Neural Networks and Robotics

Neural networks has a unique feature of robust processing and adaptive capability in changing even in noisy environments. It is estimated that the human brain contains over 100 billion neuron. Several application has been implemented using artificial neural networks (ANN) such as pattern recognition, image processing, and machine learning.

A basic ANN model consists of a large number of neurons linked to each other with connection weights (see Figure 8).

 
Figure 8: A basic ANN model.  

The ANN processing can be divided into two phases:

Retrieving Phase:

this phase performs the iterative updating of the activation values . A generic iterative formulation for the updating phase is:

where, is the input to the PU i, is the activation value of PU is the effect of PU j on PU i, an external input to PE i, and is a nonlinear activation function at PE i.

Learning Phase:

In this phase, the synaptic weights are updated based on the input and the target training pattern using an adopted learning rule. The following is an example of learning rule:

where is the update rate parameter, and is the increment of weight change.

The application of neural networks can be grouped into two classes: optimization and associative retrieval/classification. Most robot problems can be formulated as one of the two classes. For example, stereo vision for task planning, autonomous robot path planning, and position control can be formulated as optimization problems.

A ring VLSI systolic architecture for implementing ANNs with application to robotic processing was proposed in [27]. It is demonstrated that the ANNs are suitable for several robot processing applications such as: task planning, path planning, and path control. Several models of ANNs are investigated in this paper such as: single-layer feedback neural networks, competitive learning neural networks, and multi-layer feed-forward neural networks.