Safety-Aware Optimal Control With Language-Guided Online Parameter Adjustment via Large Language Models

In this paper, we present LaMPC-CBF, a language-guided control framework that integrates model predictive control (MPC) with a control barrier function (CBF) to generate provably safe trajectories from free-form user instructions for robotic systems. Large language models translate natural language, including varying levels of user-expressed safety intent, into optimization objectives and CBF-based safety constraints. In particular, the expressed safety intent is mapped to numerical safety parameters within the CBF constraints, which directly govern the system’s obstacle avoidance behavior. This automated parameterization bridges the gap between abstract human linguistic expressions and rigorous mathematical safety guarantees, ensuring that the robot’s behavior aligns precisely with the user’s perceived safety preferences. Additionally, LaMPC-CBF supports interactive user feedback during task execution. This allows users to redefine safety intents via natural language on the fly, eliminating the need for controller recompilation. This approach allows non-experts to adjust parameters without manual tuning by experts easily and enables online adaptation to evolving user’s safety intent. Simulation results demonstrate that LaMPC-CBF improves safety and adaptability compared to the baseline, particularly in dynamic obstacle scenarios.