Digital Twin

A virtual copy of a real object, robot, or factory that's kept in sync using real data.

A digital twin is a computer model that corresponds one-to-one to a physical product, system, or process. What sets it apart from an ordinary simulation is that it continuously receives data from the real object to keep its own state synchronized, and in the strict sense it can also feed back and influence the real system. NASA used ground-based simulators paired with the actual spacecraft for troubleshooting as far back as the Apollo era, and formally defined the concept in 2010; NVIDIA positions it as a flagship use case for Omniverse, applied to factory planning and robot testing. In embodied AI, the term is often used more loosely, to mean a simulated environment that faithfully reproduces a real scene, used to train and evaluate a policy in simulation before it ever touches a real robot. The closer the replica, the smaller the sim-to-real gap, but the higher the modeling cost; an approach that only aims for similarity rather than an exact match is called a “digital cousin.”

ExampleAccording to NVIDIA, Foxconn used a digital twin to optimize a production line's layout and equipment placement in a virtual factory before rolling the changes out to the physical one.

Related

Real-to-Sim · Digital Cousin · Sim-to-Real Gap (Reality Gap) · NVIDIA Isaac Sim · NVIDIA Omniverse · Simulation Fidelity

Sources

Wikipedia: Digital twin
NVIDIA Glossary: What Is a Digital Twin?
Automated Creation of Digital Cousins for Robust Policy Learning (ACDC, CoRL 2024, arXiv 2410.07408)