High-performance real-time model predictive control via operator splitting for embedded applications

Many industrial control systems must operate under very short sampling times and on embedded hardware with limited memory and computational power. This creates a practical barrier to the deployment of model predictive control (MPC), even when MPC could significantly improve performance and help manage operating limits. This paper presents a deterministic real-time MPC framework for such settings with explicit iteration and stability guarantees. By combining relaxed recentered logarithmic barrier (RRLB) modeling with a structure-exploiting operator-splitting scheme, the approach enables a fixed-metric primal–dual iteration with provable global linear convergence and an explicit bound on the number of iterations required to ensure closed-loop stability. In contrast to conventional methods that rely on asymptotic convergence, the proposed framework guarantees predictable computational effort, recursive feasibility, and bounded constraint violations under fixed computational budgets. As a result, it is directly applicable to real-time embedded control on resource-constrained hardware. Numerical results demonstrate faster convergence and improved control performance compared to ADMM-based approaches, while embedded hardware experiments confirm deterministic runtimes and reliable real-time behavior, highlighting its practicality for real-world embedded applications such as robotics and autonomous systems.