines the generative side of spatial intelligence. It encompasses tasks such as 2D/3D scene synthesis, controllable image and video generation, trajectory and motion prediction, and dynamic world creation. These generative processes are often guided by spatial prompts, semantic layouts, or agent intentions, forming a bridge between structured knowledge and creative capability. -
•
Applications (Part IV: Applications)
Finally, the application pillar illustrates how the above methods power real-world systems. Domains include embodied agents and robotics (e.g., , navigation, grasping), autonomous driving and aerial agents, warehouse and wearable AI systems, virtual and augmented reality (XR), geospatial AI (e.g., , urban planning, satellite analysis), biomedical and scientific computing (e.g., , 3D medical reconstruction), and intelligent assistants. These chapters emphasize system-level integration and highlight how spatial AI capabilities translate into societal impact.