Underwater Marine MPC — Constrained Control for AUVs with Actuator Saturation and Coupled Dynamics

Public reference use case (by others): This page summarizes publicly available reference implementations and papers. Not client results of Dr. Rafal Noga.

Why This Matters (Executive Summary)

• Business pain: Autonomous underwater vehicles (AUVs) operate under hard thruster limits, coupled multi-axis dynamics, and limited sensing — conventional decoupled PID loops violate actuator constraints during transients, causing poor depth control, excessive overshoot, or instability.
• Solution class: Constrained multivariable MPC exploits the coupled depth-pitch model and enforces thruster force and rate limits by construction, eliminating constraint violations that single-loop controllers cannot prevent.
• Measurable outcomes: Published experiments report depth overshoot below 10% (typically around 5%) during step changes while commanding zero pitch, with stable operation across multiple tuning configurations and validated lake deployments.
• Transferable pattern: AUV control is a clean instance of the “constrained MIMO MPC” pattern that transfers to any mechatronic system with coupled outputs, hard actuator limits, limited sensing, and safety envelopes — including multi-axis positioning systems, robotic manipulators, and industrial drives.

The Design Pattern Explained

Underwater vehicles present a textbook case for constrained MPC. The depth and pitch axes are coupled through the vehicle’s hydrodynamics: commanding a depth change affects pitch, and pitch changes affect depth tracking. Thruster limits — both maximum force and maximum rate of change — are hard physical constraints that cannot be exceeded without damaging equipment or losing control. Conventional approaches decouple depth and pitch into separate PID loops and ignore actuator limits, leading to integrator windup, constraint violation, and poor transient performance.

MPC solves this by optimizing thruster commands over a prediction horizon using a coupled depth-pitch model, subject to explicit thruster force and rate constraints. At each control interval, the optimizer computes the best feasible sequence of commands, applies the first, and repeats. When an unconstrained optimum would violate thruster limits, the optimizer automatically finds the closest feasible solution — there is no windup, no saturation surprise, and no need for anti-windup logic.

The underwater environment adds further challenges: GPS is unavailable, acoustic positioning is slow and noisy, communication bandwidth is minimal, and safety constraints (depth limits, seabed collision avoidance, emergency ascent) are non-negotiable. These factors make onboard autonomy and robust state estimation essential complements to the MPC layer.

Applications & Reference Implementations

Application 1: Constrained MPC for AUV Depth Steps with Zero-Pitch Command

Researchers at the University of Southampton developed a constrained MPC controller for the vertical-plane motion of a thruster-actuated AUV, treating depth and pitch as a 2x2 MIMO system with front and rear vertical thrusters as inputs. The controller enforced hard constraints on thruster force magnitude and rate of change (delta-thrust), with thrust bounds specifically configured to avoid motor stop and the associated startup dead-band. Experimental step-change tests from the surface to 1 m depth, with pitch commanded at 0 degrees, demonstrated depth overshoot below 10% — with approximately 5% typical in stable configurations and one documented test showing 7.6% maximum measured overshoot. The implementation ran in Python within ROS on Linux, consuming depth measurements from a pressure transducer and pitch from a digital compass. A systematic tuning study across 22 experimental tests with varying prediction horizons (Np = 25 to 100) and cost weights established that Np = 50 provided the fastest stable response, while Np = 25 was insufficient for stability. ¹

Application 2: Hover-Capable Hybrid AUV — MPC Validated in Lake Deployment

The Delphin2 hover-capable hybrid AUV was equipped with MPC control laws based on a vehicle model with parameters identified from experimental data. Unlike conventional torpedo-shaped AUVs that rely on forward speed for control authority, the Delphin2 can hover in place using multiple thrusters, creating an over-actuated control problem. The MPC was evaluated in a large lake deployment scenario, targeting practical applications such as environmental monitoring and invasive species detection (zebra mussel surveys). The lake trials provided real-world validation under currents, temperature gradients, and sensor noise — bridging the critical gap between simulation results and operational deployment. The over-actuated nature of the vehicle makes MPC particularly valuable because it can optimally allocate thrust across redundant actuators while respecting individual thruster limits. ²

Application 3: Depth-Pitch MPC with Explicit Actuator Constraint Handling

A complementary study focused on the depth-and-pitch control problem with particular attention to how MPC handles actuator constraints during transient maneuvers. The controller was designed to execute depth step changes while maintaining zero pitch — a requirement that prevents the vehicle from tilting during vertical repositioning, which is critical for sensor payload stability and collision avoidance near the seabed. The Hildreth programming procedure was used to solve the constrained quadratic program at each control interval, providing a computationally tractable method for real-time constraint enforcement. Monitoring the number of Hildreth iterations per sample provided an operational indicator of constraint activity — effectively a “constraint stress” metric that operators can use to assess how close the vehicle is operating to its actuator limits. ³

Application 4: Transferable MIMO MPC Pattern for Industrial Mechatronics

The AUV depth-pitch problem is structurally identical to many industrial mechatronic control challenges: multi-axis positioning stages with coupled dynamics and actuator limits, robotic arms with joint torque constraints, and multi-drive systems with shared power budgets. The linearized model with augmented integrator states, explicit input and rate constraints, and Hildreth-based QP solution represents a minimal but complete constrained MIMO MPC implementation that can be adapted to any system where actuator saturation and cross-coupling define the control problem. The key insight is that once you have a credible coupled model and well-defined actuator limits, the MPC formulation is largely standard — the engineering effort is in the modeling and validation, not in the control algorithm itself.

