Abstract
Despite great successes, model predictive control (MPC) relies on an accurate dynamical model and requires high onboard computational power, impeding its wider adoption in engineering systems, especially for nonlinear real-time systems with limited computation power. This paper develops a unified online data-driven predictive control pipeline to efficiently control a system with guaranteed safety without incurring large computational complexity. The framework learns not only the real system but also the control policy, resulting in a reasonable computational cost for data-driven predictive controllers. Specifically, we develop a spatial temporal filter (STF)-based concurrent learning scheme to identify system dynamics, a robust control barrier function (RCBF) for safety guarantees in the presence of model uncertainties, and an online policy correction scheme through perturbation analysis. Extensive simulations on cart-inverted pendulum and automotive powertrain control demonstrate the efficacy of the proposed framework, showing comparable performance with much lower computational cost than benchmark algorithms.
Introduction
Nonlinear Model Predictive Control (NMPC) is a powerful methodology for controlling complex systems while respecting operational constraints. However, its real-world deployment faces two significant hurdles: the requirement for an accurate mathematical model of system dynamics and the high computational cost of solving optimization problems at every time step. These challenges often exceed the capabilities of low-power onboard embedded systems used in robotics and automotive engineering.
This work introduces a unified framework designed to bridge this gap by learning both the system model and the control policy from data, ensuring that the resulting controller is both computationally efficient and robustly safe.
The Unified Framework
The proposed framework consists of a multi-stage pipeline that separates heavy computation from real-time execution. It integrates offline system identification and policy learning with an online adaptation layer for real-time deployment.
System Identification and Safety
To remove the need for manual modeling, the framework uses Spatial-Temporal Filters (STF) and a Concurrent Learning scheme to identify system dynamics. Unlike traditional methods that require artificial noise injection (Persistence of Excitation), this approach uses a history of recorded data to satisfy rank conditions, making it practical for real-time operation.
Safety is maintained through Robust Control Barrier Functions (RCBF). These functions are specifically designed to be robust against three types of uncertainty: system identification errors, control policy learning errors, and unknown external disturbances. This holistic approach ensures the system remains within safe boundaries even when the learned models are imperfect.
Policy Learning and Real-Time Performance
The core innovation for reducing computational overhead is the transition from online optimization to offline policy learning. Instead of solving a complex NMPC problem at every time step, the framework solves it offline for various scenarios and uses a function approximator to learn the resulting control policy.
By learning the mapping from states to optimal controls, the framework turns a heavy optimization task into a simple function evaluation. To mitigate any performance degradation from the learned policy, an Online Adaptive Correction layer is employed. This layer utilizes KKT sensitivity analysis and an ancillary feedback controller to keep the real system trajectory close to the ideal nominal path.
Key Results
The framework was evaluated on a cart-inverted pendulum and a high-fidelity turbocharged internal combustion engine simulator. Key findings include:
- Massive Speedup: In the automotive task, the proposed controller achieved a loop time of approximately 78.9 microseconds, compared to 25.2 milliseconds for standard NMPC—a speedup of over 300x.
- High Fidelity: The learned system models achieved over 96% accuracy, and the learned control policies matched ground-truth NMPC with over 98% accuracy.
- Guaranteed Safety: Simulations confirmed that the RCBF-based approach successfully maintained safety constraints (e.g., cart position bounds) even under external disturbances where uncorrected policies failed.
Conclusion
This unified framework demonstrates that data-driven methods can effectively bypass the computational bottlenecks of NMPC without sacrificing safety. By combining robust control theory with efficient policy learning, it provides a viable path for implementing advanced predictive control on resource-constrained hardware. Future research will focus on extending these safety guarantees to unbounded stochastic disturbances and deepening the theoretical analysis of recursive feasibility.
Summary updated