The sea is a liar to machines. It pretends to be calm, then swells with a gust of wind, a school of chop, a current that redefines your whole path. For autonomous vessels, that unreliable theatre is the ultimate test. A recent study from TU Delft andDemcon Unmanned Systems puts a new control idea to the test on a tiny, electric, seagoing workhorse called the DUS V2500. The authors—Jelmer van der Saag, Elia Trevisan, Wouter Falkena, and Javier Alonso-Mora—embark on a journey from simulation to salty air to ask: can a control strategy that treats disturbances as something to be actively rejected keep a ship on a precise course even when the ocean is throwing everything at it?
What makes this question gripping isn’t just the technical trick, but the promise. If you can make a USV follow a trajectory with less drift in waves and currents, you unlock safer hydrographic surveys, quicker inspections, and more productive near-shore missions—all without assuming perfect knowledge of the water’s whims. The paper demonstrates that Active Disturbance Rejection Control (ADRC) can significantly improve trajectory tracking relative to the tried-and-true PID approach, but it does so at a cost: more energy use, and more torque demands in real-world sea conditions. The study gives us a window into a future where autonomous boats adapt on the fly to whatever the sea throws at them, rather than relying on a perfect model of the world.
The ADRC idea: letting disturbances do the talking
ADRC is a different way of thinking about control. Instead of trying to perfectly model every hydrodynamic jagged edge or every gust of wind, ADRC treats the unknown dynamics and disturbances as a single, catch-all term—the “total disturbance.” The system then actively estimates and counteracts that disturbance through a three-part architecture: an Extended State Observer (ESO) that sneaks up on the disturbance hidden inside the system, a Tracking Differentiator (TD) that shapes how the controller moves toward a target, and a nonlinear state error feedback that converts those estimates into actual control actions. In practical terms, the controller is always asking: what is pushing the vessel off its intended path, and how can I push back preemptively? The paper lays out the math in a compact way, but the spirit is more intuitive: you don’t have to know exactly what the sea is doing to compensate for it; you just have to estimate its effect well enough to cancel it out.
In the paper’s design, ADRC isn’t a single magic switch. It’s deployed as three parallel controllers, one for each planar degree of freedom—surge (forward motion), sway (lateral motion), and yaw (turning). The heading of the vessel, which guides steering toward a path, is fed by an L1 guidance law. The mission speed is fixed at 1.4 m/s, but the system adapts during corners to improve maneuverability. The clever trick is to treat the coupling between motion axes as manageable—ADRC’s decoupling tends to hold up in practice even when the true dynamics are messy. This is especially important for the DUS V2500, a 2.5-meter, underactuated, rudderless vessel with two stern thrusters and a bow thruster, designed to operate near shore with minimal crew. The authors emphasize that accurate hydrodynamic parameters are hard to pin down in real time, so a model-free or model-light approach like ADRC can be appealing in this domain.
What ADRC buys you is robustness. It can estimate and compensate for not only external disturbances like waves and currents but also internal uncertainties inside the vessel’s dynamics. That makes it an attractive candidate for USVs, where environmental conditions swing unpredictably and model fidelity drifts from one mission to the next. But the same robustness that helps the ship fight disturbances also comes with a price tag: more aggressive control actions, higher energy use, and the risk of torque saturation when the disturbances spike beyond what the actuators can handle. The authors were careful to quantify this trade-off, setting up a direct comparison with a standard PID controller across a range of conditions, including calm water, a steady current, significant waves, and a combination of both.
From lab to sea: building a vessel and a testing ground
The DUS V2500 is a compact, electric USV, purpose-built for inspection and hydrographic tasks near shore. Its small size belies the complexity of controlling it on a moving surface: three degrees of freedom in the plane, underactuation, and thruster delays that complicate any feedforward effort. The vessel’s position and orientation are estimated by fusing GNSS with an Inertial Measurement Unit, achieving centimeter-scale pose accuracy in ideal conditions. But those same sensors only tell part of the story when seas rise and currents push sideways. The authors acknowledge that fully characterizing the hydrodynamics would require expensive, elaborate measurements, something not readily available on this platform in real time. That limitation makes ADRC a compelling option, at least in theory and in early practical tests.
To test ADRC rigorously, the authors built a dual testing ground: a Unity-based software-in-the-loop simulator that delivers a digital twin of the onboard control computer and a physics-based “Dynamic Water Physics 2” (DWP2) ocean model for six-degrees-of-freedom vessel dynamics. They augmented DWP2 with the CREST ocean renderer to generate a Pierson–Moskowitz wave spectrum, capturing sea states up to Douglas Sea State 4 (waves up to about 2.5 meters). Currents, wind, and wave-induced momentum transfer were included to create a realistic trial bed, with sea conditions adjustable to mimic the unpredictable wind and water profiles a USV would meet in Scheveningen harbor and just offshore.
