近年来，我国新能源汽车产业发展迅猛，智能化、网联化成为了汽车发展的趋势。自动驾驶是智能网联汽车的终极目标之一，其愿景是从方向盘上解放双手，根治车祸，提升出行安全水平。随着传感器、人工智能、计算机视觉等领域的突破，现有自动驾驶技术能解决多数驾驶场景，但已有自动驾驶技术基于单车智能，无法在发生遮挡时等危险场景进行有效感知与决策。车联网技术的发展为这一挑战提供了解决方案，它已经针对数个多智能体协同的自动驾驶场景做出了标准。以这些技术为基础的协同自动驾驶成为了解决自动驾驶长尾挑战，迈向更高安全水平的重要方案。自动驾驶技术有高度数据驱动的特征，但是协同自动驾驶商业化程度低，数据稀少；相比真实场景，基于仿真生成数据成本更低，且不会面临安全、伦理方面的风险。然而已有多智能体协同自动驾驶仿真平台或智能体数量较少，或支持任务较少，或未能完成自动驾驶的完整闭环，因此生成的数据仍未能满足协同自动驾驶研究的需求。本文从多智能体协同的自动驾驶仿真出发，设计实现多智能体协同自动驾驶仿真平台，基于平台生成协同自动驾驶数据集。具体成果包括以下两方面：1. 针对多智能体协同自动驾驶的仿真，设计并实现多智能体协同的自动驾驶仿真平台，仿真平台中可同时独立运行的智能体数量超过10 个，且各智能体均具有原始数据生成、感知、规划控制、安全性评估等自动驾驶的各主要模块，完成了自动驾驶的完整闭环。仿真平台为研究协同策略对自动驾驶安全性的影响提供了基础。2. 针对多智能体协同自动驾驶的数据，生成数据WHALES。它基于成果1 的仿真平台生成，平均每场景有8.4 个智能体，在已有数据集基础上支持额外的任务。本文以在候选智能体之间进行选择的调度任务为例进行实验，结果表明合理的调度策略可将平均精度均值最多可提升25% 以上，验证了数据集的有效性。该数据集是该领域目前智能体数量最多的数据集之一，为协同自动驾驶的实验提供了支撑。
In recent years, the new energy vehicle industry in China has experienced rapid growth, with intelligence, and connectivity emerging as the main trends. Autonomous driving is one of the ultimate goals of intelligent connected vehicles, with the vision of liberating hands from the steering wheel, eradicating traffic accidents, and promoting the level of safety in travel. With advancements in the realms of sensors, artificial intelligence and computer vision, contemporary autonomous driving technologies are capable of resolving the majority of routine driving situations. However, current autonomous driving technology on the basis of the stand-alone vehicular intelligence, encounters limitations in its perception and decision abilities during dangerous scenarios, such as occlusion. The progress in vehicle-to-everything technology offers a solution to these challenges with standards for several typical scenarios of multi-agent cooperation in autonomous driving. The cooperative approach to autonomous driving is emerging as a pivotal strategy to tackle the long-tail challenges and achieve higher levels of autonomous driving. Autonomous driving technology is highly data-driven. However, the commercialization of cooperative autonomous driving is at a nascent stage with a scarcity of data. In contrast to the complexities and risks associated with real-world data collection, synthetic data presents a cost-effective alternative without concerns related to safety and ethics. However, existing multi-agent cooperative autonomous driving simulation platforms often support a limited number of agents and tasks, or fail to complete the entire loop of autonomous driving. Data generated by the simulators above still does not meet the needs of cooperative autonomous driving research. This paper begins with the simulation of multi-agent collaborative autonomous driving. This paper designs and implements a simulation platform for multi-agent collaborative autonomous driving. Based on this platform, a collaborative autonomous driving dataset is generated. The specific achievements include the following two aspects:1. Aimed at the simulation of multi-agent cooperative autonomous driving, this paper designs and implements a simulation platform for multi-agent cooperative autonomous driving. The simulation platform is capable of operating over 10 intelligent agents independently and simultaneously, with each agent equipped with major modules of autonomous driving, including raw data generation, perception, planning，control, and safety assessment, thus completing the entire closed loop of autonomous driving. The simulation platform provides a foundation for studying the impact of cooperative strategies on the overall safety of autonomous driving.2. Aimed at the data for multi-agent cooperative autonomous driving, the WHALES dataset is generated. The dataset is generated by the simulation platform of Achievement 1, with an average of 8.4 agents per driving scenario, and supports additional tasks on the basis of existing datasets. This paper conducts experiments on cooperative driving with the task of scheduling among candidate agents as an example. The results show that a reasonable scheduling strategy can increase the average accuracy by more than 25%, therefore verifying the effectiveness of the dataset. The dataset is one of the datasets in autonomous driving with the largest number of agents, providing support for further experiments in multi-agent cooperative autonomous driving.
