【摘要】 基于仿真场景的自动驾驶功能测试，能够有效地还原真实交通场景中车辆的行为和状态，从而更加准确地评估自动驾驶系统在复杂交通场景中的决策规划和控制能力。本文基于Triple-GAIL和GRU提出了一种GAIL-GRU多模态轨迹生成的自动驾驶仿真场景构建方法，首先从人类驾驶数据集中提取车辆行驶轨迹与多维度特征，然后根据车辆行为将行驶轨迹划分为直行、左换道、右换道3类行为轨迹并设置行为标签，其次构建车辆状态-行为标签-车辆动作对输入到GAIL-GRU模型，训练后形成具有拟人化驾驶行为的驾驶策略模型，最后将自动驾驶模型控制的自动驾驶车辆与驾驶策略模型控制的环境车辆部署于静态仿真环境当中，建立了多种仿真场景。在环境车辆变道切入场景与自动驾驶车辆变道切入的验证实验中，本方法成功复现了人类驾驶员在驾驶过程中采取的多种驾驶行为。实验结果表明，所提出的自动驾驶仿真场景构建方法构建的场景能够有效暴露自动驾驶算法在复杂交互情境中的决策缺陷，具有良好的拟人性和较高的风险覆盖度，通过对抗式模仿学习框架融合GRU单元可以解决多模态轨迹生成中的模式坍缩问题与时序信息丢失问题，实现了换道切入场景中车辆交互意图与运动连续性的耦合建模，突破了传统方法对交互博弈关系建模不足的局限。更多还原

【Abstract】 Simulation-based autonomous driving function testing can effectively reproduce vehicle behaviors and states in real-world traffic scenarios, thereby enabling more accurate evaluation of autonomous systems′ decision-making, planning, and control capabilities in complex environment. In this paper, a GAIL-GRU-based multimodal trajectory generation method for constructing autonomous driving simulation scenarios, integrating TripleGAIL and GRU is proposed. Firstly, vehicle trajectories and multidimensional features are extracted from a human driving dataset. The trajectories are then categorized into three behavioral types—lane keeping, left lane change, and right lane change—with corresponding behavior labels. Next, vehicle states, behavior labels, and vehicle actions are formed into input triplets for the GAIL-GRU model. After training, a driving policy model with human-like behavior is obtained. Finally, autonomous vehicles controlled by an autonomous driving model and environment vehicles driven by the learned policy model are deployed in a static simulation environment to construct diverse traffic scenarios. In validation experiments involving lane-change cut-in scenarios by both environment vehicles and autonomous vehicles, the proposed method successfully reproduces a variety of human driving behaviors. The experimental results show that the generated scenarios by the proposed autonomous driving simulation scenario construction method effectively expose decision-making flaws of autonomous driving algorithms in complex interactive situations, with strong human-likeness and high-risk coverage. Moreover, the integration of the GRU unit into the adversarial imitation learning framework addresses issues such as mode collapse and loss of temporal information in multimodal trajectory generation. This enables coupled modeling of interaction intention and motion continuity in lane-change cut-in scenarios, overcoming limitation of traditional methods in modeling interactive game dynamics.更多还原