The challenges of model mismatch, system noise, and lag at high speeds make it difficult to ensure that a distance-constrained Model Predictive Controller (MPC-DC) system remains recursively feasible while executing obstacle avoidance maneuvers at high speed. Integrating Control Barrier Functions (CBF) into the MPC framework (MPC-CBF) significantly improves upon MPC-DC. Unlike MPC-DC, MPC-CBF proactively maintains safety without necessitating the intersection of the reachable set with obstacles, thereby enhancing safety and reducing computational complexity by eliminating the need for a large optimization horizon. Despite the considerable improvement with MPC-CBF, there are cases of optimization failure and suboptimal behavior. The reason may be an overly simplified dynamic model without adequate system identification and disturbance modeling, leading to large error propagation at the limits of handling. This study explores the development of a residual model to compensate for model mismatches and disturbances. A simplified dynamic bicycle model is used as the nominal model. The residual model, a Multi-Layer Perceptron (MLP), is designed to predict the difference between the optimal system output and the prediction from the nominal model. The learned residual model is incorporated into the baseline MPC and MPC-CBF formulations for comparison. We test our algorithm in the CARLA simulator with ACADOS as the solver framework. We train the residual model using the model-free RL algorithm known as Proximal Policy Optimization (PPO). Our findings demonstrate that incorporating the learned residual model into the MPC formulation enhances the stability and feasibility of the controller, resulting in improved overtaking maneuvers at high speeds. |