Self-Initialized Locomotion Mode Prediction With GPU-Free Terrain Reconstruction

Lower-limb wearable robots require accurate Locomotion Mode Prediction (LMP) to provide appropriate assistance across diverse terrains. Recent reconstruction-based LMP methods fuse high-dimensional multimodal sensor data to model the relationship between human motion and terrains, improving prediction accuracy and cross-terrain adaptability. However, they typically incur high computational cost, including GPU dependency, and often require initialization procedures involving wearer participation or professional supervision. This paper proposes a self-initialized, GPU-free LMP method to overcome these deployment constraints. Our method adopts a gravity-aligned world coordinate frame as a unified geometric reference: a self-initialization procedure first establishes this reference, upon which a progressive plane representation enables GPU-free terrain reconstruction. Together, these two components form a pipeline that achieves reconstruction-based LMP on an onboard CPU without manual intervention. Comprehensive experiments across various terrains and subjects evaluate the system in terms of LMP accuracy, initialization success rate, computational efficiency, and memory footprint. We also compare the proposed method with a lightweight end-to-end baseline to further examine the role of the GPU-free terrain reconstruction. The results show that the proposed method achieves prediction accuracy comparable to state-of-the-art methods, including GPU-dependent counterparts, while operating entirely on a CPU without manual intervention.