Model Predictive Control (MPC)

The system model is a mathematical representation of the controlled process dynamics, used to predict future states. For fleet orchestration, this is often a kinematic or dynamic model of each agent (e.g., differential drive robot).

• Key Types: Linear Time-Invariant (LTI) models, non-linear models, or data-driven surrogate models.
• Purpose: Provides the predictive capability essential for the controller's forward-looking optimization.
• Example: A state-space model predicting an Automated Guided Vehicle's (AGV) position and velocity based on wheel velocity commands.

The cost function quantifies controller performance, defining what 'optimal' means for the optimization solver. It is minimized over the prediction horizon.

• Typical Terms: Penalties for tracking error (deviation from target), control effort (energy/actuation), and constraint violation.
• Scheduling Context: Can include terms for minimizing total travel time, maximizing throughput, or balancing fleet utilization.
• Design: A weighted sum of quadratic or linear terms, where tuning the weights balances competing objectives like speed vs. energy consumption.

Constraints are hard or soft limits imposed on system states and control inputs, ensuring solutions are physically realizable and safe.

• Common Types:
  • State Constraints: Limits on position, velocity, or battery level.
  • Input Constraints: Limits on actuator commands (e.g., max wheel torque).
  • Spatial-Temporal: Collision avoidance (minimum separation), time windows for task completion, and capacity constraints for loading zones.
• Role: The optimization solver finds control actions that satisfy all constraints over the prediction horizon.

The optimization solver is the computational engine that, at each control step, solves the finite-horizon problem defined by the model, cost function, and constraints.

• Algorithm Types: For linear/quadratic problems, Quadratic Programming (QP) solvers (e.g., OSQP, qpOASES). For non-linear problems, Nonlinear Programming (NLP) solvers or Sequential Quadratic Programming (SQP).
• Real-Time Requirement: Must compute a solution within the sampling period (often <100ms). This drives the use of efficient, specialized solvers.
• Output: The optimal sequence of control inputs for the horizon; only the first step is applied.

The prediction horizon (N) is the number of future time steps the MPC looks ahead. The control horizon (M ≤ N) is the number of steps over which control inputs are optimized.

• Prediction Horizon: A longer horizon improves anticipatory control but increases computational cost.
• Control Horizon: Often shorter; beyond M, control inputs are assumed constant or zero. This is a common simplification to reduce problem complexity.
• Tuning: Critical for stability and performance. In scheduling, the horizon must be long enough to foresee potential conflicts or bottlenecks.

The state estimator provides the initial state for the optimization problem at each time step, fusing sensor data (e.g., LiDAR, odometry) and model predictions.

• Function: Compensates for noisy measurements and unmeasured states.
• Common Technique: A Kalman Filter (linear systems) or Extended Kalman Filter (non-linear systems).
• Importance: MPC's predictions are only as good as its initial state. Accurate estimation is vital for fleet state estimation in dynamic environments.

MPC is the industry standard for controlling complex chemical plants, refineries, and distillation columns. It excels here because it can:

• Handle multiple interacting variables (temperature, pressure, flow rates) simultaneously.
• Explicitly enforce critical safety and quality constraints (e.g., reactor temperature limits).
• Optimize for economic objectives, like maximizing yield or minimizing energy use, while maintaining stable operation.

Example: Controlling a catalytic cracking unit to maximize gasoline production while staying within safe operating envelopes.

In autonomous driving, MPC is used for local path planning and motion control. The controller:

• Uses a vehicle dynamics model to predict future poses over a short horizon.
• Solves for optimal steering, throttle, and brake inputs.
• Incorporates constraints like lane boundaries, collision avoidance with predicted obstacle trajectories, and actuator limits.
• Continuously re-plans as new sensor data arrives, enabling smooth, safe navigation in dynamic environments.

MPC is applied to aircraft for energy-efficient maneuver execution and autonomous guidance. Key applications include:

• Trajectory tracking for unmanned aerial vehicles (UAVs) in windy conditions.
• Formation flying control for multiple aircraft.
• Managing constrained actuator dynamics (control surface deflection limits, engine thrust bounds). Its ability to handle multivariable coupling (e.g., between pitch, roll, and yaw) and directly optimize for fuel consumption makes it superior to classical PID cascades for complex flight regimes.

