Multi-Agent 5G Architecture for Task-Level Robotic Control in Simulation

Abstract

This paper presents and evaluates a multi-agent architecture for natural-language control of mobile robots in simulation. The system targets an autonomous mobile robot operating in an industrial-like scenario, where high-level textual commands are provided by a human operator. An emulated fifth-generation non-public network (5G NPN) transport segment connects the robot and edge domains, while a multi-access edge computing (MEC) orchestrator interprets each command through a pluggable large language model (LLM)-based semantic parser. The resulting target pose is then forwarded to the robot through a gRPC-based agent exposing Robot Operating System 2 (ROS2) Navigation 2 (Nav2) capabilities. The architecture includes timing instrumentation that decomposes the end-to-end (E2E) control loop into semantic parsing, emulated 5G transport, middleware, and navigation execution components. Experiments were conducted in Gazebo with a TurtleBot, using 20 trials per model and independent network measurements. A cloud-supported model was compared with locally deployed alternatives under identical robotic and network conditions. Results show that the emulated 5G segment contributes only a few milliseconds to the total latency, whereas the semantic parsing stage determines the main differences in responsiveness. These preliminary findings support the feasibility of combining 5G-oriented edge communication, MEC, and LLM-based multi-agent orchestration for intuitive interaction with mobile robots, while highlighting the trade-off between deployment autonomy and E2E latency.