A fast and accurate compound collision detector for RRT motion planning

Abstract

The application of artificial intelligence tools has led to newly developed collision detectors which have better computational efficiency than the kinematics-and-geometry based collision detectors (KCD) to improve robot motion planning strategies. However, new detectors are not very accurate in some cases. To improve the accuracy, a trade-off between efficiency and accuracy is required. We propose a novel compound collision detector (CCD) for collision queries that modifies the planners of rapidly-exploring random tree (RRT) to improve the classical probabilistic collision detector (PCD). It is composed of an exact collision detector (ECD), an inference collision detector (ICD) and a strategy to determine ECD or ICD based on some conditions. In our CCD, we use a sphere-ellipsoidal pseudo distance (SEPD) in the determination strategy to alleviate the problem of highly-frequent outputs of false-positive in narrow passages of PCD, and a node based bounding method (NBB) to increase the speed of data storage and loading for the sub-algorithm ICD. Experiments on a Kinova Jaco assistive robotic arm are taken to evaluate the performance of our CCD, which show an improved accuracy with a small reduction of speed in comparison with PCD. So, it is a promising tool in robot motion planning.