For robotic arms and mobile manipulators, MPC provides high-precision, torque-optimal control. It is particularly effective for:

• Pick-and-place operations with dynamic obstacles.
• Tasks requiring force control and physical interaction (e.g., polishing, assembly).
• Controlling under-actuated or non-holonomic systems like quadrupeds or wheeled robots. The controller uses a rigid-body dynamics model to predict joint torques and positions, ensuring smooth, collision-free motion that respects motor torque and velocity constraints.

MPC is crucial for optimizing energy flow in complex, interconnected systems. Primary uses include:

• Microgrid operation: Balancing generation from renewables (solar, wind), storage (batteries), and load demand.
• Building HVAC control: Minimizing energy consumption while maintaining occupant comfort zones across multiple zones.
• Power plant load-following: Adjusting generator output efficiently to meet fluctuating grid demand. MPC's strength lies in predicting disturbances (e.g., weather changes affecting solar output) and preemptively adjusting control actions to minimize cost or maximize efficiency over a rolling horizon.

In the context of Heterogeneous Fleet Orchestration, MPC acts as the real-time replanning engine for spatial-temporal scheduling. It:

• Models the kinematic and operational constraints of mixed fleets (AMRs, AGVs, manual vehicles).
• Predicts future system states (agent positions, task completion) based on current assignments and velocities.
• Continuously solves for the optimal sequence of task allocations, routes, and velocities over a receding horizon.
• Dynamically incorporates new orders, agent failures, or delays, recalculating plans to minimize makespan or total travel distance while avoiding deadlocks and collisions.

The foundational mathematical framework for modeling sequential decision-making under stochastic (probabilistic) uncertainty. An MDP is defined by a tuple: states, actions, transition probabilities, and a reward function. It provides the theoretical underpinning for both Reinforcement Learning and stochastic optimal control, of which stochastic MPC is a variant.

• Core Concept: The Markov property—the future state depends only on the present state and action, not the full history.
• In Scheduling: A fleet's state (agent positions, battery levels) evolves probabilistically based on actions (move, pick, charge), with rewards for efficient task completion.

A mathematical optimization technique where an objective function is minimized or maximized subject to constraints, with some decision variables required to be integers. It is the workhorse solver for many MPC formulations in scheduling, where discrete decisions (e.g., assign task 'yes' or 'no', sequence order) must be made.

• Key Use: Encoding binary variables for task assignment and integer variables for sequencing in the MPC's online optimization problem.
• Example: A MIP solver determines which robot picks which order and in what sequence over the next 10-minute horizon.

A programming paradigm for solving combinatorial problems by declaring relations (constraints) between variables. Instead of using a monolithic objective function, CP uses powerful propagation and search algorithms to prune infeasible solutions. It excels in highly constrained scheduling problems with complex logical rules.

• Key Strength: Native handling of complex global constraints (e.g., 'all different' for task sequencing, cumulative resource constraints).
• vs. MIP: Often more efficient for feasibility problems or those with many logical constraints; can be integrated with MIP in hybrid solvers.

A methodology for optimization under uncertainty that seeks solutions which remain feasible and perform well for all possible realizations of uncertain parameters within a defined uncertainty set. It is a key approach to making MPC schedules resilient to disturbances like agent delay or task duration variability.

• Core Idea: Prioritize worst-case performance over average performance.
• In MPC: The optimization is solved with constraints that must hold for any demand fluctuation or travel time within a bounded set, producing a conservative but reliable plan.

A virtual, dynamic representation of a physical system (e.g., a warehouse fleet) that uses real-time data and simulation models to mirror its state and behavior. It is the ideal prediction model for an MPC controller, providing the high-fidelity environment in which future system trajectories are simulated during the optimization step.

• Key Function: Enables 'what-if' analysis and safe evaluation of candidate control sequences.
• Integration with MPC: The digital twin's simulation engine is called repeatedly by the MPC solver to evaluate the outcome of different action sequences over the prediction horizon.