Sequential Convex Programming based Model Predictive Control for Unprotected Left-Turn

In autonomous driving, an unprotected left turn is a highly challenging scenario. It refers to the situation where there is no dedicated traffic signal controlling left turns; instead, left-turning vehicles rely on the same traffic signal as the through traffic. This presents a significant challenge, as left-turning vehicles may encounter oncoming traffic with high speeds and pedestrians crossing against red lights. To address this issue, we propose a Model Predictive Control (MPC) framework to predict high-quality future trajectories, while utilizing Sequential Convex Optimization (SCP) to approximate the original problem. Optimization-based trajectory planning methods can generate high-quality trajectories, but they are characterized by their non-convexity of obstacle avoidance constraints. In this paper, the proposed model predictive control architecture is based on a bicycle model which is nonconvex, our way to solve it is to employ SCP to approximate the original non-convex problem near certain initial solutions. Our method performs well in the comparison with the widely used sampling-based planning methods.