Neural Network Based Obstacle Avoidance Using Simulated Sensor Data

This study characterized the design and implementation of a low-cost autonomous robot capable of performing obstacle avoidance using neural networks trained with simulated sensor data. The only sensor used for detecting the environment was an infrared distance sensor attached to a hobby servo, allowing for 180° of sensor visibility. In order to train the neural networks, simulated sensor data was created using LabVIEW and presented to a user, who selected the expected robot operation in that specific situation. The simulated sensor data and expected robot operation data was then used to create training data. The trained neural networks were then verified by testing the actual network output with the training data, and additional random sensor data. The robot was controlled wirelessly by a computer running LabVIEW, which processed the sensor data through the networks and controlled the robot’s subsequent movements. The networks were able to produce accurate obstacle avoidance actions during the simulated network analysis, and on the test bed, allowing the robot to avoid obstacles while successfully performing its mission. Keywords—neural network, backpropagation, infrared, LabVIEW, autonomous robot, dsPIC, microcontroller, obstacle avoidance, extended delta bar delta, simulation