摘要
自动泊车运动规划需满足安全性、舒适性、最终泊车位姿等多目标最优.提出一种基于模型的强化学习运动规划方法,以最大限度摆脱人类泊车经验,并综合考虑上述需求.建立了用于逼近实车的仿真模型;构建了基于加速度和距离控制的纵向策略;基于蒙特卡洛树搜索和神经网络,结合构建的纵向策略强化学习,最终收敛得到最优的泊车策略,迭代过程中的奖励函数综合考虑安全性、舒适性及最终泊车位姿等因素;通过实车实验对获得的泊车策略进行了验证.结果表明,规划策略能够满足对安全性、舒适性、最终泊车位姿等多目标最优的需求.
The multi-objective optimization including safety,comfort and final parking performance should be considered in the automatic parking motion planning.A model-based reinforcement learning motion planning method that tries to get rid of the human parking experience is proposed in this paper.The model for simulating the real vehicle is established.A longitudinal strategy based on acceleration and distance control is constructed.The parking strategy finally converges to the optimal based on Monte Carlo tree search and neural network,combined with the longitudinal strategy for reinforcement learning.The multi-objective including safety,comfort and final parking performance are taken into account in the iterative process by the reward function.The convergence strategy is verified by field tests.The results show that the designed planning strategy can meet the demands of multi-objective optimization including safety,comfort and final parking performance.
作者
张继仁
陈慧
宋绍禹
胡峰伟
ZHANG Jiren;CHEN Hui;SONG Shaoyu;HU Fengwei(School of Automotive Studies,Tongji University,Shanghai 201804,China)
出处
《同济大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2019年第S01期186-190,共5页
Journal of Tongji University:Natural Science
关键词
自动泊车
运动规划
强化学习
蒙特卡洛树搜索
神经网络
automatic parking
motion planning
reinforcement learning
Monte Carlo tree search
neural network
