Bio-inspired terrain perception for snake robots: spectral analysis of vertical undulation and sidewinding

While snake robots offer exceptional potential for exploration in confined and unstructured environments, their capacity for active adaptation is currently hindered by a lack of robust terrain classification. Inspired by the mechanosensory perception of biological snakes, this study proposes a proprioceptive terrain classification framework relying solely on a single uniaxial accelerometer. To capture the coupled robot-terrain interaction dynamics from the complex, non-stationary signals inherent in floating-base locomotion, we introduce the band-limited spectral magnitude feature, which quantifies vibratory energy within specific frequency bands. The framework was validated across four canonical terrains (concrete, gravel, grass, and sand) using two distinct locomotion modes with varying physical impact mechanisms: vertical undulation and sidewinding. However, because sidewinding generates dominant shear forces with diminished normal impacts, it introduces classification ambiguity on yielding, low-stiffness terrains. To resolve this, we implemented a hierarchical classification strategy that first categorizes bulk terrain stiffness before finely differentiating terrains within the low-stiffness group. This approach achieved classification accuracies exceeding 98.8% for vertical undulation and up to 92.5% for sidewinding. Notably, data-driven optimization reveals that the most effective frequency bands for terrain identification closely align with the peak sensitivity range of biological snake mechanoreceptors. By ensuring high reliability through a single scalar feature, this work establishes a practical, computationally efficient baseline for real-time, terrain-adaptive locomotion control in field applications.