Tensor-voting-based navigation for robotic inspection of 3D surfaces using lidar point clouds

This paper describes a solution to robot navigation on curved 3D surfaces. The navigation system is composed of three successive subparts: a perception and representation, a path planning, and a control subsystem. The environment structure is modeled from noisy lidar point clouds using a tool known as tensor voting. Tensor voting propagates structural information from points within a point cloud in order to estimate the saliency and orientation of surfaces or curves found in the environment. A specialized graph-based planner establishes connectivities between robot states iteratively, while considering robot kinematics as well as structural constraints inferred by tensor voting. The resulting sparse graph structure eliminates the need to generate an explicit surface mesh, yet allows for efficient planning of paths along the surface, while remaining feasible and safe for the robot to traverse. The control scheme eventually transforms the path from 3D space into 2D space by projecting movements into local surface planes, allowing for 2D trajectory tracking. All three subparts of our navigation system are evaluated on simulated as well as real data. The methods are further implemented on the MagneBike climbing robot, and validated in several physical experiments related to the scenario of industrial inspection for power plants.