Semantically parameterized MPC-CBF control for safety-critical obstacle avoidance in autonomous vehicles

Aligning autonomous-vehicle control with human intent remains challenging when safety-related constraints must be considered in closed-loop motion control. This paper proposes a Large Language Model (LLM)-assisted parameterization framework for motion control based on Model Predictive Control (MPC) and Control Barrier Function (CBF). Natural-language instructions are converted into structured semantic parameters that adjust CBF decay rates, while the execution layer solves an online semantically parameterized MPC–CBF problem for trajectory tracking and obstacle avoidance. Stationary and dynamic obstacles are handled through different discrete-time CBF update rules, and a direction-aware lateral incentive is included in the quadratic programming formulation. Parameter-associated differences in sampled clearance and tracking metrics were observed in the nominal co-simulation cases and repeated drive-by-wire trials under the evaluated operating conditions.