Model Predictive Control (MPC)

The defining feature of MPC is its receding horizon principle. At each time step, the controller:

• Solves a finite-horizon optimal control problem using the current state as the initial condition.
• Predicts the system's trajectory over a future window (the prediction horizon).
• Applies only the first control input from the computed optimal sequence.
• At the next time step, the horizon 'recedes' forward, and the process repeats with updated measurements. This provides continuous feedback and inherent robustness to disturbances and model inaccuracies.

MPC's predictive capability is grounded in an explicit mathematical model of the system's dynamics (e.g., differential or difference equations). This model is used to simulate future states. Key model types include:

• Linear Time-Invariant (LTI) Models: Enable fast, convex optimization (e.g., for process control).
• Nonlinear Models: Capture complex dynamics like robot kinematics or aerodynamics, requiring more computationally intensive Nonlinear MPC (NMPC).
• Hybrid or Piecewise-Affine Models: Handle systems with discrete modes (e.g., gears, contact). Model fidelity directly trades off with computational cost and real-time feasibility.

Unlike classical controllers, MPC explicitly handles constraints as part of the optimization problem solved in real-time. These constraints are fundamental to safety and performance:

• State Constraints: Keep the system within safe operating limits (e.g., joint angles, battery charge).
• Input Constraints: Respect actuator limits (e.g., torque, voltage).
• Path Constraints: Enforce conditions over trajectories (e.g., obstacle avoidance, speed limits). The optimizer finds control inputs that minimize a cost function (e.g., tracking error, energy use) while satisfying all constraints over the prediction horizon.

The controller's objective is encoded in a scalar cost function (often quadratic) minimized at each step. A typical formulation penalizes:

• Tracking Error: Deviation from a desired reference trajectory.
• Control Effort: Magnitude of control inputs to save energy and reduce wear.
• Terminal Cost: A penalty on the predicted state at the end of the horizon to ensure stability. The solution is the optimal control sequence for the finite horizon. While globally optimal over the infinite horizon is not guaranteed, well-tuned MPC with a sufficiently long horizon provides excellent performance.

MPC is inherently a feedback strategy. It requires accurate knowledge of the current system state to initialize its predictions. Since not all states are directly measurable (e.g., internal temperatures, friction), an observer or state estimator is critical:

• Kalman Filter: Optimal for linear systems with Gaussian noise.
• Extended Kalman Filter (EKF) / Unscented Kalman Filter (UKF): Used for nonlinear systems.
• Moving Horizon Estimation (MHE): An estimation counterpart to MPC that solves a similar optimization problem over a past horizon. This closed-loop structure corrects for model mismatch and external disturbances.

The main challenge of MPC is its computational demand. Solving an optimization problem at every control step (often at 10-1000 Hz) requires:

• Specialized Solvers: Use of fast, real-time optimization algorithms like QP solvers (for linear/quadratic problems) or SQP/IPOPT (for nonlinear problems).
• Hardware Acceleration: Implementation on FPGAs, GPUs, or dedicated control units.
• Approximation Techniques: Methods like explicit MPC (pre-computes the solution offline) or using simplified models to meet timing deadlines. The trade-off between prediction horizon length, model complexity, and sampling time is central to practical MPC design.

MPC is the industry standard for controlling complex, interconnected units like distillation columns, chemical reactors, and heat exchanger networks. It manages:

• Multiple Input, Multiple Output (MIMO) interactions between temperature, pressure, flow, and concentration.
• Long time delays and non-linear dynamics inherent in reaction processes.
• Hard constraints on valve positions, product purity, and safe operating temperatures. The controller solves an optimization problem every few seconds to maximize yield, minimize energy use, and ensure safety, often achieving efficiency improvements of 5-15% over traditional PID loops.

In self-driving cars, MPC is used for local trajectory planning and tracking. It computes the optimal steering, throttle, and brake inputs over a short time horizon (1-3 seconds) by solving an optimization that:

• Minimizes deviation from a global path while ensuring passenger comfort (smooth acceleration).
• Enforces kinematic and dynamic constraints (e.g., maximum steering angle, tire friction limits).
• Incorporates predictions of surrounding vehicles' motion to maintain safe distances.
• Respects traffic rules and lane boundaries as hard constraints. This allows the vehicle to react smoothly and safely to dynamic environments.

For robotic arms, MPC enables dynamic, constraint-aware motion. It is crucial for tasks requiring:

• Contact-rich manipulation (e.g., assembly, polishing) where forces must be carefully regulated.
• High-speed motion (e.g., pick-and-place) where actuator torque limits and jerk constraints are critical.
• Mobile manipulation where the base and arm must be coordinated. The controller uses a dynamic model of the robot to predict future states and computes joint torques that drive the end-effector along a desired trajectory while explicitly avoiding self-collision, joint limits, and torque saturation.

