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

TL;DR

The paper tackles lane-change in dense traffic where inter-vehicle gaps are narrow, a scenario demanding cooperative behavior and real-time planning. It introduces NNMPC, a two-stage framework that couples MPC-based trajectory planning with SGAN-based predictions of surrounding vehicles, augmented by an adaptive safety boundary and sensor-noise mitigation via a Kalman Filter. Key contributions include pre-computed driving-intention trajectories, an SGAN-driven interactive motion-prediction module, a receding-horizon optimization that enforces safety under prediction errors, and a complete algorithm with guarantees for recursive feasibility. The approach yields 100% success across tested scenarios, reduces merge time by about 27% relative to baselines, and maintains smooth lateral maneuvers, demonstrating practical potential for dense-traffic autonomous driving in real-time CARLA simulations. The work advances cooperation-aware motion planning by integrating neural predictions with optimization-based control and by explicitly accounting for prediction errors and sensing noise, paving the way for safer, more comfortable autonomous lane changes in challenging traffic conditions.

Abstract

This paper presents an online smooth-path lane-change control framework. We focus on dense traffic where inter-vehicle 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 lane-change 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.

Figures (13)

• Figure 1: The autonomous-driving vehicle (in green) intends to change lanes, within a restricted merging area. The traffic is dense with narrow inter-vehicle intervals that are spatially insufficient for a vehicle to merge into. The autonomous-driving vehicle would get stuck in the merging area, unless other drivers slow down to make space for the vehicle.
• Figure 2: Diagram for two-stage planning and control framework. The trajectory planner in the upper layer generates waypoints, a sequence of positions coupled with velocities. The trajectory tracking controller in the lower layer determines a pair of steering angle and throttle/braking. The trajectory planner evaluates cooperative behaviors conditioned on possible choice of trajectory.
• Figure 3: Diagram of the planning framework with a Recurrent Neural network, SGAN. The optimization is solved with respect to accelerations and steering angles which are then converted to waypoints that the lower level controller in Fig. \ref{['fig:diag-planning-control']} tracks.
• Figure 4: Examples of driving intentions.
• Figure 5: Vehicle shapes are modeled by three circles. Note that multiple (more than three) circles of various radius can be applied to any vehicle shape and size. The minimum distance $g$ is computed as the minimum Euclidean distance between any pair of circle, as in Eqn. \ref{['eq:const_circle']}

Theorems & Definitions (1)

• Proposition 1