智能汽车主动避撞工况的高实时预测控制

High real-time predictive control for active collision avoidance of intelligent vehicles

为满足复杂交通场景下智能汽车轨迹跟踪避撞控制的高实时性要求,该文采用了一种循环模型预测控制算法(RMPC)将在线优化问题转化为循环策略参数的离线求解,并进行了仿真试验。根据车辆主动避撞的约束条件,引入惩罚函数将约束型主动避撞优化控制问题转化为无约束有限时域最优控制问题;进而利用循环函数逼近得到不同预测步长控制问题的最优解;最后将算法部署到原型控制器,结合CarSim平台验证了算法的避撞性能以及在线计算的高效性。结果表明:预测步数从12增加到20步,避撞过程最小车距由0.34 m提升至1.38 m,千次实验碰撞次数由44下降到0;与常用在线优化求解器相比,该算法在预测步数为15时,其计算效率提升超过5.6倍。

A recurrent model predictive control(RMPC)algorithm was adopted to meet the high real-time requirements for active collision avoidance control in complex traffic scenarios.RMPC transformed the online optimization problem into an offline solution of recurrent policy parameters,and simulation experiments were conducted to validate its effectiveness.By introducing a penalty function,the constrained active collision avoidance optimization control problem was formulated as an unconstrained finite-time optimal control problem.The optimal solution of the control problem with different prediction steps was represented by a recurrent neural network.The learned policy to the prototype controller was deployed and its collision avoidance performance and online computational efficiency was verified by using CarSim.The results shows that the minimum vehicle distance during the collision process increases from 0.34 m to 1.38 m with the prediction step increasing from 12 to 20,and the number of collisions in a thousand experiments decreases from 44 to 0.Furthermore,the algorithm demonstrates a computational efficiency improvement of over 5.6 times compared to commonly used online optimization solvers when the prediction step is 15.