Synthesis of Certifiable Control Software for Lane Change in Autonomous Vehicles

Description du projet

Solution correcte par conception pour la conduite autonome

La technologie de la conduite autonome progresse rapidement et suscite un vif intérêt de la part des secteurs publics et privés. Toutefois, la mise au point d’un logiciel capable de garantir de manière fiable un comportement de conduite sûr au niveau des voies reste un défi majeur. Les approches correctes par construction peuvent remédier à ce problème en concevant des systèmes de contrôle qui garantissent un comportement sûr dès le départ. Le projet CertiLane, financé par le CER, vise à développer un planificateur et un contrôleur de changement de voie automatique innovant, ainsi que des preuves vérifiables par machine afin d’améliorer la synchronisation et la sécurité. Le projet associera des méthodes de construction correctes à une chaîne d’outils efficace et complète.

Objectif

Autonomous driving keeps advancing, yet one obstacle persists: building lane-level behaviors whose software is provably safe. In current practice, requirements are written in natural language, interpreted inconsistently, and implemented with ad-hoc heuristics. Human error and limited test coverage leave edge cases unaddressed, while mileage-heavy validation is slow and costly. The result is fragile behavior and eroded public trust. Building on our research project AutoCPS, we will turn correct-by-construction methods into a product focused on a single high-impact feature: a certifiable Automatic Lane Change capability.

The goal is an Automatic Lane Change planner and controller accompanied by machine-checkable evidence for safety and timing, plus explicit operating assumptions. Requirements are captured as formal, reviewable rules; software is then automatically synthesized and delivered with certificates and runtime monitors for safety cases. Validation proceeds in stages: first in the MORAI driving simulator, then on a scaled roadway, and finally on an instrumented vehicle with an NVIDIA Jetson Orin. The toolchain reuses and hardens assets from our prior work—pFaces for compute acceleration, PIRK for predicting safe envelopes around other actors, AMYTISS for handling uncertainty, and OmegaThreads for compiling temporal rules.

End users are planning, safety, and verification teams at car makers and suppliers. The outcome is a licensable kit generating verified Automatic Lane Change software and certificates, with optional integration support. This proof of concept will (i) deliver the prototype and evidence package; (ii) quantify reductions in validation mileage and computing effort; (iii) complete a freedom-to-operate and intellectual-property plan; and (iv) secure early-access partners. Success will demonstrate a proof-driven alternative to mileage-based validation, improve road safety, and open a path to a family of certifiable functions beyond lane change.