Spatial Reasoning Benchmarks

These benchmarks measure whether AI systems can go beyond recognizing objects to reason about geometry, direction, topology, perspective, and how scenes change over time. They use controlled synthetic puzzles, annotated images and videos, maps, and simulated or real embodied tasks, often with hierarchical tests and executable checks for precise answers or actions. Current models can handle some local relations but often fail at multi-step composition, consistent spatial maps across viewpoints, metric judgments, and long-horizon navigation. Structured representations, tools, active exploration, and targeted training can help, but reliable grounding in visual evidence and fair, standardized evaluation remain open challenges.