Digital twins and high-fidelity simulators are becoming central to the design, testing, and operation of modern robotic systems in industry, logistics, and human–robot interaction. However, existing comparative studies typically focus on a single robot type, task, or simulator, and rarely consider digital–twin–specific aspects, such as telemetry pipelines, operator interfaces, or cloud connectivity. This article proposes a unified evaluation framework for robotics-oriented digital twin platforms. It applies it across five representative use cases: (i) small UAV inspection, (ii) an industrial hand-robot cell, (iii) a warehouse logistics cell with autonomous mobile robots (AMRs), (iv) HCI/VR-based teleoperation of a mobile manipulator, and (v) a general indoor patrol and docking scenario.The framework standardises inputs (robot and environment profiles, task definitions, digital-twin data contracts) and outputs (physics/task metrics, performance metrics, and DT metrics) and explicitly encodes task complexity via structured phases and control steps per episode. It is implemented on several widely used robotics simulators (Gazebo/gz-sim, Webots, NVIDIA Isaac Sim, CoppeliaSim) and on an Unreal Engine 5 (UE5)–based stack that combines Datasmith asset import, Chaos Physics, Blueprints/Control Rig, and UMG/VR interfaces.As a detailed case study, we instantiate the warehouse logistics use case as a UE5-based digital twin and report quantitative results on order throughput, path lengths, real-time factor, frame times, telemetry bandwidth, and UI-to-actuation latency, along with observed bottlenecks such as Blueprint CPU overhead and visual-load–induced slowdowns. Across all platforms and use cases, the results highlight trade-offs between physical fidelity, scalability, interaction richness, and DT responsiveness, and lead to practical guidelines for selecting and composing simulation and digital-twin stacks for UAV, manipulation, logistics, and HCI-centric applications. The paper concludes with a discussion of limitations, including the simplification of physics and hardware dependence, and outlines future steps toward sim-to-real correlation and automated benchmarking.
Volodymyr et al. (Mon,) studied this question.