When a Ship Learns to Dock Itself: The Promise of Simpler Models for Autonomous Berthing
The hardest part of sailing isn't the open ocean. It's the last hundred meters.
That's where pilots describe their palms going white-knuckled, where tugboat captains earn their reputations, where even veteran mariners feel their stomachs tighten. Berthing—guiding a massive vessel safely alongside a pier or quay—is the maritime equivalent of threading a needle while riding a bronco. It requires reading wind, current, water depth, and the invisible push of neighboring hulls. It demands split-second adjustments based on decades of gut feel. It is, in short, deeply, irreducibly human.
And it is precisely the task that may soon surrender to automation.
A team of researchers from Japan—Agnes N. Mwange, Taichi Kambara, Kouki Wakita, Kazuyoshi Hosogaya, and Atsuo Maki—has published a study that nudges us closer to autonomous port operations. Their work, published on arXiv in July 2026, attacks one of the central obstacles: the infuriating complexity of modeling how ships behave at low speeds, in shallow water, near docks. The researchers propose a surprisingly elegant solution. Instead of building ever-more-complicated mathematical models of ship behavior, they show that a simpler, linear model—one built from real ship data, not theory alone—can capture low-speed maneuvering with remarkable accuracy.
The implications ripple outward. If we can model a ship's behavior at low speed with a compact set of equations, we can then feed those equations into a control system that tells the vessel how to steer itself. The dream is a ship that glides into port with the same precision your car parks itself in a garage. The reality is closer—but still a few nautical miles away.
The Science
Why Low Speed Is Hard
Before diving into the research, it helps to understand why low-speed ship maneuvering has stymied automation efforts for decades.
A ship sailing at cruising speed in open water is, in some ways, a physics student's dream. The vessel moves steadily forward. Water flows past the hull in predictable patterns. Forces like drag and lift can be modeled with reasonable accuracy using equations developed over a century of naval architecture. The ship responds to its rudder in ways captains can anticipate and adjust for.
At low speeds—approaching and entering port—the physics becomes anarchic.
Consider what happens when a vessel decelerates from 10 knots to near-zero. Its massive hull still carries enormous momentum. The water around it, disturbed by the hull's passage, continues moving in complex patterns that interact with the seabed, with the dock face, with other vessels. The rudder, which provides steering authority at speed, becomes nearly useless when the ship is barely moving. Propeller thrust, water jets, bow thrusters, and tug assistance must compensate—but these forces interact in ways that are deeply non-linear. Small inputs produce disproportionate effects. The same control input that works flawlessly at 2 knots may fail entirely at 0.5 knots. Time delays, hysteresis, and coupling between different axes of motion (heading, drift, yaw) create a system that defies the clean differential equations engineers love.
The traditional response to such complexity has been to build more complex models. Researchers have developed elaborate mathematical frameworks that attempt to capture every relevant force—the influence of shallow water on the flow patterns beneath the hull, the interaction between the ship's underwater geometry and nearby structures, the non-linear relationship between rudder angle and turning force. These models can be extraordinarily detailed, incorporating dozens or hundreds of parameters that encode the specific characteristics of a particular vessel.
But complexity has costs. More parameters mean more opportunities for error. More intricate equations mean more computational burden. And crucially, more sophisticated models require more sophisticated control systems to exploit them—systems that may be fragile, difficult to tune, and hard to verify for safety-critical applications.
A Different Philosophy
The Japanese research team took a different tack. Instead of asking "what forces govern ship motion at low speed?" and building equations to capture them, they asked a simpler question: "what does the ship actually do?"
This is the essence of system identification—a branch of control theory that attempts to construct mathematical models directly from observed data. Instead of deriving equations from physics, you measure inputs and outputs (thrust commands and the resulting motion) and then find the equations that best reproduce those observations.
The approach has a long history in control engineering. When you can't derive a model from first principles, or when your first-principles model is too complex to be useful, system identification offers an alternative. Measure. Fit. Predict.
But system identification has its own challenges. Finding the right model structure—the right form of equations—is itself a non-trivial task. And fitting the parameters of those equations to data requires optimization algorithms that can navigate high-dimensional spaces without getting stuck in local minima.
