A comparative analysis of two model predictive control architectures for autonomous driving systems

In response to the growing demand for reliable and efficient autonomous driving systems, NMPC has emerged as a prominent control strategy, capable of handling both linear and nonlinear constraints while predicting future states. The objective of this study is to develop and evaluate two distinct NMPC architectures for lane centering and obstacle avoidance. The key difference between the two architectures is that in the base one, both path planning and control computation are performed by the same NMPC, whereas in the double-layer system the higher layer handles path planning while the lower layer computes the control inputs. The decision to use Model Predictive Control for path planning stems from its model-based nature, this ensures that the generated path is not only feasible but also dynamically consistent with the vehicle’s capabilities. Therefore, detailed vehicle models, including a four-wheeled model, a two-wheeled model, and a point mass model are developed starting from the vehicle dynamics and then used in simulations. The two NMPC architectures are assessed for their efficacy in lane centering and managing both static and dynamic obstacle avoidance. Performance is evaluated in terms of computational efficiency, driving comfort, and overall control effectiveness. The results demonstrate that while both architectures effectively manage vehicle control, the double-layer architecture exhibits superior performance, it is faster, produces smoother trajectories, and performs better at higher velocities compared to the single-layer architecture.