From Artificial Intelligence to Embodied Intelligence: The Evolution from Abstract Problem-Solving to 3D Scene Understanding

Classical artificial intelligence has long centered on symbolic reasoning frameworks and statistical pattern recognition, with most prior studies confined to deterministic, closed settings that decouple cognitive computation from real-world physical interaction. This paper systematically explores the pivotal paradigm shift toward embodied intelligence, a framework where cognition is fundamentally grounded in physical or simulated three-dimensional spaces via tight, continuous perception-action coupling. It generalizes the core methodological innovations across data acquisition pipelines, 3D representation learning, and embodied learning algorithms, and validates relevant empirical findings through standard benchmark tasks including robot navigation, dexterous manipulation, and object-centric interaction tasks. The paper evaluates seminal enabling technologies such as RGB-D data sensing, learning-based 3D perception, and high-fidelity simulation platforms, along with matched data augmentation approaches. It also discusses emerging fundamental conflicts between end-to-end and modular architectures, and between geometric and semantic 3D representations. Finally, it outlines frontier research trends including multimodal foundation models, differentiable physics, and sim-to-real transfer techniques, while clearly defining key unresolved bottlenecks: sample efficiency deficits, causal reasoning limitations, and cross-scenario transfer validity gaps.