Scalable Edge–Cloud Cooperative Digital Twin Control for Multi-Robot Systems

While edge-enabled Digital Twin networked control systems (E-DTNCSs) perform perception at the network edge and transmit compact semantic states instead of raw sensor streams, prior validations have been confined to small fleets. In this paper, we investigate whether these advantages persist as the fleet grows, investigating this issue in terms of perception complexity, communication load, and closed-loop responsiveness. We redesign the Smart Observer so that its per-frame cost is dominated by resolution-dependent operations, with only lightweight per-robot decoding growing with the number of robots. A dual-channel architecture separates fleet-state observation from robot-specific control: the observation path keeps a single periodic source whose payload grows by approximately 36 bytes per robot per update, and robots can join or leave a running Digital Twin (DT) session without reinitializing perception. Physical experiments with two to eight robots show that Smart Observer processing increases from 30.61 to 34.28 ms and DT synchronization latency increases from 34.00 to 37.65 ms, while the compression ratio relative to per-robot raw-image uplinks increases from 811:1 to 1298:1. The system remains collision-free through eight robots at a conservative safety radius, achieves higher task throughput than a raw-image DTNCS that becomes inoperable beyond five robots, and handles repeated join, leave, and rejoin events without interrupting observation. Calibrated component-wise scaling models give a nominal decision-availability latency of approximately 95 ms at N=100, an analytical stress point rather than a physically validated fleet size.