SensingAgents: Multi-Agent LLM Framework for IMU Activity Recognition

A new multi-agent system called SensingAgents has been developed by researchers to enhance Human Activity Recognition (HAR) using Inertial Measurement Unit (IMU) sensors, utilizing Large Language Models (LLMs). Existing deep learning HAR models struggle with issues such as reliance on labeled data, ambiguity based on position, and insufficient transparency in reasoning. SensingAgents assigns distinct roles to LLM-driven agents: Analyst Agents for analyzing position-specific sensor data (arm, wrist, belt, pocket), Advocate Agents that address sensor discrepancies through both dynamic and static debates, and a Decision Agent that maintains reliability amid sensor drift or failure. Evaluated on the Shoaib dataset, this framework demonstrated notable improvements and addresses crucial challenges in IMU-based HAR, with potential uses in mobile health, smart environments, and human-computer interaction.