Terrain Adaptive Trajectory Planning and Tracking on Deformable Terrains

• Published in: IEEE transactions on vehicular technology Vol. 70; no. 11; pp. 11255 - 11268
• Main Authors: Dallas, James, Cole, Michael P., Jayakumar, Paramsothy, Ersal, Tulga
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.11.2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptive control; Algorithms; autonomous ground vehicles; Autonomous vehicles; Bicycles; Collision avoidance; Computing time; Control algorithms; Control theory; Controllers; Deformation; Formability; Maneuvers; Mathematical models; model predictive control; Navigation; Neural networks; Parameterization; Predictive control; Terrain; terrain estimation; terramechanics; Tire force; Tracking control; Trajectory control; Trajectory planning; Vehicle dynamics; Vehicle safety
• ISSN: 0018-9545; 1939-9359
• DOI: 10.1109/TVT.2021.3114088

In this work, a novel single-level adaptive trajectory planner and tracking controller is developed for off-road autonomous vehicles operating on deformable terrains. Trajectory planning and tracking algorithms often rely on a simplified vehicle model to predict future vehicle states based upon control inputs, hence requiring accurate modeling and parameterization. On off-road deformable terrains this is a challenging task due to unknown terrain parameters and the complex interactions at tire-terrain interfaces, which pose issues in continuous differentiability, operating conditions, and computational time. To address these difficulties, in this paper, a neural network deformable terrain terramechanics model and its implementation within a terrain adaptive model predictive control algorithm is presented to improve vehicle safety and performance through more accurate prediction of the plant response. It is shown in simulations that the neural network is able to predict the lateral tire forces accurately and efficiently compared to the Soil Contact Model as a state-of-the-art model and is able to yield accurate bicycle model predictions. It is demonstrated that the implementation of the neural network within model predictive control can outperform both a baseline Pacejka-based and a rapidly exploring random tree controller by improving performance and allowing for more severe maneuvers to be completed that otherwise lead to failure when terrain deformations are not explicitly taken into account. The improved performance achieved through estimating terrain parameters online in an adaptive controller is highlighted against the nonadaptive realization. Finally, it is shown the algorithm is conducive to real-time implementation.