Lidar based ground plane estimation, hybridVisual-lidar odometry and navigation of Autonomous trucks

Autonomous trucks are gaining significant attention as they aim to address challengessuch as driver shortages, operational safety, and efficiency. To harness thepractical and financial benefits of these systems, both industry and academia areincreasingly focusing on the application of autonomous truck technologies. One primaryuse case is in mining operations, where trucks must navigate dynamic, uneventerrains. The natural characteristics of these environments pose unique challengesfor robotics, requiring innovative research to address them.This thesis focuses on several fundamental problems in robotics, including groundplane and traversability estimation, motion planning and control, and localization.While these are broad topics with extensive prior research, we propose novel approachesthat differ from existing works. Specifically, we introduce a polar gridrepresentation of the ground that adaptively arranges its bins based on the terrain’sunique characteristics. This adaptive representation, combined with traversabilityestimation tailored to truck dynamics and dimensions, enables more accurate occupancymapping of the environment, leading to safer motion planning. Additionally,this thesis presents the design of a system architecture for autonomous trucksoperating in mines. Known as GPENS, this system consists of several essentialmodules for navigation, including point-to-point travel and reversing in loading andunloading areas. At a high level, GPENS integrates perception, localization, planning,and control modules. The perception module incorporates ground plane andtraversability estimation, as well as road boundary detection, to ensure completeivness and identify safe driving paths. A dataset was created using real truck-trailercombinations operating in a mining environment, allowing us to validate the effectivenessof the proposed algorithms. Our ground plane and traversability estimationsolution improves precision by 2% compared to competing algorithms.Motion planning for mining truck navigation can often be adapted from structuredsolutions. However, reversing truck-trailer combinations in loading and unloadingareas presents a unique challenge. Efficient coordination between navigation andparking planners is critical to ensure the parking tracker performs optimally, as reversingsuch combinations is inherently unstable. In this work, the proposed planneralgorithms operate offline to determine maneuver switching points, and paths, whichare then tracked using optimal control and geometric tracker approaches. The effectivenessof the designed planner and trackers is demonstrated through real-worldtests with truck-trailer combinations in actual mining environments.Another key focus of this thesis is localization, specifically addressing the odometryproblem. Recent advances in deep learning have led to widespread interest inend-to-end solutions. However, hybrid approaches, which combine deep learningestimations with mathematical models, have proven more reliable due to their abilityto accurately represent underlying phenomena. Building on this foundation, weenhance hybrid odometry solutions by incorporating LiDAR point clouds. Whilemonocular depth estimation has advanced significantly, it often suffers from scaledrift errors when used for scale recovery in odometry systems. In contrast, LiDARprovides metric-scale but sparse depth measurements. This thesis proposes a fusionsystem that combines monocular depth estimation with LiDAR sparse depthdata, improving overall effectiveness. Two alternative fusion approaches are introduced,achieving up to a 40% improvement over baseline performance on the KITTIodometry dataset.