Computing controllable sets for safe and stable automated vehicles

Abstract

Adaptive Cruise Control (ACC) is an important driver-assistance function in automated and semi-automated vehicles. A well-designed ACC system can improve longitudinal safety, traffic flow, and driving comfort by controlling the ego vehicle in response to the motion of a preceding vehicle. Maintaining safety becomes particularly challenging when the lead vehicle behaves unexpectedly or when rapid braking is required. This thesis presents a framework that combines Model Predictive Control (MPC) and Active Learning (AL) for longitudinal safety analysis. The MPC controller computes ego-vehicle acceleration commands while accounting for the system dynamics, input limitations, and safety constraints. A more detailed plant model is used to introduce effects such as actuator dynamics, aerodynamic drag, rolling resistance, road grade, and reaction delay. In addition to controlling the vehicle, MPC is used as a feasibility oracle to classify operating conditions as feasible or infeasible. AL then selects informative states, particularly near the transition between these regions, thereby reducing the number of MPC evaluations required to approximate the safety boundary. Simulation results indicate that the proposed framework can identify safety-critical operating conditions and approximate the MPC-feasible region while concentrating computational effort near the feasibility boundary. The results also illustrate the influence of model mismatch and critical braking conditions on the estimated safety region.