Classifying terrain for quadruped robots based on acoustic features

Quadruped robots offer significant advantages in outdoor exploration missions due to their superior ability to adapt to variable terrains. Accurate terrain recognition enhances the robot’s motion control and safety. In recent years, acoustic signals have emerged as a powerful supplement for terrain recognition, owing to their robustness and fine-grained classification capabilities. This paper presents a terrain classification method for quadruped robots based on acoustic signatures, employing a ResNet-18 deep residual network to learn Mel spectrograms for efficient terrain classification and probability prediction. A directional microphone mounted on the robot captures sound signals generated during interaction with different terrains, and the Mel spectrograms are extracted as training data. The deep residual network is then trained on these spectrograms to achieve accurate terrain classification. To optimize real-time performance, this study also explores the impact of different audio segment lengths on classification accuracy, determining the optimal segment length to balance efficiency and response time. Experimental results demonstrate that the proposed method achieves high accuracy and real-time performance in classifying three types of terrain-hard surface, grass, and sand—effectively meeting the terrain perception needs of quadruped robots in complex outdoor environments.