Smoother Robot Control from Convolutional Neural Networks Using Fuzzy Logic
A recent development in robotic control systems is the use of classification neural networks to produce discrete control signals. While the output of the neural network can be mapped directly to control signals, the possible control decisions are generally limited to the number of output classes. Fortunately, the neural network is also capable of producing a set of probabilities that a given input belongs to each class. This additional information can be used to expand the range of possible control signals. In this experiment, a fuzzy control system was applied to a robot that previously implemented a navigation-by-classification approach. The results of the experiment showed that the network is indeed capable of providing extra information in its probability designations, and this information can be exploited to smooth the discrete outputs of the system using fuzzy logic. In the case of the robot studied, the fuzzy control system outperformed the original discrete control system in navigating new courses. The robot's movements also appeared smoother when compared to the original solution. Thus, the additional information gained from the probabilities enabled the system to be more generalized and robust.
Robots, Neural networks, Navigation, Fuzzy logic, Training, Cameras, Control systems
W. Born and C. Lowrance, "Smoother Robot Control from Convolutional Neural Networks Using Fuzzy Logic," 2018 17th IEEE International Conference on Machine Learning and Applications (ICMLA), Orlando, FL, USA, 2018, pp. 695-700, doi: 10.1109/ICMLA.2018.00110.