A live connection to one specific machine
A simulation, of the kind used for sim-to-real training, is built once and then run independently. It models a class of robot or a class of environment, and once training is done it has no ongoing relationship to any particular physical unit rolling off the line. You could delete the physical robot entirely and the simulation would keep running unaffected, because it was never tethered to that one machine's actual state. A digital twin inverts this relationship entirely. It exists to represent a specific, individual, serial-numbered physical asset, and it only remains meaningful as long as it stays synchronized with that asset's real, current condition.
That synchronization is the defining feature, not the modeling technique. A twin ingests a continuous stream of live data (joint positions, motor currents, temperatures, vibration, cycle counts) from its physical counterpart and updates its internal state to match. If robot unit #14 develops a slight misalignment in its second joint, unit #14's twin should reflect that misalignment, while a generic simulation of 'that model of robot' would have no way to know or care. The twin is disposable in a different sense too: when the physical robot is decommissioned, its twin's purpose ends with it, because the twin was never a general-purpose model, it was a specific one's shadow.