Intelligent Vehicle High Real-Time Predictive Control For Diverse Reference Trajectories

Motion control is a critical factor in determining the driving performance of intelligent vehicles.While Model Predictive Control(MPC)is an effective method for trajectory tracking control,its real-time performance is poor.Approximate dynamic programming(ADP)can solve real-time issues by training neural networks to optimize control policy for online applications.However,it suffers from slow offline training and is difficult to apply to diverse traffic scenarios.This paper proposes a time-splitting iterative framework based on the optimal relaxation coefficient,proposes a time-splitting based approximate dynamic programming training method.Furthermore,a trajectory tracking algorithm for diverse reference trajectories is developed,providing a foundation for achieving high real-time trajectory tracking control for autonomous vehicles.Firstly,to address the issue of slow convergence in time-splitting for optimal control problems,a time-splitting iterative framework based on the optimal relaxation coefficient is proposed.By utilizing the optimality conditions of the alternating direction multiplier method,a range of relaxation coefficient values that satisfy convergence requirements is obtained.Additionally,a two-stage selection criteria is established for the relationship between the relaxation factor and convergence rate.Furthermore,by combining the mathematical form of the optimal control problem with a single optimization variable,a positive correlation between the relaxation coefficient and the iteration efficiency is identified,leading to the derivation of the optimal relaxation coefficient.The proposed framework with the optimal relaxation coefficient can reduce up to 63.7%of the iteration times,compared to existing time-splitting iterative frameworks.Aiming to address the challenge of long computation time for policy approximation in long-horizon optimal control problems,a parallelizable timesplitting based approximate dynamic programming method is proposed.The method introduces a consistency variable to decouple the system state and splits the original problem into multiple sub-problems along the prediction horizon.Redundant states are designed as network inputs to ensure the independence between sub-problems.Combined with the time-splitting framework with the optimal relaxation coefficient,a parallel training mechanism for the policy network parameters is established.Simulation results demonstrate that for long-horizon problems,the training time of the policy network using this approach was reduced by 83.6%.Furthermore,for the motion control of vehicles in complex road scenarios,a tracking control strategy suitable for diverse reference trajectories is designed.By designing four typical reference trajectories and two reference speeds,the state space of diverse traffic scenarios is approximated.The simulation results using Baidu Apollo simulation shows that the policy network achieved continuous trajectory tracking tasks such as straight driving,lane changing,and U-turn.Compared with model predictive control,the speed tracking error was reduced by 0.017 m/s,the trajectory tracking error was increased by 0.9 cm,and the decision efficiency was improved by 153.1 times,and the cumulative tracking loss was smaller.In summary,under the premise of comparable tracking performance,the ADP algorithm greatly improves the control real-time performance.Finally,relying on the Baidu’s Apollo real vehicle platform,a policy network for tracking diversified reference trajectories was deployed to complete trajectory tracking and control experiments on a circular road in a campus scenario.The experimental results demonstrate that the learned policy network can smoothly complete diversified trajectory tracking and control tasks,such as straight driving and turning.Compared with MPC,the trajectory tracking error is reduced by 2.1cm,the decision efficiency is increased by 65.8 times,and the speed tracking error only increases by 0.021m/s.This further proves that the proposed approach greatly improves online control real-time performance without sacrificing tracking performance.