Continuous Integration of Neural Networks in Autonomous Systems

The perception of the autonomous driving software of the FS223, a low-level sensor fusion of Lidar and Camera data requires the use of a neural network for image classification. To keep the neural network up to date with updates in the training data, we introduce a Continuous Integration (CI) pipeline to re-train the network. The network is then automatically validated and integrated into the code base of the autonomous system. The introduction of proper CI methods in these high-speed embedded software applications is an application of state-of-the-art MLOps techniques that aim to provide rapid generation of production-ready models. It further serves the purpose of professionalizing the otherwise script-based software production, which is re-done almost completely every year as the teams change from one year to the next.