WEEK 10

PREFACE TO WORK

As the next step of the project , it is necessary to have a optimal prediction method to predict the contact distance and the material of the object. For that purpose Artificial neural networks was chosen with the ideas gain from the reference papers.

PROGRESS

For the real time implementation of the tactile sensors to the autonomous robot , the language that used to code should have speed execution time and process time. C++ and the python were selected as the suitable coding language. Since the neural network is totally new area to touch , it is necessary to get the proper idea of the neural networks and how they work.

Therefor as the initial step to the neural networks , reading materials were referred and try to get clear idea about the CNN (convolution neural networks ) and ANN (Artificial neural networks).

The classification task is done for 5 materials (Aluminum, PVC, Brass, Wood and Steel). Therefore a pattern recognition tool was used here. Since the aim is to find a domain which is classified with a higher accuracy at higher dimension. Before going to classify five materials it was tested using only 2 materials(Aluminum and Brass).The data set was created by changing the contact distance by 5mm along the antenna(only the 80% distance of the antenna). Initially the neural architecture was built with a single hidden layer and less number of neurons. The activation function was kept as default (tanh function).some learning algorithms were tested here. Such as gradient descent, gradient descent back propagation etc. The number of iterations were 5000 epochs. There was also some earlier stopping algorithm used to stop the training process when the model is over fitted.