The development of safe autonomous vehicles remains a challenge. Modern autonomous driving controllers are able to operate safely in most operating conditions, however, rare conditions may cause undesired behaviors. To alleviate this issue, a complex controller may be paired with a simpler, yet verified, controller in a simplex architecture. To switch between the two, we take inspiration from formal methods, namely reachability analysis, to develop an online method called neural reachability that estimates the reachable sets of the system. The method predicts the possible states of the system under uncertain initial states and disturbances in control. The approach utilizes deep neural networks for conservative approximation of reachable sets in bounded time. If an intersection of the reachable set with an unsafe set is detected, a simple verified controller is utilized. We show how a deep neural network is trained using existing reachability analysis algorithms and demonstrate how neural reachability may be utilized in a simplex architecture. Also, we provide probabilistic guarantees based on statistical model checking approaches. Finally, the approach is evaluated as part of a resilient safety architecture for autonomous vehicles in a simulated environment with several maneuvers. Our evaluation demonstrates that reachability analysis can be done within a fraction of a second and outperforms traditional nonlinear reachability tools by two orders of magnitude.
Runtime Assurance for Autonomous Driving with Neural Reachability
2022-10-08
1092690 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Runtime Assurance for Autonomous Aerospace Systems
AIAA | 2020
|REACHABILITY ESTIMATION IN DYNAMIC DRIVING SCENES FOR AUTONOMOUS VEHICLES
British Library Conference Proceedings | 2020
|Reachability Estimation in Dynamic Driving Scenes for Autonomous Vehicles
BASE | 2020
|Runtime Safety Assurance Using Reinforcement Learning
IEEE | 2020
|