Safe Reinforcement Learning for Autonomous Lane-Changing: Considering the Collision Time and Lane Boundary Constraints

To ensure the safety of autonomous driving decision-making, a lane-changing decision-making framework is proposed based on safe reinforcement learning, it considers collision time constraints and lane boundary constraints. By minimizing the cumulative cost value and constraining unsafe actions, the framework effectively avoids potentially dangerous behaviour and guarantees the safety of lane-changing decisions. The agents are trained in CARLA to provide realistic driving scenarios and evaluate the proposed algorithm. Finally, the randomly generated scenarios are used for the tests. Simulation results demonstrate that the method proposed in this paper enables continuous lane-changing decisions while maintaining a safe driving distance in a dual-lane scenario. Compared with the reinforcement learning algorithm that does not consider safety constraints, it performs higher safety.