Terrain-based Place Recognition in Quadruped Robot SLAM Using 4D SoC Radars

This study presents a simultaneous localization and mapping (SLAM) framework for quadruped robots equipped with 4D system-on-chip (SoC) radar sensors. Because of their compact size and lower power consumption, 4D SoC radars are suitable for legged robot platforms operating in visibility-degraded environments. However, the low spatial resolution and low signal-to-noise ratio of radar point clouds make reliable place recognition and loop closure challenging. To address this issue, we propose a terrain-based place recognition method that exploits terrain-response characteristics of quadruped robots. Specifically, proprioceptive dead reckoning is fused with baro-radar-inertial odometry to stabilize vertical position estimates, thereby improving the consistency of terrain reconstruction from foot contact locations. The reconstructed terrain models are matched to identify revisited locations, and the accumulated radar point clouds are subsequently registered to estimate relative poses for loop closure. Experimental results from indoor and outdoor environments using a quadruped robot equipped with dual 4D SoC radars demonstrate that the proposed method improves terrain reconstruction consistency, enables reliable loop closure, and reduces accumulated localization drift compared with conventional radar odometry.