SEAE: Stable end-to-end autonomous driving using event-triggered attention and exploration-driven deep reinforcement learning

Abstract

In self-driving cars, significant losses can be caused by various unstable factors. Thus, the use of the reinforcement learning self-driving technology with stability constraints is essential. The proposed multi-input stable autonomous driving based on exploration-driven with attention and event-triggered (SEAE) helps the agent program better autonomous driving with stability. This paper optimizes the input information processing of deep reinforcement learning using the multi-head self-attention mechanism, enhances the spatial exploration ability of the agent using the exploration-driven network. It combines the acceleration stability with the event-triggered mechanism to ensure a high driving safety while taking the driving stability into account. More precisely, the proposed multi-input approach treats the instantaneous acceleration as a constraint specified by the agent and optimizes the reward function, while taking into consideration the rate of motion change. Weights are then assigned to the data sequences through a multi-head self-attention mechanism, allowing the agent to focus on the environmental information part that is more important for the autonomous driving task received by the sensors. In addition, the proposed multi-input SEAE method is compatible with SAC and DDPG algorithms to verify its effectiveness in driving stability. The obtained results show that the proposed method has the highest performance in average reward value, average episode length, driving speed and driving stability in complex scenarios of autonomous driving tasks.