GPU-Enabled Parallel Trajectory Optimization Framework for Safe Motion Planning of Autonomous Vehicles

This letter presents a GPU-enabled parallel trajectory optimization framework for model predictive control (MPC) in complex urban environments. It fuses the advantages of sampling-based MPC, which can cope with nonconvex costmaps through random sampling of trajectories, with the advantages of gradient-based MPC, which can generate smooth trajectories. In addition, we leverage a generalized safety-embedded MPC problem definition with a discrete barrier state (DBaS). The proposed framework has three steps: 1) a costmap builder to generate the barrier function map, 2) a seed trajectory generator to choose randomly generated trajectories to send to the optimizers, and 3) a batch trajectory optimizer to optimize each of the seed trajectories and select the best trajectory. Experiments with real-time simulations compare the effectiveness of the proposed framework, sampling-based MPC, and gradient-based MPC, which optimizes a single trajectory. The experiments also compare the application of two different control sequence sampling schemes to the proposed framework. The results show that the proposed framework performs gradient-based optimization but can plan a better trajectory even in complex environments by providing various initial guesses. We also show that the proposed framework can perform more accurate control actions than sampling-based MPC.