Predictive racing control using model-mismatch learning for autonomous racecars

Autonomous racing remains an extremely challenging problem due to the strong nonlinearity of tire forces and unknown road conditions. The strong nonlinearity of tire forces and unknown road conditions result in model-mismatches, degrading the performance of model-based approaches. A predictive racing controller using model-mismatch learning is proposed to enhance the performance of autonomous racing. To balance model accuracy and complexity, a single hidden-layer fully neural network is employed to learn model mismatches, and the learned model mismatches are then incorporated directly into the prediction model of model predictive control(MPC). A CarSim/Simulink co-simulation test is conducted to verify the effectiveness of the proposed control scheme. Result shows a 3.36% improvement in the average lap speed of racing cars compared to the nonlinear MPC.