A Probabilistic Models Fusion Based Contact Detection for Quadruped Robot

Abstract

Compared with wheeled or crawler robots, legged robot has inherent advantages when traversing irregular and discontinuous terrain. However, because of the lack of environment sensor and foot force measuring sensor, it is an extremely challenge for legged robot to perceive current terrain, and further switch control strategies of swing leg and contact leg. In this paper, a Kalman filter based multiple probabilistic models fusion method is applied to detect the actual contact state of robot foot. To avoid the computation cost of whole-body dynamics, an external torque observer of knee joint is proposed, which can estimate an accurate contact probability according to the external torque received by the knee joint. And a terrain detection based probabilistic contact model of foot height is proposed to improve the estimation accuracy of contact probability in unstructured terrain situation. Compared with single probabilistic contact model, this method improves the reliability and robustness of contact detection. Finally, experiments are conducted on Unitree A1 quadruped robot to verify the effectiveness of this algorithm. With the contact detection algorithm, the robot could traverse complex terrains, such as slope and step.