One of the paper’s technical innovations is a practical approach to the onboard thruster delay. Real boats don’t respond instantaneously to a command—the propellers have inertia, gears take time to spin up, and the water provides resistance. The authors approximate these delays with a first-order low-pass filter and a delay-compensation term that nudges the commanded thrusts to anticipate their own latency. In a field like marine robotics, the inclusion of such a delay model is not a cosmetic add-on; it’s the difference between a controller that looks good on a screen and one that behaves reliably on a real boat in rough water.
In the field, two kinds of tests were conducted: harbour-based trials with relatively light disturbances and near-shore trials about a kilometer from shore where conditions were calmer than the open sea but still nontrivial. The same trajectories were run under ADRC and PID control in both arenas, with the same mission speed and the same baseline controller values for a fair apples-to-apples comparison. Across simulation and field tests, the researchers tracked a handful of metrics, notably the cross-track error (XTE)—the lateral deviation from the intended path—and battery usage. The parcel of data they collected allowed them to separate how well each controller held the line from how much energy it burned to do so.
Beyond the direct experiments, there’s a broader methodological point worth noting. The study demonstrates a thoughtful end-to-end workflow: a Unity-based digital twin for rapid iteration, a hydrodynamics-informed wave model for realism, and a real vessel for ground truth. That combination—simulation-informed design with real-world validation—feels like a blueprint for how to push autonomy from clever ideas to deployable systems in a field where every extra knot of disturbance matters.
What the results teach us about the future of autonomous shipping
The headline finding is surprisingly straightforward: ADRC delivers a meaningful improvement in trajectory tracking compared with a traditional PID controller across all tested conditions. In simulation, cross-track error shrank by about 30% to 40%. In the real world, the improvement was more modest—roughly 10% to 20%—but still statistically significant and practically meaningful when every centimeter of drift can translate into missed survey lines or failed inspections. In short, ADRC makes USVs sit a little closer to their intended path, even when the ocean refuses to play along.
But there’s a crucial caveat. ADRC’s stronger performance comes with higher energy consumption. In calm water, the energy cost rose noticeably for ADRC; when currents were present, ADRC completed trajectories faster, which mitigated some energy use, but in near-shore trials the energy penalty was pronounced—roughly a 50% increase in power draw relative to PID. The root cause, the authors explain, is the feed-forward term that compensates the total disturbance. Since x3 blends internal dynamics with external disturbances, the controller sometimes overcompensates, driving torque demand up and causing spikes that push the batteries harder than PID would under the same conditions.
The team doesn’t stop at diagnosing the issue. They suggest a practical remedy: scale the disturbance compensation to avoid overreacting to every gust or wave. They also note that ADRC could be deployed more wisely when combined with partial, known physics of the system. In other words, ADRC is powerful, but it’s not a silver bullet. The real power may lie in a hybrid approach that uses ADRC’s robust disturbance rejection where model accuracy is weak, while leaning on traditional, well-characterized dynamics where they’re reliable. That balance—robustness where the model fails, efficiency where the model is trustworthy—feels like the right tempo for autonomous shipping as a practical, scalable technology.
From a broader perspective, the study contributes a clear, practical signal: when you push autonomous ships into the real world—near the shore, where currents shift and waves crash unpredictably—the story is not just about clever algorithms. It’s about managing energy budgets, actuator limits, and safety margins. ADRC’s strength is that it doesn’t pretend to know the sea perfectly; it tries to sense what the sea is doing to the ship and respond in real time. The weakness is that, in doing so, it can gnaw more on the power supply and push the hardware harder than conventional controllers in some situations. The authors’ honesty about these trade-offs matters, because it frames the engineering challenge not as a single fix but as a careful engineering choice with context and consequences.
Finally, the paper is a reminder that the best ideas in autonomous systems often come from cross-pollination between fields. ADRC originated in control theory and has found a receptive home in robotics, but its application to marines requires embracing oceanography-scale uncertainty, field-testing rigor, and engineering discipline around energy use and actuator limits. The work stands as a concrete demonstration that machine autonomy can get better at “listening” to disturbances and adjusting in real time, even if the ship sometimes pays in the moment for it with a few more amp-hours in the battery. The sea’s test isn’t just about whether a controller can stay on a line; it’s about whether we can design controls that sustain performance and reliability in the long voyage ahead.
As the authors note, the team behind this study hails from the Cognitive Robotics Department at TU Delft, supported by Demcon Unmanned Systems. The lead authors—Jelmer van der Saag, Elia Trevisan, Wouter Falkena, and Javier Alonso-Mora—helped bridge theory, simulation, and field validation. Their work doesn’t claim to have invented the perfect control system for every sea state; it offers a carefully measured step toward that horizon. If ADRC can be tuned to scale disturbances without draining energy, it could become a standard tool in the autonomous-maritime toolkit, a quiet enabler of safer, more reliable operations in the challenging, beautiful, and occasionally merciless ocean.