The researchers chose a specific optimization method: the Covariance Matrix Adaptation Evolution Strategy, or CMA-ES. This algorithm, developed by Nikolaus Hansen and Andreas Ostermeier in the 1990s, belongs to a class of optimization techniques inspired by biological evolution. It maintains a population of candidate solutions, selects the best performers, and uses statistical properties of the winning solutions to guide the next generation of candidates toward improved fitness.
CMA-ES has proven particularly effective for what engineers call "difficult" optimization problems—those with many local optima, non-smooth landscapes, or poorly understood parameter sensitivities. It requires relatively little hand-tuning, adapts its search strategy based on the problem structure, and has demonstrated success across a remarkable range of applications, from training neural networks to optimizing chemical processes.
For the ship modeling problem, CMA-ES served to estimate the parameters of a linear state-space model from full-scale maneuvering data.
The Model Structure
State-space models represent dynamic systems using two sets of equations. The state equations describe how the system's internal state evolves over time based on current state and external inputs. The output equations describe how the system's measured outputs relate to its internal state.
For ship maneuvering, the state variables typically include the vessel's position, heading, velocity components, and rotational rate. The inputs include thrust commands to propellers, rudder angles, and perhaps thruster forces. The outputs—things you can measure—might include position from GPS, heading from compass, and speed from log or GPS.
The researchers chose a time-invariant, continuous-time linear state-space model. "Time-invariant" means the model structure doesn't change over time. "Continuous-time" means the equations describe motion at every instant, not just at discrete sampling intervals. "Linear" means the relationships between variables are proportional—no squared terms, no products of variables, no sudden discontinuities.
This last choice is the boldest. Low-speed ship dynamics are notoriously non-linear. The forces involved don't scale linearly with speed. Interactions between different motion components are complex. Yet the researchers' results (discussed below) suggest that a linear model, properly identified from data, can capture the essential behavior surprisingly well.
The model was identified using data from actual ship maneuvers—full-scale trials rather than wind-tunnel or towing-tank experiments. This is crucial because the behavior of a real vessel in real port conditions may differ substantially from predictions based on scale-model testing or theoretical calculations. Wind, current, waves, water depth, and the presence of nearby structures all influence low-speed handling in ways that are difficult to capture without direct measurement.
The Research Context
The study emerges from a productive research environment. Co-author Atsuo Maki and colleagues have published extensively on ship maneuvering prediction, including work on neural network approaches to berthing guidance and systematic methods for maneuver model identification from full-scale data. The team combines expertise in maritime engineering, control theory, and applied mathematics—a necessary triangulation for attacking a problem that spans all three domains.
The work also connects to broader efforts in maritime autonomy. The shipping industry faces mounting pressure to reduce emissions, improve safety, and address crew shortages. Autonomous vessels—initially for repetitive open-water tasks like cargo transport between ports, eventually for port operations—represent one vision for addressing these challenges. But port operations, with their tight spaces, complex interactions with infrastructure and personnel, and zero-tolerance safety requirements, remain among the hardest problems in maritime automation.
What They Found
The core result is straightforward to state but represents significant engineering effort: a simple linear model, identified from full-scale maneuvering data using CMA-ES, reproduces ship motion at low speeds with strong agreement to measured data.
The researchers validated their approach by comparing model predictions against held-out data—maneuvers that weren't used to train the model. This is the essential test of any predictive model: can it generalize to situations it hasn't seen before?
The results demonstrate that the identified linear model captures the key features of low-speed maneuvering behavior: the way the ship drifts as it loses headway, the rotational dynamics as it turns in confined waters, the coupling between lateral motion and heading changes. The agreement between model output and empirical data is described as "strong" in the paper's abstract—a characterization that, in the context of system identification for low-speed ship dynamics, represents meaningful progress.
The exact metrics of this agreement—root mean square errors, correlation coefficients, maximum deviations—would appear in the detailed results section of the full paper. The abstract's phrasing suggests performance that is measurably better than baseline approaches, though specific comparisons aren't detailed in the available summary. The comparison is implicit: against more complex non-linear models, against simpler linear models identified through other methods, against purely theoretical predictions from naval architectural calculations.
