Safety-Regulated Reinforcement Learning Compensated Model Predictive Control for Ground Vehicles with Unknown Disturbances

This paper proposes an MPC-RL-CBF control framework that leverages the individual strengths of MPC (Model Predictive Control) schemes and Deep RL (Reinforcement Learning) techniques. This allows using a model mismatched computationally inexpensive optimal controller with a compensating learning technique to handle the uncertainties in system dynamics and unknown external disturbances. The controller is evaluated in simulation for a vehicle tracking a path with a lane change, subjected to unknown crosswinds. The results show that the MPC-RL-CBF approach helps track the path better than the purely model-based approach and does so safely, through safety guided training. This framework can be extended to off-road driving controls under changing terrain types and properties, tire-terrain interaction behavior, steep slopes etc.