TÉCNICAS E INSTRUMENTOS DE RECOLECCIÓN DE DATOS

The proposed work is based on combined effect of localization with recurrent neural net- work (RNN). This thesis explores; how to plans mobile robot to predict its position period- ically using odometry system as well as how to updates its coordinates systems.

The RNN is implemented for path planning and navigation after sensor fusion as well as updates its navigation map according to sensor data. In addition, ‘Se’ stands and includes the sampling data for further reference and express robot configuration in the Cartesian coordinates system (O, x, y) based on rotational principle.

In fact, the robot calculates its future or current configuration with reference to previously position and navigation action. Indeed, RNN stands for position parameters (X,Y,θ) and

work already executed by the robot as the output ‘S’. Accordingly, the network provides the new configuration to the robot (X, Y,θ), depends upon previous network. After that,

combined all network to evaluate the final output as presented in ‘Fig. 6.5, 6.6 and 6.7’. In addition, hyperbolic tangent sigmoid transfer function is used for the hidden layer and linear transfer function used for the output layer. According to the means squared errors, measurement of the network’s performance has been done well “Fig. 6.7”. Four neurons are chosen for hidden layer. At the time of training phase navigation, significant improvement has been considered with robot movement and it is all about due to communication with RNN over simply neural network. Finally, the fixed architecture with different data of RNN and the outputs of test phase has been represented by “Fig. 6.6 and 6.7”. In this Fig straight line represents the desire position at each interval. Including four inputs for path planning (i.e. ‘X-Xre f’,‘Y-Yre f’, ‘θ’ and function of obstacles ‘F’ and one RNN output (i.e. ‘S’)

the overall RNN processes has been projected. With RNN, after setting the weights and preferences; system tested with different base. “Fig. 6.6 and 6.7” represents the result of MSE for test phase of RNN with curves.

Figure 6.6: Value in test phase

Figure 6.7: Curves represent the errors of train (blue line), validation (green line), test (red line) and best (dotted line) data

Figure 6.8: Multilayer RNN for implementation of robotic behaviours

the mobile robot, at that time inaccuracy of the robot configuration substantially increases the inaccuracy due to cumulative errors, which created during the integration of different elements with displacement of robot. Accordingly, the use of time learning methodology reduces these types of errors and learning creation achieve preciously by robot.

The whole process of RNN is represented by Fig. 6.8. In this Fig, first layer is used as input layer at which four neurons (LOD = Left obstacle distance, FOD = Front obstacle distance, ROD = Right obstacle distance and T.A = Target Angle) are combined with the network with individual sensors data (S1, S2.). Then, the robot network contains two hid- den layers as shown in Fig 6.8 which tune the weight of neuron; as with one hidden layer it is difficult to train the network within a definite error limit. The training error has been defined as the variation between desired output and real output. From above Fig, there are two hidden layers where first hidden layer has eighteen neurons and the second one has five neurons. Finally, an output layer is generated with a single neuron which provide steering angle to control the direction of movement of the mobile robot. Real time recurrent learn- ing method is used to decrease the error and improve the path at time of online navigation towards target point.

Table 6.1 represents the simulation results for robot navigation using RNN technique. The errors between experimental and theoretical path analysis with respect to time also pre- sented.

Table 6.1: Overall path length, time taken and errors between results

Algorithm TSR (P)* ESR (P)* TSR (T)* ESR (T)* ERROR

FLC 0.739 0.745 00:28:192 00:28:876 0.8053

“TSR (P)” = Theoretical Simulation Result Related to Path Length (in meter). “ESR (P)” = Experimental Simulation Result Related to Path Length (in meter). “TSR (T)” = Theoretical Simulation Result Related to Time Taken (in minute). “ESR (T)” = Experimental Simulation Result Related to Time Taken (in minute).

“ERRORS” = Between Path Length i.e. TSR and ESR

6.6

Summary

This investigation is based on RNN for path planning as well as development of autonomous navigation learning algorithm for a mobile robot, functioning in a known and partially unknown environment in presence of obstacles. Firstly, we learnt obstacle avoidance al- gorithm and secondly developed the localization technique for mobile robot navigation, which is essential stage for development of path planning algorithm. In addition, obstacle avoidance algorithm plays an important role while mobile robot developing the collision free path. To control the robot, two integrated RNNs algorithm are developed and both are connected in series. Environmental map develop by robot depends upon sensor fusion and learning of RNN. The advantage of RNN based methodology is that, it can’t require any heavy mathematical model. The robot motion depends upon RNN network which is connected in series and integrated over time to time. In this chapter, firstly RNN helps to develop the localization technique after that, embedded for path planning. As a result, the developed intelligent algorithm offers the mobile robot to construct its collision free path as well as robot able to find its target in an environment. In addition, the developed algorithm is easy to implement in realistic world. Finally, it offers abundant capability to the robot, to estimate its position inside environment with precious rate and attain target in efficient manner.

H

ARDWARE

A

NALYSIS

The Hemisson is a cheap mobile robot which has been principally developed for the require- ments of research institute. Hemisson has attractive features for students and researchers in any science and engineering institute.

7.1

Introduction

Two Differential drive motors has been engage with robot system. Motors are receiving data from eight IR (Infrared) light and distance sensors mounted on the robot body. CeeBot- Hemisson communicates with computer through a serial RS-232 interface or via Bluetooth. Required programming can be done with CCS C compiler and also can be downloaded in controller of robot by the serial port using the Hemisson Uploader. Other equipment’s include programmable LED, buzzer and switches. The Hemisson robot equipped with 8bit microcontroller unit.