Jian Teng

Augmented reality (AR) learning cards present a promising medium for fostering spatial cognition in early STEM education, yet prevailing systems rely on conventional convolutional neural networks that discard critical spatial hierarchies during object recognition. This study proposes CapsAR‐STEM, a novel framework that integrates Capsule Networks (CapsNet) with marker‐based AR learning cards to scaffold elementary children's spatial reasoning and STEM vocabulary acquisition. The CapsNet module preserves part‐whole spatial relationships through dynamic routing between capsules, enabling robust recognition of card‐bound 3D objects across rotations, scales, and partial occlusions. A pedagogically sequenced AR interaction pipeline renders context‐sensitive 3D STEM models and delivers adaptive spatial vocabulary prompts aligned with each child's proficiency profile. Experimental evaluation with 36 parent–child pairs (children aged 6–9 years) demonstrated significant improvements in spatial vocabulary comprehension ( p < 0.01, Cohen's d = 0.91) and STEM concept retention ( p < 0.01) compared to a conventional AR flashcard baseline. The CapsNet recognition module achieved 92.4% classification accuracy across 68 card categories with notably superior robustness to viewpoint variation (88.7% under 45° rotation) relative to a MobileNetV2 baseline (79.3%). Usability assessments revealed high parental satisfaction ( M = 5.3/7.0) and sustained child engagement ( M = 4.6/5.0). Because the experimental condition differs from the baseline in both recognition architecture and adaptive scaffolding, the present study evaluates CapsAR‐STEM as an integrated system; the monotonic relationship between capsule‐relevant vocabulary dimensions and learning gains is interpreted as suggestive, rather than definitive, mechanistic evidence, and a recognition‐only ablation is identified as necessary future work. Results indicate that capsule‐based spatial hierarchy preservation, combined with pedagogically grounded AR interaction design, provides effective scaffolding for children's STEM spatial cognition development.

Skeleton‐based hand gesture recognition in augmented reality (AR) environments demands models that are both accurate and computationally efficient to support real‐time stereoscopic interaction. Existing graph convolutional network (GCN) approaches construct dynamic adjacency matrices primarily through simple pairwise subtraction between joints, which constrains their capacity to capture the nuanced spatiotemporal dependencies inherent in stereoscopic AR gesture data. Furthermore, many current architectures impose substantial parameter overhead, thereby limiting deployment on resource‐constrained AR devices. To address these challenges, this paper proposes a Mamba‐Enhanced Graph Convolutional Network (ME‐GCN), which integrates the selective state space model (SSM) Mamba into a lightweight graph convolutional framework for stereoscopic hand gesture recognition. Specifically, we design a Spatial Graph Mamba Module (SGMM) that leverages the linear‐complexity temporal scanning capability of Mamba to model long‐range inter‐frame dependencies, and a Stereoscopic Depth‐Aware Topology Refinement (SDTR) mechanism that dynamically adjusts the graph topology using binocular depth cues. A Channel‐Temporal Dual‐Stream Fusion (CTDF) strategy further integrates spatial, temporal, and depth features into a unified representation. Extensive experiments conducted on the SHREC'17 Track, DHG‐14/28, and a custom AR‐Stereo‐Gesture dataset demonstrate that ME‐GCN achieves competitive recognition accuracy while maintaining significantly fewer parameters and lower inference latency compared to state‐of‐the‐art methods. Specifically, ME‐GCN attains recognition accuracies of 97.38% and 94.16% on the SHREC'17 Track 14‐class and 28‐class tasks, and 94.28% and 92.14% on DHG‐14/28, while requiring only 0.68 M parameters and 1.08 GFLOPs—a 48.5% parameter reduction relative to TD‐GCN—at an average inference latency of 11.52 ms, the lowest among all compared methods. Future research will extend ME‐GCN to continuous gesture recognition in streaming AR scenarios, multi‐modal RGB‐depth fusion, and lightweight model‐compression techniques for real‐time on‐device AR deployment.