ChatMPC: a language-driven model predictive control framework for adaptive and personalized autonomous driving

Abstract Model Predictive Control (MPC) has emerged as one of the most widely adopted and effective approaches in autonomous driving systems. Conventional design methodology of MPC systems, however, often rely on static rule-based architectures and predetermined control strategies, limiting their flexibility and responsiveness to complex and dynamic traffic environments. To enhance the system’s understanding of driver intentions and improve strategy adaptability, this paper proposes a novel autonomous driving framework, ChatMPC, that integrates Natural Language Processing (NLP) with MPC. The framework employs a Transformer-based sentence embedding model, Sentence-BERT (SBERT), to parse driving intents embedded in natural language commands (e.g., “overtake,” “follow”), and dynamically updates the MPC controller’s objective functions and constraints. This enables the generation of personalized driving behaviors aligned with user preferences. Simulation experiments conducted on the Matlab platform show that ChatMPC completes the full cycle from instruction parsing to control optimization in an average of 15 seconds, with MPC prediction requiring an average of 13.5 ms and a worst-case time of 22.2 ms, well within the 50 ms real-time budget. In typical traffic scenarios, the system achieves high tracking accuracy, with a following error of 0.827% and overtaking error of 1.67%, validating its real-time performance and effectiveness.