Experimental evaluation demonstrates modular sensing and accurate narrative reporting in an indoor robotic platform, indicating viable functional decoupling for environmental monitoring.
Key Points
To design and implement a distributed, service-oriented architecture that decouples sensing, data storage, rule evaluation, and narrative report generation for robotic environmental monitoring.
Integrated a Raspberry Pi gateway, dedicated motor controller, five sensor modules, Node-RED middleware, a database, and deterministic alert rules with optional local large language models (LLMs).
Conducted two indoor monitoring campaigns, including a residential deployment recording 234 minutes of SCD41 CO2 data across 78 three-minute bins to evaluate a 15-minute persistence forecasting model.
Assessed report generation using 270 reports derived from nine deterministic synthetic scenarios across fixed templates and nine locally hosted LLMs.
The persistence forecast model achieved a 15-minute CO2 prediction mean absolute error of 58.3 ppm and a root mean square error of 78.8 ppm across observed concentrations of 679 to 1471 ppm.
All reporting systems preserved deterministic alert identifiers, with fixed templates and seven of nine local LLMs reaching 100% numerical fidelity across 270 synthetic scenario evaluations.
Qwen 3.5 9B was the only tested LLM that delivered all required measured content without claim-review flags while producing perfectly identical outputs across repeated scenario runs.