The adoption of robots in inspection tasks is rapidly expanding thanks to recent advances in autonomous navigation technologies and the availability of affordable sensors and electronic systems. However, in harsh and dynamic environments, achieving robust and safe autonomous navigation remains a major challenge, especially in GNSS-denied areas. In this work, we propose a novel navigation method based on a wall detection and following strategy to enable autonomous robotic inspection of vertical surfaces in GNSS-denied environments, such as caves, tunnel-like environments, or row-crop fields like orchards and vineyards. In contrast to GNSS-based navigation methods, our approach is based only on local measurements from a noisy 3D point cloud of the environment acquired by an on-board RGB-D sensor. The proposed method features two main modules, namely a self-localization module and a control module. For localization, two different plane estimation algorithms are investigated, which exploit the geometric properties of the point cloud to estimate the robot’s relative pose in terms of distance and orientation. Then, according to the kinematic model of the autonomous vehicle, a wall-following algorithm is developed for lateral vehicle control, and a wall-following control law is built, after its theoretical stability in the presence of noisy measurements is proven. The whole system is implemented in ROS2 and validated first in the laboratory and then under field conditions in a vineyard environment. Although validated in a vineyard environment, the approach is applicable to other vertical-surface scenarios such as tunnels, caves, and more. Furthermore, the plan estimation methods are compared in terms of statistical metrics. In particular, in field conditions, the Least Squares method achieved an RMSE of 0 . 021 m with an average computational time of 0 . 5 ms , outperforming RANSAC-based methods, making it suitable for real-time implementation. We prove that the methodology is a reliable and efficient solution, suitable for real-time execution, thus resulting particularly useful for robotic platforms operating in hostile contexts or for resource-limited applications. • A self-localization strategy based on various plane estimation methods, such as the Least Squares, the Sample Consensus Model Plane, and the Sample Consensus Model Normal Plane. It is shown that different methods lead to different accuracy and computational time. • A control law fed by the relative position error, obtained by comparing the estimated pose of the robot with the desired pose. It is demonstrated both theoretically and experimentally that the convergence of the control law depends on the accuracy of the pose estimation, which in turn is influenced by the noise level in the reconstructed 3D point cloud. • An integrated robotic system implemented in ROS2, featuring a mobile robotic platform and an RGB-D camera, for experimental validation of the proposed strategy under field conditions. The code is made available as an open-source ROS2 package to facilitate reuse and further development by the research community. Although other wall-following packages are available that mostly employ laser scanners for indoor applications, the software package developed in this work extends the wall-following task to consumer-grade RGB-D cameras and outdoor environments.
Rana et al. (Fri,) studied this question.