权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

Neural Networks Based Sensor Fusion and Adaptive Control for Robotics

基于神经网络的机器人传感器融合和自适应控制

基本信息

批准号：
01550199
负责人：
INABA Masayuki
金额：
$ 1.28万
依托单位：
The University of Tokyo
依托单位国家：
日本
项目类别：
Grant-in-Aid for General Scientific Research (C)
财政年份：
1989
资助国家：
日本
起止时间：
1989 至 1990
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-01550199/
关键词：
Neural Network Robot Vision Backpropagation Pattern Recognition Sensor Fusion Robot Control Parallel Processing Transputer ロボットセンサ-適応制御

项目摘要

The purpose of this research is to study techniques of neural network system for sensor fusion and adaptive control of intelligent robot and develop an neural networks simulator on parallel distributed system of multiple transfers. The following summarize the several topics in this research.1. Development of Neural Net Simulation System on Multiple Transputer System with Sensor Interface. The simulator uses eight transputers with 4MB for neural net calculation. One board in this system has one transputer T800 with 1MB memory and 32bit memory-mapped I/O to make interface with sensor devices. In order to develop several kind of neural works on this simulator, multi-lisp system on multi-transputer is developed.2. Experiment of pattern recognition with several human face images. This aims checking the performance of the simulation system. Each image has the size of about 64 by 64. It takes few minutes to finish learning of the images of six persons.3. Experiment of neural networks for motion control with feedforward and feedback. The input data to neural networks with three layers are the goal reference velocity to the robot control system, the past output torque and the past actual velocity. The teaching data to the network is the goal torque of the robot system. As the past data of output torque and velocity are also feedbacked to the networks, the network can learns feedback control process model in this configuration. In the experiment, a simple mobile robot which has two degree of freedom work under this networks.4. Neural networks simulation to correlate sensor information derived from different kind of sensors. In this networks, several layered neural networks give its output signal to another network as teaching data. In order to separate the substantial information between difference sensors, an independent factor is defined and an expansion method of the network output is introduced.

本研究的目的是研究智能机器人的传感器融合与自适应控制的神经网络系统技术，并开发一个基于并行分布式多传输系统的神经网络仿真器。以下是本研究的几个主要内容.带传感器接口的多传感器系统神经网络仿真系统的开发。模拟器使用8个4 MB的晶片机进行神经网络计算。该系统中的一块板上有一个具有1 MB内存和32位内存映射I/O的Transputer T800，用于与传感器设备进行接口。为了在该模拟器上开发多种神经网络，开发了多Transputer上的多Lisp系统.几幅人脸图像的模式识别实验。目的是检验仿真系统的性能。每个图像的大小约为64乘64。完成六个人的图像的学习需要几分钟。前馈与回馈之类神经网路运动控制实验。三层神经网络的输入数据是机器人控制系统的目标参考速度、过去的输出转矩和过去的实际速度。网络的示教数据是机器人系统的目标转矩。由于输出转矩和速度的过去数据也被反馈到网络，网络可以学习反馈控制过程模型。实验中，一个简单的两自由度移动的机器人在该网络下工作.神经网络模拟关联来自不同类型传感器的传感器信息。在这种网络中，几个分层的神经网络将其输出信号作为教学数据提供给另一个网络。为了分离不同传感器之间的大量信息，定义了独立因子，并引入了网络输出的扩展方法。