SenseAgent Uses an LLM to Adapt IMU Sensing Across Domains

Summary

Deep learning models for inertial measurement unit (IMU) sensing often lose reliability when users, devices, or body positions change. The SenseAgent paper presents an LLM-guided agent that treats cross-domain activity recognition as a closed-loop sensing problem rather than a fixed inference pipeline. The LLM does not classify raw IMU signals directly; it plans over sensing tools, source-domain experience memory, online target memory, and verification modules. From an unlabeled target stream, the agent builds a diagnosis report and chooses whether to retain raw inference or invoke gravity-aware sensing, prototype transfer, or style normalization. Verifiers check source calibration, target-memory reliability, and no-harm criteria before high-risk adaptations are accepted. The system can also use scarce user feedback without retraining its backbone or abandoning the label-free route. Across multiple IMU datasets and deployment shifts, the verified route-selection process improves cross-domain sensing, with particularly strong benefits under harder body-placement and compound shifts. Limited user feedback provides additional gains.