A New Robot Path Planning Method Based on LSTM Neural Network and Rapidly-Exploring Random Tree Algorithm

2021 ◽  
Author(s):  
Weifei Hu ◽  
Feng Tang ◽  
Zhenyu Liu ◽  
Jianrong Tan

Abstract As an important field of robot research, robot path planning has been studied extensively in the past decades. A series of path planning methods have been proposed, such as A* algorithm, Rapidly-exploring Random Tree (RRT), Probabilistic Roadmaps (PRM). Although various robot path planning algorithms have been proposed, the existing ones are suffering the high computational cost and low path quality, due to numerous collision detection and exhausting exploration of the free space. In addition, few robot path planning methods can automatically and efficiently generate path for a new environment. In order to address these challenges, this paper presents a new path planning algorithm based on the long-short term memory (LSTM) neural network and traditional RRT. The LSTM-RRT algorithm first creates 2D and 3D environments and uses the traditional RRT algorithm to generate the robot path information, then uses the path information and environmental information to train the LSTM neural network. The trained network is able to promptly generate new path for randomly generated new environment. In addition, the length of the generated path is further reduced by geometric relationships. Hence, the proposed LSTM-RRT algorithm overcomes the shortcomings of the slow path generation and the low path quality using the traditional RRT method.

Author(s):  
Yongmin Zhong ◽  
Bijan Shirinzadeh ◽  
Xiaobu Yuan

This paper presents a new methodology based on neural dynamics for optimal robot path planning by drawing an analogy between cellular neural network (CNN) and path planning of mobile robots. The target activity is treated as an energy source injected into the neural system and is propagated through the local connectivity of cells in the state space by neural dynamics. By formulating the local connectivity of cells as the local interaction of harmonic functions, an improved CNN model is established to propagate the target activity within the state space in the manner of physical heat conduction, which guarantees that the target and obstacles remain at the peak and the bottom of the activity landscape of the neural network. The proposed methodology cannot only generate real-time, smooth, optimal, and collision-free paths without any prior knowledge of the dynamic environment, but it can also easily respond to the real-time changes in dynamic environments. Further, the proposed methodology is parameter-independent and has an appropriate physical meaning.


2019 ◽  
pp. 491-511
Author(s):  
Yongmin Zhong ◽  
Bijan Shirinzadeh ◽  
Xiaobu Yuan

This paper presents a new methodology based on neural dynamics for optimal robot path planning by drawing an analogy between cellular neural network (CNN) and path planning of mobile robots. The target activity is treated as an energy source injected into the neural system and is propagated through the local connectivity of cells in the state space by neural dynamics. By formulating the local connectivity of cells as the local interaction of harmonic functions, an improved CNN model is established to propagate the target activity within the state space in the manner of physical heat conduction, which guarantees that the target and obstacles remain at the peak and the bottom of the activity landscape of the neural network. The proposed methodology cannot only generate real-time, smooth, optimal, and collision-free paths without any prior knowledge of the dynamic environment, but it can also easily respond to the real-time changes in dynamic environments. Further, the proposed methodology is parameter-independent and has an appropriate physical meaning.


Sign in / Sign up

Export Citation Format

Share Document