RAG-HAR+ Boosts Cost-Efficient Human Activity Recognition on Edge

This paper introduces RAG-HAR+, an advanced framework for Human Activity Recognition (HAR) from wearable sensors, optimized for cost-efficient edge deployment. Building upon the existing RAG-HAR, which frames HAR as a training-free, retrieval-augmented task, RAG-HAR+ significantly strengthens its retrieval capabilities while minimizing reliance on Large Language Model (LLM) inference. Traditional deep learning HAR methods often demand extensive labeled data and retraining for new sensor settings or activity taxonomies. RAG-HAR+ addresses these challenges through a "retrieval-first" approach. It incorporates an offline Retrieval Designer Agent that intelligently selects dataset-specific feature groups from a diverse pool of motion descriptors. This allows for more precise comparison of sensor windows based on activity patterns. During inference, the system primarily uses majority voting from retrieved neighbors for samples with high confidence. Only cases deemed uncertain are then passed to an LLM-based Ambiguity Resolver Agent. This strategy enables RAG-HAR+ to achieve competitive or improved performance across six HAR benchmarks while substantially reducing LLM usage, token consumption, and inference time, demonstrating practical feasibility for mobile sensing.