Safe-DRL: A Safety-conscious Deep Reinforcement Learning Decision-making Algorithm for Unmanned Platforms

To address the safety issue caused by unpredictable behaviors in traditional deep reinforcement learning (DRL) during inference, this paper proposes a safety-enhanced deep reinforcement learning (DRL) algorithm for autonomous driving in unmanned platforms across multi- task scenarios. The algorithm integrates an improved Markov process with an action recognition network for pre-execution safety assessment, and adopts a parallel dual-thread network architecture to suppress hazardous driving behaviors. Additionally, a novel kinematics-based reward function is designed to take into account driving safety and efficiency. In the highway-env environment, a comparative experiment is conducted on the proposed algorithm in three typical driving scenarios— single-lane roads, intersections, and roundabouts. It is shown that the proposed algorithm significantly improves driving safety and generalization capability. The results verify its effectiveness and potential for supporting the application of unmanned platform in remote deployment, cargo transportation, and regional penetration.