Terrain-Aware Autonomous Navigation

Abstract

Autonomous ground vehicles (AGVs) are considered to be critical for the future of the military. As military vehicles often need to operate with high mobility, even on off-road deformable terrains, AGVs need to meet the same requirements. However, current navigation strategies hinder the application of AGVs. This is because AGV mobility critically depends on the properties of the terrain, but these properties can vary significantly during the operation of the vehicle and cannot be directly measured with the sensors available on-board. If the trajectory planning algorithms that navigate these vehicles are not aware of such changes, they can devise plans that are difficult or even infeasible for the vehicles to execute, leading to performance and safety issues. Therefore, the overarching goal of the thesis is to develop, implement, and evaluate a terrain-adaptive trajectory planning algorithm for improved safety and performance of autonomous vehicles in off-road conditions.

To achieve this goal, this thesis addresses four major research goals. First, due to limitations in operating conditions, computational efficiency, and continuous differentiability of existing state-of-the-art terramechanics models, a novel deformable terrain terramechanics model is developed to satisfy the differentiability requirements of optimal control while also remaining efficient and capable of dynamic operation. Second, state-of-the-art terrain estimation algorithms face limitations in application due to sensor setup, model simplifications, computational efficiency, and requiring low speed vehicle operation. To address this, a method is developed to estimate the terrain properties online from measurements available on-board the vehicle. Third, a single-level adaptive model-predictive trajectory planning and tracking algorithm is developed that uses the estimated terrain properties, and developed terramechanics model, to continuously adapt to the environment and make its model-based mobility predictions while optimizing the trajectory. Finally, Uncertainty-based Contingent Model Predictive Control is presented to extend the terrain-adaptive framework to be robust to uncertainties in the terrain estimate.

Therefore, this dissertation addresses the following original contributions:

1. Development of an efficient, dynamic, and twice continuously differentiable deformable terrain terramechanics model.

2. Development of a terrain estimation algorithm capable of real-time performance.

3. Development of a single-level terrain adaptive trajectory planning and tracking algorithm for off-road autonomous vehicles operating on deformable terrains.

4. Development of Uncertainty-based Contingent MPC, a novel selectively-robust MPC formulation to increase robustness to terrain estimate uncertainty.