Uncertainty-aware motion planning for ground vehicle in unstructured uneven off-road terrain

Abstract

Navigating large, unstructured, and uneven off-road environments, such as those encountered in search and rescue missions or planetary exploration, presents significant challenges. These environments are characterized by varying terrain semantics and complex geometries. Furthermore, initial map representations are often uncertain, as they are typically generated from aerial scans or other remote sensing techniques that may provide incomplete or outdated data. Existing algorithms that focus on planning a single path through the environment frequently overlook the opportunity to incorporate future information gathered during navigation, which can be used to reduce the expected traversal cost.

In this thesis, we propose an uncertainty-aware motion planning framework. The framework starts by integrating both geometric and semantic terrain data to assess terrain traversability. We then utilize an unsupervised region clustering algorithm to segment uncertain regions and group grids with similar visual and spatial features. Following this, our approach is structured into three stages: generating a network of pathways, constructing a stochastic graph, and developing an optimal navigation policy. A multi-query sampling-based planner is used to create a comprehensive network of pathways between the start and goal points, efficiently exploring multiple potential routes. These pathways are then converted into a topological stochastic graph representation of the environment, capturing uncertainty through probabilistic edge representations. The stochastic graph is modeled as a Canadian Traveler Problem (CTP), which is a decision-making framework designed for navigating graphs where some edges have a probability of being blocked. To minimize the expected traversal cost, we extend the state-of-the-art CTP solver CAO*, introducing Complete CAO* (CCAO*), which guarantees to produce a navigation policy that minimizes the expected traversal cost, even when no deterministic path exists.

We validate our framework through extensive simulations using real-world off-road data, testing both small and large environments to assess scalability. Results demonstrate that our approach consistently generates compact graph representations, unaffected by uncertain regions that do not impact the robot's movement. These findings highlight the framework's computational efficiency, robustness, and ability to reduce expected traversal costs when compared to traditional baseline methods.