AutoResearcHAR: A Multi-Agent Framework for Dataset Adaptive IMU Based Human Activity Recognition

Building robust IMU-based human activity recognition (HAR) systems requires tuning the full sensing pipeline, not only the model. We present AutoResearcHAR, a multi-agent framework for dataset-adaptive HAR optimization. A Data Agent refines source selection, preprocessing, windowing, and sampling; a Model Agent improves architecture and training choices; and an Augmentation Agent designs IMU augmentation and synthetic-data strategies. An Orchestrator composes compatible proposals, while an LLM-based Judge filters candidate edits before empirical evaluation. The system performs iterative hill climbing over bounded code and configuration level changes, retaining candidates that improve Macro-F1. We evaluate AutoResearcHAR on 10 IMU-HAR datasets using leave-one-subject-out evaluation and 3 held out datasets using leave-one-dataset-out evaluation. AutoResearcHAR improves over the best non-agent baselines by 1.2 to 12.9 Macro-F1 points in LOSO and 2.3 to 8.9 points in LODO, showing the value of coordinated multi-agent pipeline search for wearable sensing.