A Novel Online Safe Reinforcement Learning with Control Barrier Function Technique for Autonomous vehicles

This paper introduces an innovative reinforcement learning-based controller that incorporates the control barrier function (CBF) approaches for an optimal constrained control. This method predominantly guarantees safe operation and optimal performance which is applicable to safety-critical systems. The safe reinforcement learning (RL) is formulated to minimize a specific cost function involving a customized CBF alongside a Lagrangian multiplier via Karush-Kuhn-Tucker (KKT) conditions. This approach approximates the optimal cost and control policy through the solution of the Hamilton-Jacobi-Bellman (HJB) equation, thereby ensuring that the system states are remained within the specified safe region while simultaneously preserving stability. To alleviate the challenge posed by the system’s inability to guarantee safety, we propose a novel actor-critic-Lagrangian neural network (NN) framework. This approach enables the system to effectively learn optimal policies while remaining to safety constraints. The efficacy of the proposed method is carried out on a lateral vehicle dynamic in autonomous vehicles.