One particularly notable aspect is the continuous-time formulation. Many system identification approaches work with discrete-time models—equations that describe how state evolves from one sampling instant to the next. But continuous-time models have advantages for control applications: they more naturally represent the underlying physics, they don't introduce artifacts from sampling rate choices, and controller designs developed for continuous-time plants often exhibit better robustness when implemented digitally. By identifying a continuous-time model directly from data, the researchers have produced a representation that's more readily usable by control engineers.
The CMA-ES optimization, applied to full-scale data, enabled parameter estimation that would have been difficult with gradient-based or local search methods. The algorithm's ability to handle the complexity of the parameter landscape—potentially involving interactions between hydrodynamic coefficients, environmental factors, and vessel-specific characteristics—without getting trapped in poor solutions is a significant practical contribution.
Why This Changes Things
The Paradox of Simplification
There is a paradox at the heart of this work that deserves unpacking. Complex systems often require complex models. This is the conventional wisdom, and it's not wrong—modeling turbulent fluid-structure interaction, for instance, demands more detail than modeling a mass on a spring. Yet in many engineering domains, the most successful predictive models are surprisingly simple. F = ma works for everything from falling apples to orbiting satellites, not because the underlying physics is simple, but because it's fundamental. Newton's law captures the invariant relationships; everything else is perturbation.
The ship maneuvering result suggests something similar may be true at low speed. The researchers' linear state-space model is, in a sense, the "mass on a spring" of maritime dynamics—a first-order approximation that captures the dominant dynamics well enough to be useful. Non-linear effects, time-varying behavior, and environmental complexities may introduce corrections, but the linear skeleton explains most of the variance.
This matters for several reasons.
First, simpler models are more robust. A model with 50 parameters, each estimated from limited data, accumulates error at every parameter. A model with 10 parameters has fewer sources of error. When you deploy a model in novel conditions—different wind, different current, different loading configuration—the simple model's predictions may degrade more gracefully than the complex model's, which may fail catastrophically when pushed beyond its training conditions.
Second, simpler models enable simpler controllers. The history of control theory is partly a history of trading model complexity for control complexity. If you have an accurate model, you can design a controller that exploits its structure. If your model is approximate, you can use robust control techniques that guarantee performance despite model uncertainty—but these guarantees often come at the cost of conservatism. A model that captures 80% of the behavior with 20% of the complexity enables a controller that's cheaper to implement, easier to verify, and more predictable in edge cases.
Third, the data-driven identification approach means the model is grounded in reality rather than theory. Naval architects have developed equations for ship resistance and maneuvering based on decades of experiments and experience. These equations encode valuable knowledge. But every ship is different—hull shape varies, loading conditions change displacement and trim, fouling affects drag, and operational conditions deviate from test conditions. A model identified from the specific vessel's behavior, in its actual operating environment, may simply be more accurate for that vessel than a model derived from general principles.
Implications for Maritime Autonomy
The path to autonomous ships runs through several milestones. The first is remote operation—human pilots controlling vessels from shore, with varying degrees of automation handling routine tasks. The second is supervised autonomy—the vessel makes its own decisions but a human overseer can intervene. The third is full autonomy—the ship operates without expectation of human intervention, at least in defined operational domains.
Berthing and unberthing have remained firmly in the human domain for good reason. The consequences of failure—damage to the ship, damage to infrastructure, potential injuries to port workers—make automation both high-value and high-risk. The marginal benefit of autonomous berthing is real: ports operate around the clock, crews don't fatigue, and specialized pilot skills could be replaced by algorithms. But the risk profile demands extraordinary reliability.
The current study advances the foundation for autonomous berthing by demonstrating that low-speed ship dynamics can be captured with models simple enough for control design. This is not the same as demonstrating autonomous berthing itself—that would require not just a model but a complete system including perception, planning, control, and safety monitoring. But a reliable model of ship behavior is a prerequisite.
