交互合作式智能汽车决策规划方法研究

Intelligent vehicles will be the develop trend and core of the future transportation system. And the “brain” of an intelligent vehicle, the decision making and planning algorithm, will determine whether an intelligent vehicle can make reasonable behaviors, which is related to the landing process. However, how to design a decision making and planning algorithm adapt to complex traffic and capable of being understood by various traffic participants still faces serious challenges. To this end, the project conducts research on interactive and cooperative decision making and planning methods for intelligent vehicles which including the followings. The coupling mechanism of "human-vehicle-road" is studied, and based on this, the intention recognition model of each type of traffic participants is established. A hierarchical decision making method is proposed. In the long-term decision making layer, the long-term behavior decision making algorithm is obtained based on deep reinforcement learning. In the short-term decision making layer, short-term action decision making algorithm is obtained through continuous interaction and cooperation with other traffic participants. The research of this project will provide new theoretical methods and technical support for the unresolved anthropomorphic decision-making and planning problem in the world, and accelerate the development and landing of intelligent vehicles.

智能汽车将是未来交通的发展趋势和核心，而智能汽车的“大脑”，决策规划算法，将决定着智能汽车是否能做出合理的行为动作，关系着智能汽车的发展与落地进程。然而如何做出适应复杂交通的、又能够被各个交通参与者理解的决策规划算法却仍然面临着严峻的挑战。为此，本项目针对交互合作式智能汽车决策规划方法开展研究，内容包括：对“人-车-路”耦合机理进行研究，并基于此建立各类型交通参与者的意图识别模型；提出分层的决策规划方法，在长期决策层，基于深度强化学习的方法得到长期的目标行为决策；在短期决策层，通过不断的与其他交通参与者进行交互与合作，在协作与博弈中得到拟人化的短期动作决策规划。本项目研究将为国际上尚未解决的智能汽车拟人化决策规划技术提供新的理论方法和技术支撑，加速智能汽车的发展与落地。

决策规划是智能汽车核心算法的重要组成部分之一，它决定着智能汽车将做出何种动作、并沿着什么样的轨迹行驶。然而，面对复杂的、充满不确定性的交通场景，如何做出合理准确的动作决策以及实时最优的运动规划仍然面临着严峻的挑战。本项目首先对“人-车-路”耦合机理进行研究，提出了基于动态安全场的交通环境评估方法，该方法能够将交通场景中不同交通参与者、道路结构、行为预测等因素对智能汽车的影响进行统一量化建模，解决了交通环境要素多带来的决策规划算法复杂性大的问题，为决策规划算法的有效性和实时性提供基础。然后进行智能汽车长期目标行为决策研究，提出了基于动态安全场和深度强化学习的智能汽车行为决策方法，该方法能够在仿真环境中通过不断的自我学习改进策略，实现在不同复杂场景下的有效决策。最后进行智能汽车短期运动规划研究，提出了基于动态安全场和闭环动力学的运动规划方法，解决了智能汽车规划路径与实际轨迹一致性差的问题，实现了实时最优的智能汽车运动规划。本项目研究提出的交互合作式智能汽车决策规划方法可满足鲁棒性和实时性的要求,并适应复杂交通流，为加速智能汽车的发展与落地提供助力。