Bayesian and Convex Optimization of Racing Line and MPC Path Following Control for 4WIS&4WID Autonomous Vehicle

This paper focuses on the racing line optimization of a 4-wheel independent steer (4WIS) and 4-wheel independent drive (4WID) vehicle. And the minimum lap time problem, can be reformed as finding the minimum of a nonlinear function
$f(\alpha)$
, where
$\alpha$
is a vector representing the geometry of a possible racing line, and
$f(\alpha)$
is its minimum lap time. To balance the trade-off between computation efficiency and vehicle model precision, a two-stage optimization structure is adopted, where the preliminary stage uses genetic algorithm (GA) to narrows down the range of racing line with a simplified vehicle model and the efficient “three-step” method. And with the results from GA, the second stage uses Bayesian optimization for a precise racing line, where the cost function
$f(\alpha)$
, specifically the minimum lap time, is produced by a convex optimization function with a detailed 4WIS&4WID vehicle model considering the torque and steering of each wheel. In addition, to validate the optimized racing line and lap time, a path tracking model predictive controller (MPC), with 4WIS&4WID strategies is built for CarSim evaluation. Results have shown that compared to using Bayesian directly(>1000), our two-stage structure has better efficiency with fewer iterations(about 200). And the proposed method can better estimate the actual lap time, where the lap time gap between optimization and CarSim is below 3%, meanwhile the optimized wheel forces are also generally consistent with simulation.