RCS: Automatic Navigation Framework for Legged Robots with Incremental RGB and Consistent Semantics Information

Autonomous movement of legged robots in outdoor scenes requires accurate state estimation and environment navigation maps. Different from wheeled and tracked mobile robots, legged robots can adjust parameters such as gait type, step height, gait frequency, torso height, foot friction, and other coefficients to better adapt to the geometric characteristics and physical properties of verstile terrains. To this end, in this paper, we propose a framework for autonomous navigation of legged robots, including the construction of geometric-property RGB consistent semantic map (RCS) and the autonomous navigation method under the RCS map. The method for building RCS map formulates lidar-inertial odometry based on a factor graph. Through end-to-end semantic segmentation of images, terrain information is projected and saved while mapping. For achieving pose and point cloud consistency, local and global lidar boudle adjustment (BA) is introduced to the optimization process. We demonstrated the effectiveness of RCS through outdoor navigation experiments on a quadruped robot. The experiment shows that our terrain detection method can achieve an average detection time of 30 ms on an embedded board.