Three Pillars of Embodied AI: Data, Model, Embodiment

具身智能三要素（数据 / 模型 / 本体）Common

An industry shorthand for breaking embodied AI down into three parts: data, model, and embodiment.

This is a breakdown commonly used in industry reports and talks, not a formal framework from any single paper. Data refers to the demonstrations, simulations, and human videos used to train a robot; model refers to the policy that turns observations into actions, such as a VLA model; embodiment refers to the robot hardware itself — its joints, dexterous hands, and sensors. The three constrain each other: the embodiment determines what data can be collected and what actions are possible, the amount of data determines how well the model can learn, and the model's capability in turn determines how valuable the embodiment is. Companies also tend to position themselves along these same three lines, as embodiment makers, “brain” companies, or data service providers.

Also called: Data / Model / Embodiment Framework
Related: Embodiment · Vision-Language-Action Model · Data Scarcity · Robot Body Maker · Robot-Brain (Model-Only) Company · Embodied AI Data Service Provider
Sources: Open X-Embodiment: Robotic Learning Datasets and RT-X Models