Agent-centric temporal logic planning for autonomous vehicles

Abstract

To responsibly introduce autonomous vehicles in practical use, we first need ensure their safety. However, testing autonomous vehicles require more time than what is actually feasible. Instead we can create a controller using synthesis, which provides a safe implementation for a formal description of the environment and system demands. Its practical use is limited due to how it is strongly affected by state space blowup, a factor heavily restricting the size of model variables. We design an agent-centric model for a vehicle with the objective to avoid obstacles and navigate towards a target, believing that this novel perspective will limit the state space issues while it, unlike other models with similar purpose, does not require live synthesis or preplanning on a scaling number of specifications. Further, we showcase a possible framework for integrating the controller into a physical vehicle system. We also present a simulator, allowing one to freely explore the behaviour of the agent-centric model. The final result is a model that is fully synthesised to a single controller in preplanning. The navigational model shows promise in its adaptability to unexpected map changes but fails to navigate through every task, getting stuck in certain situations and acting against common traffic norms in certain situations. We believe that the agent-centric model could potentially be used in a more flexible vehicle controller, and that both the agent-centric spatial approach and the directional target navigation are possible perspectives worth considering for work in this area.