Open Access

ARTICLE

A 5G-MEC-Enabled, Digital-Twin-Trained Framework for Autonomous Mobile Robots on the ROSMASTER R2 Platform

Daniel Šolc^1,*, René Ivančák¹, Juraj Gazda¹, Eva Chovancová¹, Eugen Šlapák¹, Gabriel Bugár²

1 Department of Computers and Informatics, Faculty of Electrical Engineering and Informatics, Technical University of Košice, Košice, Slovakia
2 Department of Computer Networks, Faculty of Electrical Engineering and Informatics, Technical University of Košice, Košice, Slovakia

* Corresponding Author: Daniel Šolc. Email:

Computers, Materials & Continua 2026, 89(2), 13 https://doi.org/10.32604/cmc.2026.088224

Received 30 June 2026; Accepted 10 August 2026; Issue published 15 September 2026

Abstract

Autonomous mobile robots increasingly rely on three tightly coupled capabilities: low-latency wireless connectivity, edge-side compute acceleration, and simulation-based pre-training of perception and control models. Each has been studied extensively in isolation, but their joint deployment on a single platform remains rare. This paper presents an integrated framework combining a private 5G Stand-Alone (5G SA) access network, a Multi-Access Edge Computing (MEC) layer with adaptive offloading, and a digital-twin training pipeline in NVIDIA Omniverse Isaac Sim. It is realised on the Yahboom ROSMASTER R2 with an NVIDIA Jetson Orin NX, a Quectel RM530N-GL 5G modem in band n78, and a Raspberry Pi 5 load-balancing orchestrator. Perception models are trained entirely in simulation: a one-dimensional residual network with Squeeze-and-Excitation blocks (ResNet-1D-SE) for Light Detection and Ranging (LiDAR) localization, and a Soft Actor-Critic (SAC) agent for vision-guided lane following. At deployment time, a utility-based node selector on the Raspberry Pi 5 selects in real time which heterogeneous MEC node runs the You Only Look Once version 26 (YOLO26) detector from central processing unit (CPU), random-access memory (RAM), graphics processing unit (GPU) and round-trip-time (RTT) telemetry, and the robot streams directly to it. We evaluate end-to-end latency and stability under one-to-four concurrent streams, adaptive MEC switching, onboard-vs.-offloaded energy, and sim-to-real localization accuracy. The private 5G SA link delivers seven times lower per-packet latency jitter than Wi-Fi (2.1 vs. 14.8 ms standard deviation) at the cost of a 5 ms higher single-stream end-to-end latency (67 vs. 62 ms); this overhead is already erased at three concurrent streams and becomes a 14 ms advantage for 5G SA at four. Offloading YOLO to a GPU MEC node cuts the robot’s compute-related power draw by 5 W (40% of the compute power budget, propulsion excluded); and the sim-trained localizer transfers to the physical robot with a 7.8 cm mean error and no real-world fine-tuning. The framework provides a reproducible reference design and a quantitative deployment guideline for connected autonomous robots.

---

Keywords

---