Loading ePaper
n powered by Artificial Neural Networks with fischertechnik TXT 4.0 — Thorsten Leimbach, Fraunhofer IAIS Version: 1 27.07.2024
Artificial Neural Networks with fischertechnik TXT 4.0 xNN This tutorial shows you how to teach a robot to maintain a minimum distance using a neural network. You should be familiar with programming in the Open Roberta Lab. You should have already created and applied a simple neural network in the Open Roberta Lab. Page 2 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks Maintain a minimum distance To make this experiment a bit simpler, we use the simulated robot from the Open Roberta Lab. However, you can also run the program on a real fischertechnik TXT 4.0 robot. Page 3 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot Clicking on opens the simulation environment. 2D representation of a simulated TXT 4.0 robot: Left wheel Right wheel The simulated TXT 4.0 robot has the following sensors: • Line camera • Color camera • Ultrasonic sensor • Infrared line sensor Page 4 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot - Ultrasonic sensor The simulated robot in the Open Roberta Lab is equipped by default with an ultrasonic sensor, among other things. The ultrasonic sensor allows the robot to measure the distance to objects. The ultrasonic sensor is located at the front of the simulated robot. Facts: • • • • The dashed lines simulate ultrasonic beams. The darker dashed line indicates the shortest distance to an object. The ultrasonic sensor cannot detect whether an object is to the left or right of the sensor. The maximum range of the ultrasonic sensor is 400 cm." Ultrasonic sensor Object Page 5 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot - Ultrasonic sensor Clicking on displays, among other things, the sensor values of the robot. This display is helpful for determining the distance to an object. Page 6 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot - Experiment Minimum Distance The goal of the experiment is to train the neural network so that the robot maintains a minimum distance of 30 cm from an object. The robot should drive straight ahead until it measures a distance of 30 cm. Distance Robot <-> Object Page 7 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab As an introduction to the 'Minimum Distance' experiment, the robot should first be programmed to adjust its speed based on the distance to the object. The closer the robot gets to an object, the slower it should move. This can be achieved by using the distance values measured by the ultrasonic sensor as the speed for the motors. Measured Distance = 100 Measured Distance = 50 Page 8 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab The use of the ultrasonic sensor values for the robot's speed can be easily programmed as follows. Speed = 50% Measured Distance = 50cm Page 9 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab Clicking on the tab takes you to the view of the neural network The neural network looks as follows by default:: Input neutron Edge weight Output neutron Bias / Threshold Page 10 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab To make the program easier to 'read,' the neurons should be given different names. By clicking on , you can change the names of the input and output neurons. • • The input neuron has been renamed to US (for Ultrasonic Sensor). The output neuron has been renamed to Speed. Speed Page 11 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab The program for controlling the speed can also be programmed with a neural network. Page 12 08.10.2024 © Fraunhofer IAIS Public Speed
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab The robot now moves faster or slower according to the measured distance. At a distance of 0 cm, the robot stops. Next, a minimum distance of 30 cm needs to be maintained. Idea: We simulate that the robot is already in front of an object, even though it is still 30 cm away Measured Distance = 30 → Speed should be = 0% Page 13 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab This idea can be implemented very easily by always subtracting the value 30 from the measurements of the ultrasonic sensor. A typical program might look as follows: Page 14 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab The idea can also be implemented with a neural network. The program would then look as follows: In addition to the program, the neural network must also be created. Page 15 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab Looking at the programs and the neural network together, it becomes clear that the value 30 is subtracted from the measured values of the ultrasonic sensor (US) due to the negative edge weight from Dist to Speed. Page 16 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab The goal now is to train the neural network so that it calculates a value of 0 when a distance of 30 cm is measured. For this, we need to switch to the tab . The view looks as follows: → The edge weights and the bias should initially be set with random values in the NEURAL NETWORK Define tab Page 17 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab To train the neural network, training data is needed. This can be entered via the button . A table for entering the training data will appear as follows: Page 18 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab The following values can be entered as training data: What do the values mean? • • • If the ultrasonic sensor (US) measures a distance of 100 cm, the robot should move at a speed of 70%. If the US measures a distance of 60 cm, the robot should move at a speed of 30%. If the US measures a distance of 30 cm, the robot should move at a speed of 0%, meaning it should stop. Page 19 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab Clicking on starts the training Page 20 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab Clicking on starts the training → This means that after 500 epochs, the neural network calculates a speed of 3.74 for an input distance of 30 cm. This indicates that the robot is still far from maintaining the minimum distance of 30 cm. Page 21 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab Clicking on starts the training → This means that after 1001 epochs, the neural network calculates a speed of 0.51 for an input distance of 30 cm. This indicates that the robot is getting closer to maintaining the minimum distance of 30 cm. Page 22 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab Clicking on starts the training → After approximately 2400 epochs, the neural network calculates a speed of 0.00 for an input distance of 30 cm. This means that the robot is maintaining the minimum distance of 30 cm. Page 23 08.10.2024 © Fraunhofer IAIS Public
Training artificial neural networks TXT 4.0 Robot – Neural Network in the Open Roberta Lab "What happens exactly? ▪ The adjustment is done via a process called backpropagation. • The training data is fed into the neural network. • The calculated result is compared with the result from the training data. is called the training loss. ▪ A complete pass through the training data is called an epoch. ▪ The learning rate is the speed at which the adjustment of the edge • If the result deviates from the training data, the edge weights and bias occurs. ▪ The deviation between the neural network's result and the training data weights and bias are adjusted. Page 24 08.10.2024 © Fraunhofer IAIS Public
Contact — Thorsten Leimbach Tel. +49 2241 14-2404 thorsten.leimbach@iais.fraunhofer.de Fraunhofer-Institut für Intelligente Analyse- und Informationssysteme IAIS Schloss Birlinghoven 1 53757 Sankt Augustin Germany www.iais.fraunhofer.de