Synthesis of safe controller via supervised learning for truck lateral control

Correct-by-construction techniques such as control barrier function (CBF) have been developed to guarantee safety for control systems as supervisory controller. However, when the supervisor intervenes, the performance is typically compromised. On the other hand, machine learning is used to synthesize controllers that inherit good properties from the training data, but safety is typically not guaranteed due to the difficulty of analysis. In this paper, supervised learning is combined with CBF to synthesize controllers that enjoy good performance with safety guarantee. First, a training set is generated by trajectory optimization that incorporates the CBF constraint for multiple initial conditions. Then a policy is trained via supervised learning that maps the feature representing the initial condition to a parameterized desired trajectory. Finally, the learning based controller is used as the student controller, and a CBF based supervisory controller on top of that guarantees safety. A case study of lane keeping for articulated trucks shows that the student controller trained by the supervised learning inherits the good performance of the training set and the CBF supervisor never or rarely intervenes.