Autonomous lateral maneuvers for self-driving vehicles in complex traffic environment

In recent years, there has been a growing interest in the field of autonomous driving, attracting attention from both the academic community and industry sectors. This thesis primarily focuses on autonomous lateral maneuvers of self-driving vehicle in complex environment with uncertain parameters. It aims to provide a holistic understanding of the challenges and solutions associated with ensuring safe and efficient vehicle operation under a variety of conditions. The research is divided into two interconnected sections.

The first section of the study introduces a method based on Model Predictive Control (MPC) for generating safe and viable trajectories by including extra safety space in the decision making function and the state constraints for the ego vehicle to account for uncertain state measurement of the surrounding vehicles. This method allows the ego vehicle to execute various lateral maneuvers. The proposed methodology optimizes control inputs to navigate the ego vehicle in mixed traffic environments, which consist of both human-controlled and autonomous vehicles. The Extended Kalman Filter is used to estimated the states of the surrounding vehicles based on noisy measurements. The constraints has been set to take into account the uncertainties of surrounding vehicle by providing extra safety space for the maneuver. A novel reference speed function is incorporated in the proposed algorithm to adjust the position of the ego vehicle based on the estimated states of the surrounding vehicles before initiating any lateral maneuvers. Once a sufficient gap is identified, the system employs a lateral maneuver initiation function that includes a built-in threshold function. This ensures the safe execution of the maneuver and grants the ego vehicle autonomous control over initiating or ceasing lateral movements. In order to ensure accurate simulation, the dynamic bicycle model is augmented with a nonlinear tire model, which represents the ego vehicle’s actual behavior. To reduce the computational burden of the nonlinear MPC, a Linear Time Varying Model Predictive Controller (LTV-MPC) is employed to enable realtime execution of the algorithm. The simulation results demonstrate that diverse maneuvers (i.e., lane merge, lane change, overtaking and lane departure) can be accomplished in realtime without the need for controller re-tuning or parameter modifications.

The second section of the thesis aims to develop adaptive Model Predictive Controller (MPC) with parameter estimator to help the autonomous vehicle to navigate road sections with unknown road profiles including unknown road grades and adhesion coefficients. Initially, a modification to the improved gradient method is proposed as the parameter estimator for the adaptive MPC. This modification allows the improved gradient method to adapt to unknown road grades within the nonlinear dynamic bicycle model when the state measurements are accurate. However, real-world traffic scenarios present a challenge due to noisy state measurement. Such measurement noise can be non-Gaussian due to external disturbance, sensor limitations and network or communication issues and cause the state measurement to be non-zero bias with random walk errors or temperature induced errors. Thus the particle filter is proposed to estimate the uncertain parameters in the vehicle model due to its ability to handle non-Gaussian noise instead of the traditional filters such as the Kalman Filter (KF), Extended Kalman Filter (EKF) or Unscented Kalman Filter (UKF) which all require the measurement noise to satisfy the Gaussian distribution. With non-Gassian measurement noise, the proposed particle filter not only provides superior parameter estimation for the unknown road grade when compared to Extended Kalman Filter, but it also outperforms Unscented Kalman Filter in estimating the unknown road adhesion coefficient. To accelerate the convergence process of the unknown parameters estimated by the previous particle filter based adaptation law, a Dual MPC framework is proposed and evaluated in two different driving scenarios. This framework, unlike the other adaptive MPC, actively engages in environment exploration to determine control actions, rather than relying solely on passive parameter estimation. The proposed Dual MPC approach utilizes multiple prediction models, each with different system parameters determined by the distribution of unknown par meters. The exploitation aspect of Dual MPC is demonstrated by using the average values of state variables from various prediction systems to track a given reference. On the other hand, the exploration aspect is demonstrated by adding the state estimation covariance matrix to reduce the parameter uncertainty. The dual effects of exploration and exploitation in the control strategy not only accelerate the convergence of estimated parameters but also enhance the target tracking performance. Simulation results illustrate the superiority of the proposed Dual MPC over a traditional MPC paired with Unscented Kalman Filter tracking a race track with unknown road grades. Furthermore, it surpasses the previous particle filter-based adaptive MPC for tracking an 8-shaped track when unknown road adhesion coefficients are involved.