Analysis and Design of a Guided Sampling based Path Planning using CNNs

Abstract

Intelligent sampling has a strong effect on path planning performance. Learning sampling distributions from the expert planning algorithm demonstrations could contribute to an optimized and improved planning performance. In this study, we offer a novel CNN-based network to predict suitable sampling distributions for a faster path planning process. We also propose several improvements and modifications to strengthen the link between intelligent sampling networks and path planning. Our proposed method is tested against the more commonly used random sampling approach in various conditions (i.e. three different sample sizes, two different path planners). The test results showed that the proposed method is remarkably more sample efficient when compared with conventional planning approaches on large sample sets. Additionally, this novel approach results in a more user-friendly and intuitive design, with much less computational parameters, hence, in a path planning approach that is more convenient for real-time implementation.