Fast Collision-Free Multi-Vehicle Lane Change Motion Planning and Control Framework in Uncertain Environments

In this article, we focus on the design, test and validation of a hierarchical control framework capable of optimizing lane change trajectories and steering the motion of multiple automated guided vehicles (AGVs) in an uncertain environment. In the upper-level maneuver planning phase, a convex feasible set-based real-time optimization algorithm is adopted to plan the optimal motion trajectories for AGVs. The main novelty of this approach lies in its optimization process, where a sequence of convex feasible sets around the current solution is iteratively constructed such that the nonconvex collision avoidance constraints can be approximated. Subsequently, an improved sequential convex programming (SCP) algorithm is designed and applied to reshape the current maneuver trajectory in the preconstructed convex feasible sets and reduce the error caused by successive linearization of vehicle kinematics and constraints. The planned lane change trajectories are then provided to the lower-level motion controller, where a deep reinforcement learning (DRL)-based collision-free tracking control method is established and applied onboard to produce the control commands. This approach has the capability to deal with unexpected obstacles (e.g., those that suddenly appear around the vehicle). The proposed training method integrates a consensus algorithm with actor-critic deep reinforcement learning to allow multiagent training to achieve faster training speed and improved performance compared with single-agent training. The feasibility and effectiveness of the proposed design are verified by carrying out simulation case studies. Moreover, the validity of the designed hierarchical control framework is further confirmed by executing hardware-in-the-loop tests.