Robot Motion Planning Method Based on Incremental High-Dimensional Mixture Probabilistic Model

Fusheng Zha,Jingxuan Li,Yizhou Liu,Xin Wang,Fei Chen,Wei Guo

doi:10.1155/2018/4358747

Abstract

The sampling-based motion planner is the mainstream method to solve the motion planning problem in high-dimensional space. In the process of exploring robot configuration space, this type of algorithm needs to perform collision query on a large number of samples, which greatly limits their planning efficiency. Therefore, this paper uses machine learning methods to establish a probabilistic model of the obstacle region in configuration space by learning a large number of labeled samples. Based on this, the high-dimensional samples’ rapid collision query is realized. The influence of number of Gaussian components on the fitting accuracy is analyzed in detail, and a self-adaptive model training method based on Greedy expectation-maximization (EM) algorithm is proposed. At the same time, this method has the capability of online updating and can eliminate model fitting errors due to environmental changes. Finally, the model is combined with a variety of sampling-based motion planners and is validated in multiple sets of simulations and real world experiments. The results show that, compared with traditional methods, the proposed method has significantly improved the planning efficiency.

Highlights

In recent years, as robots play an increasingly important role in industrial production and daily life, the issue of motion planning has received extensive attention
In response to this problem, sampling-based motion planning algorithms have developed rapidly and received widespread attention [2,3,4]. This type of methods do not explicitly describe obstacles. They rely on the collision query module to provide feasibility information of the candidate trajectories and connect a series of collision-free samples to generate a feasible path from the initial state to the goal region
Sampling-based motion planning algorithms can generate a large number of labeled samples with spatial collision information in planning instances, which provides a necessary condition for the implementation of machine learning methods

Summary

Introduction

As robots play an increasingly important role in industrial production and daily life, the issue of motion planning has received extensive attention. For some robots with high planning dimensions, their configuration space’s obstacle region cannot be explicitly described In response to this problem, sampling-based motion planning algorithms have developed rapidly and received widespread attention [2,3,4]. Arslan and Tsiotras [12] use the kernel function to learn the feasibility and heuristics of samples generated in the previous planning instances and increase the sampling density of the areas that may make current path cost lower This method can improve the convergence rate of asymptotically optimum motion planning algorithms RRT# [13]. The main contributions of this paper include the following: (1) The influence of number of components on the prediction accuracy is analyzed, and a method for adaptively training GMM of collision region in robot highdimensional configuration space based on the convergence of log-likelihood function is proposed. The paper is organized as follows: Section 2 introduces the main motivation of the research work; Section 3 introduces the incremental training process of GMM using Greedy EM clustering algorithm, and the integration with motion planning algorithms; Section 4 shows the simulations and the real world experiments’ results; Section 5 summarizes the results and provides directions for future work

Motivations and Problem Statement

Algorithms

Simulations and Experiments

Planning Methods

Findings

Conclusion