【摘要】 自动驾驶汽车的驾驶环境是一个复杂的动态系统，所以准确预测其他交通参与者的运动状态，是安全规划自车轨迹的重要前提。当前主流方法均采用先预测后规划的框架，这导致交互场景处理不充分。本文提出一个耦合预测规划的框架，框架联合微分博弈规划和层次博弈论预测，在复杂网络风险认知的指导下，进行整体的学习训练过程。利用复杂网络理论建模车辆间的动态交互，得到车辆交互的风险认知函数。建立一个基于Transformer的复杂网络编码器与利用Transformer和层次博弈论相结合的交互解码器，迭代完善预测周边认知域智能体的未来运动轨迹，规划的效用函数和预测轨迹、车辆初步规划一起迭代实现预测规划的耦合优化。最后基于nuPlan数据集进行验证实验，结果表明本文所提方法在开环和闭环测试的各项指标如舒适度、速度等指标都有明显提升且相比于基于学习还是基于规则的方法都能够表现出优势。更多还原

【Abstract】 The driving environment for autonomous vehicles is a complex dynamic system. Accurately predicting the motion states of other traffic participants is a crucial prerequisite for safely planning the ego vehicle’s trajectory. Current mainstream methods all adopt a "predict first, then plan" framework, which leads to insufficient handling of interactive scenarios. In this paper a coupled prediction-planning framework is proposed that combines differential game planning and hierarchical game-theoretic prediction, guided by complex network risk cognition for integrated learning. Complex network theory is used to model dynamic interaction between vehicles to derive a risk cognition function for vehicle interaction. A Transformer-based complex network encoder and an interaction decoder that combines Transformer with hierarchical game theory are constructed to iteratively refine the predicted future trajectories of surrounding cognitive-domain agents. The planning’s utility function co-iterates with predicted trajectories and preliminary vehicle planning to achieve coupled prediction-planning optimization. Validation experiments on the nuPlan dataset show the proposed method significantly improves various metrics in both open-loop and closedloop tests, including comfort and speed, outperforming both learning-based and rule-based approaches.更多还原