1.2 · Model Families and Their Interfaces
VLM, VLA, world model, WAM, JEPA, and robot foundation model are not interchangeable labels. Each describes a different input-output contract, and those contracts determine what the rest of the system must provide.
- Distinguish the major Physical AI model families.
- Match each family to its inputs, outputs, and missing guarantees.
- Avoid category errors in architecture selection and evaluation.
VLM and VLA
A vision-language model produces language or semantic features from visual and textual context. A vision-language-action model extends the interface to robot actions or action distributions. Executable output still does not guarantee safe control.
World models, WAMs, and JEPA
A world model predicts possible future states, often conditioned on action. A world-action model represents futures and actions in one formulation. JEPA-style methods predict representations rather than requiring pixel-perfect reconstruction.
Foundation and infrastructure terms
A robot foundation model is defined by broad reuse across tasks or embodiments. A simulator generates synthetic interaction under a model of dynamics. A digital twin is an operational, synchronized model of a specific physical asset or process.
Choose architecture by interface and downstream decision—not by the newest label.
