Spatial reasoning is the step that turns perceived structure into decisions: given detections, depths, poses, and a task, it works out the relations, the causes, and the actions that follow, the why and how of spatial interaction rather than merely the what and where. It is where perception becomes goal-directed, supporting inference, planning, and problem-solving in physical and virtual environments.
The chapter is organized along the spectrum by which spatial reasoning has actually developed, from explicit symbolic models to learned and foundation-model-driven ones. At one end, reasoning runs over hand-built representations of geometry, physics, and logic, whose structure is inspectable and whose conclusions can be certified. In the middle, it emerges from data, as neural models acquire spatial competence directly rather than being told the rules. More recently, large pre-trained models supply reasoning as a general capability to be elicited and steered. Hybrid approaches recombine these, pairing the semantic reach of foundation models with the precision and verifiability of symbolic and learned components; a closing survey of benchmarks then records how the field measures spatial reasoning as it shifts from passive observation toward active, situated interaction.