홈 › Papers

SMART: A Decision-Making Framework with Multi-modality Fusion for Autonomous Driving Based on Reinforcement Learning

Decision-making in autonomous driving is an emerging technology that has rapid progress over the last decade. In single-lane scenarios, autonomous vehicles should simultaneously optimize their velocity decisions and steering angle decisions to achieve safety, efficiency, comfort, small impacts on rear vehicles, and small offsets to the lane center line. Previous studies, however, have typically optimized these two decisions separately, ignoring the potential relationship between them. In this work, we propose a decision-making framework, named SMART (deciSion-Making frAmework based on ReinforcemenT learning), to optimize the velocity and steering angle of the autonomous vehicle in parallel. In order for the autonomous vehicle to effectively perceive the curvature of the lane and interactions with other vehicles, we adopt a graph attention mechanism to extract and fuse the features from different modalities (i.e., sensor-collected vehicle states and camera-collected lane information). Then a hybrid reward function takes into account aspects of safety, efficiency, comfort, impact, and lane centering to instruct the autonomous vehicle to make optimal decisions. Furthermore, our framework enables the autonomous vehicle to adaptively choose the duration of an action, which helps the autonomous vehicle pursue higher reward values. Extensive experiments evidence that SMART significantly outperforms the existing methods in multiple metrics.

Code (1)

Tasks

Similar Papers 제목 키워드 기반