Trajectory Planning And Control For Autonomous Vehicle Under Uncertain Environments

The autonomous vehicle may deviate from trajectories or even become unstable due to uncertain environments in different working condition.To improve the safety,the robustness and the stability under uncertainties,this reseach focuses on trajectory tracking,trajectory planning and decision making considering uncertainties.This research proposes an automated framework considering uncertain environments to achieve the stability,the driving comfort,the safety and the high efficiency,while decision making in this framework is utilized to navigate trajectory planning.This paper proposes a robust trajectory tracking at driving limits in the presence of hazardous road conditions,parameter mismatch and poor stability at driving limits which result in large tracking errors.Based on a vehicle-road model and the analysis of steady state errors,a feasibility analysis of a robust model predictive control(MPC)is conducted against environmental uncertainties.Robust MPC is proposed which attenates disturbances of road conditions and parameter mismatch.Online and efficient approximating techniques are proposed for a robust positively invariant set to accelerate computing for robust MPC.To ensure the stability and the speed advantage at driving limits,the stability control is proposed consisting of Direct Yaw Moment Control(DYC)and Preview G-vectoring Control(PGVC).The combination of the stability control and robust MPC makes it possible to optimize the tracking errors and the stability at the same time.Simulation results show that the neural network to approximate 30 steps reachable sets is 400% faster than the support function.The lateral acceleration is 28% less than other methods at driving limits.In this proposed trajectory tracking,the lateral acceleration at low road friction coefficient is almost the same as other methods but the maximum front steering angle is about 5°.In order to enhance the driving comfort and the awareness of a map,this paper proposes a trajectory planning method considering the driving comfort and the vehicle stability.Double-tree Rapidly Exploring Random Tree(DT-RRT)is proposed for the smoothness of paths,dynamic obstacle detection and the real-time performance.To improve the sampling efficiency,original RRT returns a collision-free and piecewise path subsequently smoothened and selected as a reference path for the modified RRT.The optimal distance metric is designed to evaluate path quality and detect dynamic obstacles in a map,improving chances of trajectory planning in a dynamic environment.Optimizing path quality and updating optimal sampling region are proposed to improve path quality futher.Then a speed planning method combining the trapezoidal speed planning with PGVC is established considering the driving comfort and the vehicle stability.Simulation results show that the accumulated curvature change of the proposed Best Nearest function is 434% less than Euclidean distance along paths.The average of the accumulated curvature of DT-RRT is 33% less than that of RRT* and DT-RRT converages within 50 loops.Based on the proposed trajectory planning,the sideslip angle is 8% less at low road friction coefficient and it is 18% less at high road friction coefficient.To handle perception errors and uncertainties of traffic participants,prioritized experience replay(PER)in Double Deep Q-networks(DDQN)for reinforcement learning(RL)is proposed as decision making with scene decomposition and action uncertainty to navigate trajectory planning.Coupled longitudinal and lateral actions are designed to ensure a tight connection with trajectory planning.Reward function is designed and Markov decision process(MDP)model is built,which guarantees the traffic efficiency and the safety in a complicated environment.This research proposes integrate scene decomposition and action uncertainty into reinforcement learning,which scales the deep neural network to multiple traffic participants and transfers the uncertainty of perceptions to the uncertainty of actions to determine whether the risk is acceptable.DDQN with PER is utilized to replay transitions with high expected learning process more frequently and accelerate the training process.Simulation results show that the required time under proposed optimal policy to finish left turn at T junction is 4s less than that under conservative policy.The maximum sideslip angle reaches 9.7° during the left turn.Finally,experiments are conducted on a scaled vehicle testbed.Subsequently we compare the dynamics of the scaled vehicle with dynamic constraints.Then the working conditions for robust MPC and DT-RRT are designed.Experimental results show that,given the front steering angle,the measured value of lateral speed and yawrate are almost the same as the estimated value,which means that the scaled vehicle meets the assumpution of the dynamic model.Robust MPC can attenuate external disturbances while the lateral error of robust MPC is 7cm less than that of MPC.The real-time performance is enhanced with the parallel computing that the updating frequency increases from 1Hz to 5Hz.As the path quality of DT-RRT is guaranteed,the vehicle can follow paths generated by DT-RRT and the actual path is smooth.Static obstacle and dynamic obstacle are achieved while the maximum lateral error remains within 26 cm under obstacle avoidance.The autonomous driving platform is introduced to further investigate the feasibility of DT-RRT.When the longitudinal speed increases to 2m/s,the maximum lateral error is about 12 cm,indicating that DT-RRT meets kinematic constraints of the scaled vehicle and autonomous driving platform.