evilheart
Member level 3
hi i am a fourth year communication and electronics student ,
but my graduation project is linked with robotics , in one pat of our project we should control a lynxmotion AL5D arm with control signals from a laptop , and as we can't do this with the arm controller , we decided to make our own.
the problem we need to make the inverse kinematic model of the arm , but we are not very good in mechanics , and we don't want to waste time learning something somehow away from our field , and difficult in the same time !!!!
, so we decided to use neural networks to implement the inverse kinematics model , it's better for real time applications , and may even be implemented on a microcontroller.
our idea was to train the network virtually using matlab , then we take the weights and implement the network in C or C++.
we made the forward kinematics model for the arm , and use it to get the training data for the network.
The idea isn't new , but the resources and papers i found was not that clear about the problems that had encountered them.
the problem is performance now , Matlab nftool gives performance (mean square error) that can reach 700 or 500 !!!!!!!
it uses levenberg-marquardt training.
its our first time to work with neural networks ,
so if someone has tried this before , or have an idea about using neural networks in control , any hint or help will be appreciated !!!!!
but my graduation project is linked with robotics , in one pat of our project we should control a lynxmotion AL5D arm with control signals from a laptop , and as we can't do this with the arm controller , we decided to make our own.
the problem we need to make the inverse kinematic model of the arm , but we are not very good in mechanics , and we don't want to waste time learning something somehow away from our field , and difficult in the same time !!!!
, so we decided to use neural networks to implement the inverse kinematics model , it's better for real time applications , and may even be implemented on a microcontroller.
our idea was to train the network virtually using matlab , then we take the weights and implement the network in C or C++.
we made the forward kinematics model for the arm , and use it to get the training data for the network.
The idea isn't new , but the resources and papers i found was not that clear about the problems that had encountered them.
the problem is performance now , Matlab nftool gives performance (mean square error) that can reach 700 or 500 !!!!!!!
it uses levenberg-marquardt training.
its our first time to work with neural networks ,
so if someone has tried this before , or have an idea about using neural networks in control , any hint or help will be appreciated !!!!!