Synthesis of Three-Layer Dynamic Binary Neural Networks for Control of Hexapod Walking Robots

Motion Planning Method for In-Pipe Walking Robots Using Height Maps and CNN-Based Pipe Branches Detector

Computational Intelligence in the Internet of Things - Advances in Computational Intelligence and Robotics ◽

10.4018/978-1-5225-7955-7.ch001 ◽

2019 ◽

pp. 1-24

Author(s):

Sergei Savin

Keyword(s):

Neural Networks ◽

Motion Planning ◽

Prediction Accuracy ◽

Planning Method ◽

Walking Robots ◽

Walking Robot ◽

Deep Convolutional Neural Networks ◽

Sequence Generation ◽

Simulation Based ◽

Inner Surface

In this chapter, the problem of motion planning for an in-pipe walking robot is studied. One of the key parts of motion planning for a walking robot is a step sequence generation. In the case of in-pipe walking robots it requires choosing a series of feasible contact locations for each of the robot's legs, avoiding regions on the inner surface of the pipe where the robot cannot step to, such as pipe branches. The chapter provides an approach to localization of pipe branches, based on deep convolutional neural networks. This allows including the information about the branches into the so-called height map of the pipeline and plan the step sequences accordingly. The chapter shows that it is possible to achieve prediction accuracy better than 0.5 mm for a network trained on a simulation-based dataset.

Download Full-text

Enhanced Footsteps Generation Method for Walking Robots Based on Convolutional Neural Networks

Handbook of Research on Deep Learning Innovations and Trends - Advances in Computational Intelligence and Robotics ◽

10.4018/978-1-5225-7862-8.ch002 ◽

2019 ◽

pp. 16-39 ◽

Cited By ~ 1

Author(s):

Sergei Savin ◽

Aleksei Ivakhnenko

Keyword(s):

Neural Network ◽

Neural Networks ◽

Convolutional Neural Network ◽

Network Design ◽

Bipedal Walking ◽

Training Dataset ◽

Walking Robots ◽

Foot Placement ◽

Gait Generation ◽

Free Region

In this chapter, the problem of finding a suitable foothold for a bipedal walking robot is studied. There are a number of gait generation algorithms that rely on having a set of obstacle-free regions where the robot can step to and there are a number of algorithms for generating these regions. This study breaches the gap between these algorithms, providing a way to quickly check if a given obstacle free region is accessible for foot placement. The proposed approach is based on the use of a classifier, constructed as a convolutional neural network. The study discusses the training dataset generation, including datasets with uncertainty related to the shapes of the obstacle-free regions. Training results for a number of different datasets and different hyperparameter choices are presented and showed robustness of the proposed network design both to different hyperparameter choices as well as to the changes in the training dataset.

Download Full-text

Implementation Method of the Robot Adaptation to Contact Interaction Mode Changes Using Deep Fully Connected Neural Networks

Proceedings of Southwest State University ◽

10.21869/2223-1560-2020-24-1-206-214 ◽

2020 ◽

Vol 24 (1) ◽

pp. 206-214

Author(s):

S. I. Savin ◽

L. Yu. Vorochaeva ◽

A. V. Malchikov ◽

A. M. Salikhzyanov ◽

E. M. Zalyaev

Keyword(s):

Neural Networks ◽

Control System ◽

Contact Interaction ◽

Training Sample ◽

Bipedal Walking ◽

Walking Robots ◽

Support Surface ◽

Walking Robot ◽

Additional Advantage ◽

Harmonic Components

Purpose of research. The present paper conserns the problem of using reaction predictors in the control system of bipedal walking robots. The main advantage of using predictors is the ability to exclude unknown reaction forces from the dynamics equations and, consequently, from the robot control problem statements based on the model. An additional advantage of predictor setting of control tasks is also discussed in the paper, namely the possibility of its use to predict changes in contact interaction modes, such as slipping motion or foot lifting from the supporting surface.Methods. The following methods are used in the research: the method of dynamics of multi-mass systems is necessary for developing a mathematical model of the behavior of a walking robot and describing its contact interaction with the support surface, the method of neural networks is used to develop a predictor that allows one to forecast the values of reactions between the robot’s foot and the surface.Results. The paper shows that there is a connection between the frequencies of the harmonic components of robot movements (the ratio p of these frequencies in the experiment and the training sample) and the quality of reactions predictor operation of the support surface. This indicates the importance of applying a representative spectrum of walking robot movement frequencies in forming a training sample, and the poor generalizability of the predictor in relation to movement frequency.Conclusion. The paper has considered the use of a reaction predictor to identify the possibility of changing the mode of contact interaction, based on the measurement of discrepancies between local linearizations for various discrete steps. The results obtained in this work will be used in the development of a motion control system for a bipedal walking robot, which allows the device to adapt to the parameters of the support surface on which the movement occurs.

Download Full-text