This thesis presents steps towards development of cricket detection and classification software needed for an autonomous cricket farm. At first describing the farming process and its importance. Then showing how data samples are collected and used to train convolutional neural network models for detecting and classifying crickets. Training the models resulted in models having an accuracy equal to and higher than 99%. After that, iterations of the development of a simulated robot is presented, from which is concluded that the robot should be trained to differentiate between good and bad samples rather than between crickets and non-crickets. When tested and when the result was considered the best, the simulated robot successfully found 55% of all usable samples and of all samples found, 93.5% were usable and 6.5% were false positives, where 98% of the false positives are assumed easy to filter out.