Light-EmbodiedHAR: An Explainable IMU Sensor-Fusion Framework for Human Activity State Recognition in Assistive Robot Interaction

Abstract

Human activity state recognition is a basic perception capability for assistive robots that interact with older adults, rehabilitation users, and workers in shared environments. However, many embodied-intelligence studies depend on vision-heavy perception pipelines that are difficult to deploy in low-cost, privacy-sensitive, or resource-constrained scenarios. This paper proposes Light-EmbodiedHAR, an explainable inertial measurement unit (IMU) sensor-fusion framework for lightweight human activity recognition and policy triggering in assistive robot interaction. Rather than claiming a new classifier, the contribution is a deployment-oriented perception-to-policy design that connects feature-level IMU fusion, compact model selection, confidence-gated policy mapping, caregiver-oriented explanation, and conservative fallback actions. A reproducible offline validation was conducted on the Hugging Face UCI-HAR federated dataset, which contains 7352 training samples, 2947 test samples, 561 IMU-derived features, six activity classes, and subject-disjoint splits. Six lightweight classifiers were compared under the same training and evaluation protocol. The best model, Linear SVM, achieved 96.13% accuracy and 96.18% macro-F1 with a serialized model size of 40.46 KB and an average inference time of 0.0068 ms per sample. Logistic regression reached 95.55% accuracy with a smaller 27.39 KB model, indicating a favorable edge-deployment trade-off. The revised framework further specifies controller memory assumptions, the effect of standard scaling on Linear SVM decision boundaries, and the use of confidence margins as uncalibrated abstention signals rather than statistical probabilities. Results show that accurate, explainable, and low-compute activity-state recognition can provide an engineering route for embodied assistive systems without relying on large-scale visual sensing.