Terrain Attribute Recognition System for CPG-Based Legged Robot

In this paper, we develop a terrain attribute recognition system for CPG-based legged robots. First, a low-cost sensing hardware device is designed to be integrated into the robot, including a tactile sensor array and RGB camera. Second, for the tactile modality, a novel terrain attribute recognition framework is proposed. A data generation strategy that adapts to the motion characteristics is presented, which transforms the original tactile signal into a structured representation, and extract meaningful features. Based on unsupervised and supervised machine learning classifiers, the recognition rates reach 94.0% and 95.5%, and the switching time is 1 to 3 steps. Third, for the recognition of terrain attributes in the visual modality, a lightweight real-time mobile attention coding network (MACNet) is proposed as an end-to-end model, which shows an exhibiting an accuracy of 88.5% on the improved GTOS mobile data set, 169FPS inference speed and 6.6 MB model parameter occupancy. Finally, these two methods are simultaneously applied to the AmphiHex-II robot for outdoor experiments. Experimental results show that each modality has its own advantages and disadvantages, and the complementary relationship between multiple modalities plays an irreplaceable role in a broader scene.