AgentSense: Virtual Sensor Data Generation Using LLM Agents in Simulated Home Environments

The introduction of AgentSense marks a significant advancement in the field of Human Activity Recognition (HAR) systems for smart homes, a domain often hindered by the scarcity of large, diverse labeled datasets. By leveraging embodied AI agents guided by Large Language Models (LLMs), AgentSense creates a virtual data generation pipeline where agents simulate daily routines in smart home environments. This innovative approach not only produces rich sensor data that mirrors real-world diversity but also preserves user privacy. The effectiveness of AgentSense is underscored by its evaluation on five real HAR datasets, where models pretrained on the generated data consistently outperform traditional baselines, particularly in low-resource settings. Furthermore, combining this synthetic data with a small amount of real data achieves performance comparable to training on full real-world datasets. This development showcases the potential of LLM-guided agents for scalable and cost-effective d…