Model Predictive Control (MPC) is highly regarded for its ability to handle state and input constraints by solving an optimal control problem over a finite future horizon at each time step. In its classical formulation, MPC stabilizes a system toward a fixed, pre-determined target state. However, real-world applications demand that controllers track trajectories that change abruptly, are dynamically complex, or are physically impossible to achieve due to strict system constraints.
When a desired target suddenly shifts or violates physical boundaries, traditional MPC often suffers from a loss of recursive feasibility—the mathematical guarantee that if a valid control sequence exists now, one will exist at the next time step. If the target is unreachable, the optimization problem simply fails to find a solution.
To solve this, a powerful paradigm shift has emerged: decoupling the immediate feasibility of the control loop from the ultimate attainability of the external target. This is achieved by introducing an artificial reference (or artificial equilibrium) directly as an online decision variable within the optimization problem.
The Feasibility Trap in Standard Tracking
In standard MPC, the controller minimizes a cost function based on the error between the predicted states $x_k$ and a fixed external reference $r$. To guarantee stability, classical MPC often requires the predicted state at the end of the horizon (the terminal state) to land exactly on the reference or within a tight "terminal set" centered around it.
Standard MPC:
Current State ───> [ Predicted Trajectory ] ───> Must hit Fixed Target (r)
(Infeasible if r jumps or is blocked!)
As highlighted by Köhler et al. (2023), this direct coupling creates a severe vulnerability. If an external planner issues a sudden setpoint change, or if the target $r$ lies outside the safe operating limits of the system, the current state may be physically unable to reach the new terminal set within the prediction horizon. The optimizer fails, leaving the system without a valid control input.
The Mechanism: Co-Optimizing a Reachable Proxy
Artificial equilibrium optimization bypasses this trap by rewriting the optimization problem. Instead of forcing the system to target the true reference $r$ directly, the controller targets an auxiliary, virtual reference $r_a$.
Artificial Reference MPC:
Current State ───> [ Predicted Trajectory ] ───> Artificial Reference (ra) ───> True Target (r)
(Always chosen to be feasible)
Crucially, $r_a$ is not fixed; it is an optimization variable determined online at every single time step. The controller optimization problem now minimizes a dual-objective cost function:
$$\text{Cost} = \sum_{k=0}^{N-1} \|x_k - r_a\|_Q^2 + \|u_k - u_a\|_R^2 + V_{\text{offset}}(r_a, r)$$
- The Tracking Cost (the summation) penalizes the deviation of the predicted states and inputs from the artificial equilibrium $(r_a, u_a)$.
- The Offset Cost ($V_{\text{offset}}$) penalizes the distance between the artificial reference $r_a$ and the actual desired target $r$.
The optimization is subject to a vital constraint: $r_a$ must be a valid, steady-state equilibrium point that satisfies all physical state and input constraints. As explained in a comprehensive tutorial by Krupa et al. (2024), this framework fundamentally shifts the control loop from direct targeting to a proxy-based navigation scheme.
Why It Works: Dynamic Buffering and Steady-State Manifolds
This formulation guarantees recursive feasibility through a simple piece of intuition: the optimization problem can always choose the artificial reference to be the system's current achievable state.
If the true target $r$ is completely unreachable (e.g., asking a vehicle to drive at an impossible speed), the offset cost will pull $r_a$ as close to $r$ as physically allowed. The artificial reference will naturally settle at the closest feasible steady state on the system's boundary, and the tracking cost will safely guide the system there, preventing the controller overreaction or failure that plagues standard tracking schemes (Berberich et al. (2021); Chen et al. (2022)).
When the true target $r$ suddenly jumps to a distant but reachable location, the artificial reference $r_a$ acts as a dynamic buffer or "stepping stone." At the moment of the jump, $r_a$ remains near the current state to preserve feasibility. Over subsequent time steps, the optimizer smoothly slides $r_a$ along the system's feasible steady-state manifold toward $r$, dragging the actual system state along behind it.
By decoupling the terminal ingredients from the fixed reference, the controller gains an extra degree of freedom (Aboudonia et al. (2022)). This mechanism ensures that the system maintains a valid solution path at all times, expanding the controller's domain of attraction and ensuring smooth transient performance even without complex terminal constraints (Ehmann et al. (2025)).
Generalization Across Domains
The principle of optimizing over an artificial reference has proven highly versatile, adapting seamlessly to nonlinear, data-driven, and uncertain environments.
Nonlinear and Dynamic Trajectories
For nonlinear systems, tracking changing targets typically requires recomputing complex terminal regions offline. Köhler et al. (2020) resolved this by introducing "reference generic" terminal ingredients that maintain stability guarantees for arbitrary trajectories without offline recalculation. When tracking highly dynamic target signals like periodic orbits, simultaneously optimizing a reachable proxy trajectory ensures exponential stability and rapid convergence (Köhler et al. (2019)).
Data-Driven Control
In modern data-driven MPC, where the system model is replaced entirely by historical data trajectories, artificial equilibria play a crucial role. Berberich et al. (2019) demonstrated that optimizing an artificial equilibrium online ensures recursive feasibility for changing setpoints using purely data-driven representations.
Furthermore, as reviewed by Allgöwer et al. (2024), constraining the artificial equilibrium to the system's steady-state manifold allows data-driven controllers to safely navigate toward target states that were completely absent from the initial training data, drastically expanding the operational envelope of the system.
Data-Driven Manifold Navigation:
[ Initial Training Data Zone ] ───(ra slides along manifold)───> [ Unseen Target Zone ]
Safe Exploration
This mechanism also provides a safety net for autonomous exploration. When navigating partially unknown environments where the final destination might be obstructed or outside sensor range, Soloperto et al. (2022) showed that optimizing intermediate reachable targets prevents local minima and ensures the controller always maintains a safe, feasible backup plan.
By treating the target reference not as an unyielding constraint, but as a flexible, optimized proxy, artificial equilibrium optimization provides a mathematically rigorous yet intuitive framework for robust, real-time tracking across the entire spectrum of modern control applications.
Key Foundational Work
Data-Driven Tracking MPC for Changing Setpoints (2019) — contributed 3 concept elements to this direction.
A nonlinear tracking model predictive control scheme for dynamic target signals (2019) — contributed 2 concept elements to this direction.
A nonlinear model predictive control framework using reference generic terminal ingredients (2020) — contributed 1 concept elements to this direction.