Abstract: This paper presents a ROSGPT-based semantic navigation system for multi-user autonomous mobile robot (AMR) operation under ROS 2. An environment-specific prompt converts Traditional Chinese, English, and mixed-language requests into structured MOVE, ROTATE, and GOTO commands. JSON-format constraints, station-name whitelisting, and physical-parameter bounds are applied before command dispatch. A middleware mechanism comprising a first-in, first-out (FIFO) Task Queue, a Worker Thread, and a Dispatcher decouples request admission from cloud-based LLM inference and maintains enqueue-order task processing. Validated GOTO commands are dispatched to the Nav2 Action Server, whereas MOVE and ROTATE commands are executed by an odometry-feedback closed-loop controller. The system was evaluated through three experiments covering semantic parsing, concurrent middleware processing, and continuous physical navigation. For a predefined dataset of 150 commands, the Improved Prompt increased the Exact Match Rate from 7.3% to 100%. In a five-client workload comprising 60 requests, all requests completed semantic parsing and structured-command dispatch without observed middleware-level loss. In addition, 39 of 40 continuous physical navigation tasks were completed successfully, corresponding to a task completion rate of 97.5%. These results provide preliminary evidence of the feasibility of the proposed architecture for multi-user semantic navigation in indoor environments.

Keywords: Robot Operating System 2 (ROS 2), Autonomous Mobile Robot (AMR), ROSGPT, Large Language Model (LLM), Prompt Engineering, Semantic Navigation, Multi-User Task Scheduling

Pages: 406-418