Lane-Change in Dense Traffic With Model Predictive Control and Neural Networks

Abstract

This paper presents an online smooth-path lanechange control framework. We focus on dense traffic where intervehicle space gaps are narrow, and cooperation with surrounding drivers is essential to achieve the lane-change maneuver. We propose a two-stage control framework that harmonizes Model Predictive Control (MPC) with Generative Adversarial Networks (GAN) by utilizing driving intentions to generate smooth lanechange maneuvers. To improve performance in practice, the system is augmented with an adaptive safety boundary and a Kalman Filter to mitigate sensor noise. Simulation studies are investigated in different levels of traffic density and cooperativeness of other drivers. The simulation results support the effectiveness, driving comfort, and safety of the proposed method.

Interaction-Aware Trajectory Planning for Autonomous Vehicles with Analytic Integration of Neural Networks into Model Predictive Control

“…Let • denote the Euclidean norm. For x < x ref , we utilize the following objective (cost) function J(∆(t), α(t), Z(t)) similar to [14]:…”

“…are the reference latitude coordinate of the desired lane and desired velocity, respectively, provided by a high-level planner [17]. For a detailed description of each term, we refer the interested readers to [14].…”

“…We compare ADMM-NNMPC with a baseline method called NNMPC [14] on the two-lane and three-lane scenarios by utilizing the same cost function (cost function coefficients listed in Table I) and T p = 8 time steps. NNMPC generates trajectory candidates by computing a finite set of spiral curves from the source lane to the target lane and selects the candidate with minimum cost.…”

“…However, the neural network integration and non-linear system dynamics make the optimization highly non-convex and challenging to solve analytically. Thus, prior efforts [6], [14], [9] that integrate neural network prediction into MPC are numerical in nature and rely on heuristic algorithms to generate a finite set of trajectory candidates. In [6] and [14], the authors generate these candidates by random sampling of control trajectories, and by generating spiral curves from source to target lane, respectively.…”

“…Thus, prior efforts [6], [14], [9] that integrate neural network prediction into MPC are numerical in nature and rely on heuristic algorithms to generate a finite set of trajectory candidates. In [6] and [14], the authors generate these candidates by random sampling of control trajectories, and by generating spiral curves from source to target lane, respectively. In [9], the authors utilize a predefined set of reference trajectory candidates.…”