Agentic MPC framework lets control systems understand natural language

Traditional model predictive control (MPC) excels at handling structured, low-level specifications but cannot dynamically incorporate high-level context such as social norms, user intent, or natural language instructions. To bridge this gap, researchers Yuya Miyaoka and Masaki Inoue (affiliated with Japanese institutions) propose Agentic MPC—a framework that combines MPC with large language model (LLM)-based agents. The agent interprets heterogeneous inputs including natural language messages, environmental observations, and external knowledge, then automatically resynthesizes the control specifications to align with the current context. This enables control systems to adapt on the fly without manual reprogramming.

The framework is validated in an autonomous driving scenario. Specifically, the system can respond to social situations like yielding to an emergency vehicle—a task that requires understanding both the visual scene and the unwritten rule of giving way. It also adjusts behavior based on personal preferences conveyed via natural language (e.g., “drive more cautiously”). The paper (7 pages, 5 figures) positions Agentic MPC as a step toward semantic control resynthesis, where high-level knowledge seamlessly shapes low-level control actions. While the demonstration focuses on driving, the approach generalizes to any domain where control systems must incorporate rich, contextual information.