Traversability Analysis of Quadruped Robot Based on Sparse Point Cloud in Rough Terrain

Abstract

Building a dense and accurate traversability map in real-time is a prerequisite for robots to navigate in challenging environment with uneven terrain. Quadruped robots have strong terrain adaptability and are considered to be the first choice for locomotion over unstructured terrain in the wild. In such environments, LiDAR is frequently used to perceive the surroundings, but the sparseness of LiDAR scans may lead to incomplete perception. We present a novel approach for traversability mapping with sparse LiDAR scans collected by a quadruped robot, which can build dense terrain traversability map in real-time. We construct a Depth Completion Network (DCN) based on U-Net to obtain dense terrain descriptions through the sparse point cloud from LiDAR. We develop a method to estimate terrain traversability using four terrain features extracted from a dense terrain description: step height, surface slope, surface roughness, and vegetation density. Our method considers the physical properties of the robot, including the maximum leg lift height, maximum climbing angle, etc., which can be applied to a variety of robotic platforms. We evaluate our method in the wild, demonstrating that it can be used online in the real-world applications.