Effectiveness of Unsupervised Training for Applications in Deep Neural Network Robot Navigation

Supervised training of neural networks is time consuming, and the scenarios required for obtaining a representative dataset must be carefully considered for each task. Applying an unsupervised training approach can greatly simplify this data collection aspect. This paper explores options for the unsupervised training of a convolutional neural network for the navigation of a mobile robot and compares its benefits with respect to a supervised training approach. A simulated training environment was created, in which the robot, through random motion, gathered the required data needed for training. Two approaches to training were investigated: either selectively choosing the training data from the random set acquired or considering modifying the network output to favor improved navigation. Both methods proved successful at obtaining an optimum value of 80% efficiency of directional travel whilst maintaining a collision avoidance performance of 97.7%. The results proved our approach was comparable in performance with respect to supervised training approaches, whilst it demonstrated superiority in terms of training-data collection.