Benchmark Analysis of Sampling Methods for RRT Path Planning

Abstract

Path planning is a crucial aspect of mobile robot navigation, ensuring that robots can safely travel from their initial position to the goal. In real-world applications, path planning is essential for autonomous vehicles, drones, warehouse robots, and rescue robots to navigate complex environments efficiently and safely. One effective method for path planning is the Rapidly-exploring Random Tree (RRT) algorithm, which is particularly practical in maze-like environments. The performance of RRT depends on the sampling methods used to explore the maze. Sampling methods are important because they determine how the algorithm explores the search space, affecting the efficiency and success of finding an optimal path. Poor sampling can lead to suboptimal or infeasible paths. In this study, we investigate different sampling strategies for RRT, specifically focusing on uniform sampling, Gaussian sampling, and the Motion Planning Network (MPNet) sampling. MPNet leverages a neural network trained on past environments, allowing it to predict promising regions of the search space quickly, unlike traditional methods like RRT that rely on random exploration without prior knowledge. This makes MPNet much faster and more efficient, especially in complex or high-dimensional spaces. Through a benchmarking analysis, we compare these methods in terms of their effectiveness in generating feasible paths. The results indicate that while all three methods are effective, MPNet sampling outperforms uniform and Gaussian sampling, particularly in terms of path length. The mean path length generated, based on a sample size of 30, is 13.115 meters for MPNet, which is shorter compared to uniform and Gaussian sampling, which are 18.27 meters and 18.088 meters, respectively. These findings highlight the potential to enhance path planning algorithms using learning-based sampling methods.