What This Means for Your Operations

If your system has multiple coupled outputs, hard actuator limits, and safety constraints — and your current decoupled PID loops struggle during transients or near operating boundaries — constrained MIMO MPC offers a systematic solution. The AUV cases demonstrate the pattern at its cleanest: two coupled outputs, two constrained actuators, explicit safety limits, and experimental validation under real-world disturbances.

For DACH industrial operations, the same pattern applies to: multi-axis CNC positioning with drive current limits, robotic systems with joint torque constraints, HVAC systems with coupled temperature/humidity and limited actuator capacity, and any process where saturation-induced windup is a recurring operator headache.

Prerequisites for deployment: reliable measurements of the controlled variables, a command interface to the actuators, sufficient knowledge of the coupled dynamics to build a predictive model, and defined actuator limits and safety constraints.

How We Deliver This (Engagement Model)

• Phase 0: NDA + data request — share system specifications, actuator datasheets, sensor configuration, current control architecture, and known constraint limits.
• Phase 1: Fixed-scope discovery (2-4 weeks) — coupled dynamics identification (step tests or existing data), constraint mapping, MPC feasibility assessment with simulated constraint scenarios.
• Phase 2: Implementation + validation + commissioning — constrained MPC development, simulation validation across the operating envelope, experimental commissioning with standardized acceptance maneuvers (step tests, disturbance rejection).
• Phase 3: Monitoring + training + scaling — constraint activity monitoring dashboards, operator training on MPC behavior near limits, and extension to additional axes or operating modes.

Typical KPIs to Track

• Safety: Depth/position overshoot (% of step), constraint violation count, emergency stop frequency
• Performance: Settling time (time to within +/-10% of setpoint), steady-state tracking error, cross-axis coupling rejection
• Actuator health: Time spent at saturation limits, actuator rate-of-change utilization, motor start/stop cycles avoided
• Operator burden: Manual interventions during transients, time spent retuning PID gains after operating point changes

Risks & Prerequisites

• Model fidelity: The coupled dynamics model must capture the dominant interactions between controlled axes; significant unmodeled nonlinearities (e.g., hydrodynamic effects at higher speeds) may require nonlinear MPC or gain scheduling.
• Actuator characterization: Thruster (or drive) dead-bands, saturation limits, and rate limits must be accurately known — incorrect constraint definitions can lead to infeasibility or conservative operation.
• Compute latency: The MPC must solve within the control interval; for the reported AUV cases, a 0.1 s sample time was used with Python/ROS, but higher-performance systems may require compiled solvers.
• Sensor noise: Underwater (or industrial) environments introduce measurement noise; state estimation or filtering may be needed to provide clean feedback to the MPC.
• Incremental deployment: Start with the most critical axis pair, validate constraint handling on standardized maneuvers, then extend to additional axes and operating scenarios.

FAQ

Q: What does MPC add versus PID for coupled multi-axis systems? A: MPC uses a coupled model to predict future behavior across all axes simultaneously and computes coordinated actuator commands that respect all constraints. PID treats each axis independently, cannot anticipate constraint violations, and requires anti-windup add-ons that are difficult to tune for MIMO systems. The AUV cases show that MPC achieves bounded overshoot and zero-pitch maintenance that decoupled PID cannot guarantee.

Q: How do you handle actuator dead-bands (e.g., thruster startup behavior)? A: In constrained MPC, dead-band avoidance is encoded as a constraint — for example, requiring thrust commands to stay above a minimum operating threshold. This is more reliable than PID-based workarounds because the optimizer plans ahead to avoid entering the dead-band region rather than reacting after the fact.

Q: Is this approach limited to underwater vehicles? A: Not at all. The constrained MIMO MPC pattern is identical for any system with coupled dynamics and actuator limits — multi-axis positioning, robotic manipulation, multi-zone thermal control, or any MIMO process where saturation and cross-coupling are the dominant challenges.

Q: What is the minimum sensor requirement? A: You need reliable measurements of all controlled variables (or a state estimator to reconstruct them from available sensors) and a command interface to the actuators. In the AUV cases, depth was measured via pressure transducer and pitch via digital compass — both are low-cost, robust sensors.

Book a 30-Minute Discovery Call

Ready to explore whether this pattern fits your system?

Dr. Rafal Noga — Independent APC/MPC Consultant

📧 Email me · 🌐 noga.es

Fixed-scope discovery · NDA-first · DACH on-site available

Public References

Footnotes

1. Steenson et al., “Experimental Verification of a Depth Controller using Model Predictive Control with Constraints onboard a Thruster Actuated AUV” (IFAC NGCUV, 2012). https://eprints.soton.ac.uk/346564/1/IFAC-Leo_20Steenson_5B1_5D.pdf ↩

2. Steenson et al., “Model Predictive Control for a Hover-Capable Hybrid AUV” (NORA/NERC). https://nora.nerc.ac.uk/507532/1/Steenson%20paper.pdf ↩

3. Steenson et al., “Constrained MPC for AUV Depth and Pitch Control” (ScienceDirect / IFAC, 2012). https://www.sciencedirect.com/science/article/pii/S1474667016306152 ↩