Development of a Lightweight Detachable Balloon Robot with Inertial Movement Recognition for Older-Adult Assistance

Abstract Service robots for older adults must balance physical manageability with sensing and mobility. We present a detachable balloon robot that separates a lightweight inflatable companion unit from a Mecanum-wheel mobile base. The approximately 2 kg upper unit includes an embedded inertial sensor for movement recognition when the head-mounted camera is occluded. Seven participants performed five prescribed movements, producing 35 recordings. Six raw inertial and orientation channels sampled at approximately 10 Hz were transformed into 14 sequence features and segmented into 5 s windows with 50% overlap. Four pipelines—radial basis function support vector machine (RBF-SVM), convolutional neural network (CNN), long short-term memory network (LSTM), and CNN-LSTM—were evaluated using nested subject-independent leave-one-subject-out (LOSO) validation. RBF-SVM achieved the highest participant-wise window-level accuracy (0.842 ± 0.128) and macro-averaged F1 score (0.840 ± 0.125). CNN-LSTM was the strongest deep model (0.742 ± 0.140 accuracy; 0.717 ± 0.135 macro-averaged F1 score). These results favor compact statistical features for the present short, pre-segmented task but do not establish superiority for longer or continuous activities. Because CNN-LSTM preserves temporal order, it remains a motivated candidate for future evaluation on longer movement sequences and action transitions. The study contributes a modular service-robot architecture and a rigorous cross-subject comparison of conventional and deep temporal recognition pipelines.