Terrain classification method for quadruped robots based on proprioception sensors

Acquiring terrain information during robot locomotion is pivotal for autonomous navigation, gait selection, and trajectory planning. Quadruped robots, owing to their biomimetic structures, demonstrate enhanced traversability over intricate terrains compared to other robotic platforms. Furthermore, the internal sensors of quadruped robots encompass rich terrain-related motion state data during locomotion across diverse terrains. This study investigates terrain perception in quadruped robots based on internal sensor data, probing the temporal signal relationships(including IMU data, body velocity data, foot-relative veloctiy data, and overall body velocity)influenced by motion state variations as quadruped robots traverse different terrains. By examining the interplay between terrain conditions, motion states, signal attributes, and locomotion patterns, we elucidate the mapping correlation between distinct terrains and internal sensor signal characteristics. This mapping correlation is acquired through the development of a deep learning model utilizing a self-attention mechanism. Employing a multi-label classification algorithm, we classify complex terrains and establish multiple physical feature labels—uneven, slippery, soft, and gradient—to depict terrain attributes. A feature-labeled terrain dataset is established by abstracting diverse terrain features across various terrains. Unlike semantic labels (e.g., grassland, sand, gravel) comprehensible only to humans, feature labels provide a more granular and precise terrain characterization, encompassing broader terrain attributes. Furthermore, a novel multi-sensor fusion model is proposed, utilizing waveform diagram as input, achieving an average classification accuracy of 98.49% across diverse terrains.