A Model Predictive Trajectory Planning Framework for Autonomous Ground Vehicles on Structured and Unstructured Roads

Trajectory planning is one of the fundamental and core components of autonomous ground vehicles (AGVs). A sequence of movement states should be planned satisfying both kinematics constraints and dynamics constraints. This makes the problem computationally demanding as the planning space dimensionally increases. Moreover, there are massive structured and unstructured roads in urban scenarios. However, it is difficult to address the structured and unstructured trajectory planning problem simultaneously. In this paper, a model predictive trajectory planning framework is proposed to address the above challenges. Firstly, a nonlinear vehicle system model considering the kinematics and dynamics constraints is established. Secondly, a generator based on the deep neural network is introduced to output intermediate trajectory points towards the goal state. Then, a discriminator is trained utilizing the expert data to select the best solution for the extension. Finally, the model predictive controller (MPC) is leveraged to further optimize the trajectory to satisfy the kinematics and dynamics constraints. Simulation results demonstrate that the proposed algorithm can implement feasible, reasonable and human-like trajectory planning on both structured and unstructured roads. Moreover, the trajectory generated by the proposed planner satisfies the kinematics and dynamics constraints of vehicles.