MPC optimizes the flow and storage of energy in complex grids, particularly with renewable sources. Key applications include:

• Building HVAC Control: Minimizing energy consumption while maintaining comfort zones across hundreds of zones, predicting thermal loads and weather.
• Microgrid Management: Balancing generation from solar/wind with battery storage and diesel backups to meet demand at lowest cost.
• Power Plant Load Following: Adjusting turbine output to follow electricity demand while respecting ramp-rate limits to reduce mechanical stress. The controller manages the inherent storage dynamics of batteries and thermal mass, and the uncertainty in renewable generation and demand forecasts.

In aircraft and spacecraft, MPC handles challenging control problems with strict safety margins:

• Quadrotor/Drone Control: Stabilizing agile flight, performing aggressive maneuvers, and ensuring robustness to wind gusts.
• Missile Guidance: Computing optimal intercept trajectories under aerodynamic and propulsion constraints.
• Satellite Attitude Control: Precisely orienting satellites using reaction wheels or thrusters while minimizing fuel consumption.
• Aircraft Engine Control: Managing fan speed, pressure ratios, and temperatures in jet engines for optimal performance across flight envelopes. MPC's ability to handle actuator saturation and multivariable coupling is critical here.

A life-critical application of MPC is in the artificial pancreas for Type 1 diabetes management. The controller:

• Uses a physiological model of the patient's glucose-insulin dynamics.
• Predicts future blood glucose levels based on continuous glucose monitor (CGM) readings, meal announcements, and physical activity.
• Computes optimal insulin infusion rates from a pump to keep glucose within a safe range (70-180 mg/dL).
• Imposes hard safety constraints to prevent hypoglycemia (dangerously low blood sugar). This is a quintessential example of MPC managing a stochastic, safety-critical, and personalized system with significant time delays.

The mathematical process of computing a time-parameterized path (a trajectory) that minimizes a defined cost function—such as energy, time, or jerk—while satisfying a system's dynamic constraints and avoiding collisions. MPC is an online, receding-horizon form of trajectory optimization, solving a finite-time problem repeatedly from the current state.

• Local vs. Global: MPC typically performs local optimization from the current state, whereas global trajectory optimization may plan an entire path offline.
• Cost Functions: Common objectives include minimizing control effort, tracking error, or final time.
• Applications: Used in robotics for smooth arm movements, autonomous vehicle path planning, and drone flight control.

The algorithmic discipline of computing a sequence of valid configurations or states that move a robot from a start to a goal while avoiding obstacles and respecting kinematic and dynamic constraints. MPC is often employed as the low-level controller that executes and dynamically adjusts the motion plan.

• Hierarchy: Motion planning sits between high-level task planning and low-level control (like MPC).
• Geometric vs. Dynamic: Path planning finds a geometric route, while motion planning incorporates dynamics and timing.
• Key Algorithms: Includes sampling-based methods like RRT and PRM, and optimization-based methods.

The defining operational principle of MPC. At each control time step, the algorithm solves an optimization problem over a finite future time horizon but implements only the immediate control input. The horizon then "recedes" forward, and the process repeats with a new measurement. This provides inherent feedback and robustness to disturbances.

• Feedback Integration: New sensor data is incorporated at every step, correcting for model inaccuracies and noise.
• Horizon Length: A critical tuning parameter balancing computational load with anticipation of future events.
• Contrast with Linear-Quadratic Regulator (LQR): LQR uses a fixed, infinite-horizon optimal gain; MPC recomputes the plan online over a finite horizon.

The broader mathematical field concerned with finding a control law for a dynamical system that optimizes a performance index. MPC is a subfield of optimal control distinguished by its use of online numerical optimization and explicit handling of constraints.

• Open-Loop vs. Closed-Loop: Classical optimal control often computes a single open-loop trajectory. MPC is a closed-loop, feedback strategy.
• Pontryagin's Maximum Principle: Provides necessary conditions for optimality in continuous-time problems.
• Dynamic Programming: Solves the Hamilton-Jacobi-Bellman equation for the optimal cost-to-go function, often intractable for high-dimensional systems.

A mathematical model of a physical system as a set of input, output, and state variables related by first-order differential (continuous) or difference (discrete) equations. MPC relies entirely on a state-space model to predict future system behavior.

• Core Equation: x_{k+1} = f(x_k, u_k) where x is the state vector and u is the control input.
• Observability & Controllability: Fundamental properties that determine if a state can be estimated from outputs and if it can be driven to a desired value.
• Linear Time-Invariant (LTI): The most common form for MPC, where f is a linear function: x_{k+1} = A x_k + B u_k.

The explicit management of physical and operational limits within the control optimization. A primary advantage of MPC over other control methods is its ability to systematically handle hard constraints (e.g., joint limits, obstacle boundaries) and soft constraints (e.g., comfort zones).

• Types of Constraints: Includes input constraints (actuator limits), state constraints (safe operating regions), and output constraints.
• Formulation: Constraints are included directly as inequalities in the online optimization problem.
• Feasibility: A central challenge is ensuring a solution exists that satisfies all constraints at each time step.