Consider the analogy to automotive self-parking. Early parking assist systems required expensive sensors and complex path planning. Modern systems can park cars with relatively simple models of vehicle dynamics, because those models have been validated extensively, parameter identification methods have been refined, and control techniques have matured. The problem didn't disappear; it was decomposed, with each sub-problem addressed incrementally.
Autonomous berthing may follow a similar trajectory. As models of low-speed dynamics improve, as parameter identification methods become standardized, as control systems incorporate better representations of environmental factors, the problem becomes tractable piece by piece.
Connections to Broader Autonomy Efforts
The research also connects to larger themes in autonomy and robotics.
System identification from real-world data is central to many modern robotics applications. Robot manipulators learn to compensate for flex and backlash. Legged robots adapt to uneven terrain. Self-driving cars build models of their own dynamics and of how pedestrians and other road users behave. The techniques developed in one domain often transfer to others.
The CMA-ES optimization approach has found applications far beyond ship modeling. It's used to train deep neural networks, to design antennas, to optimize chemical processes, and to tune controller parameters. The fact that it works for low-speed ship maneuvering adds another data point in its growing portfolio of successful applications. For engineers working on similar problems in related domains—autonomous aircraft taxiing, for instance, or autonomous ship-docking for non-commercial vessels—the success here provides evidence that CMA-ES can handle the optimization landscapes these problems present.
The continuous-time modeling approach also reflects a trend in modern control and system identification. As computational resources have grown, it's become feasible to work directly with continuous-time models, which were once dismissed as mathematically inconvenient. Continuous-time formulations connect more naturally to physics, to differential equation solvers, and to control design methods that operate in the time domain rather than in discrete sample-and-hold steps.
What Comes Next
The Road to Autonomous Berthing
The current study demonstrates a proof of concept. The next steps involve extending, hardening, and integrating.
Extension means applying the approach to more ships, more conditions, and more maneuvers. The study used data from specific full-scale maneuvers; robustness of the method will be tested by applying it to vessels of different sizes and hull forms, to operations in ports with different characteristics, and to maneuvers beyond the training distribution—different wind conditions, different tide states, different loading conditions.
Hardening means increasing reliability and reducing conservatism. If the identified model predicts ship behavior within X meters over Y seconds, what happens if we need to guarantee performance within X/2 meters? Can we tighten the bounds through better identification, through adaptive methods that update the model during operation, or through control designs that are robust to larger model uncertainties?
Integration means combining the identified model with perception, planning, and control systems to produce autonomous behavior. The model tells you how the ship will respond to control inputs. But converting desired behavior—"move to this position, facing this direction, by this time"—into appropriate control inputs is its own engineering challenge. Model predictive control, which uses an internal model to simulate many possible control sequences and choose the best, is a natural fit for this problem. The linear model identified here could serve as the internal model in an MPC framework.
Open Questions
The paper's abstract and metadata don't reveal the full scope of challenges that remain. From what's known about low-speed ship maneuvering and autonomous berthing, several open questions emerge.
Environmental uncertainty: The identified model captures ship dynamics given known inputs and measured state. But during berthing, some environmental factors—wind gusts, local currents, the influence of passing traffic—may be unknown or poorly measured. How does model performance degrade under environmental uncertainty? Can the model and controller be designed to be robust to this uncertainty?
Safety-critical verification: For any autonomous system operating in proximity to people and infrastructure, verification is paramount. How do you prove, to the satisfaction of regulators and insurers, that the autonomous berthing system will perform safely? Traditional control system verification methods exist, but applying them to systems with data-identified models requires new techniques.
Graceful degradation: When the model is pushed beyond its reliable range—when conditions differ significantly from training, when sensors provide degraded data, when system components fail—how does the autonomous system respond? The goal is typically a safe fallback rather than catastrophic failure, but designing such behavior for complex maritime operations is challenging.
Generalization vs. specialization: The identified model captures the dynamics of a specific ship, in specific conditions, based on specific data. Is it possible to identify models that generalize across ship types—perhaps parameterized by key hull dimensions and propulsion characteristics? Or is ship-specific identification necessary, as seems likely given the variation in handling characteristics across vessel types? The answer has implications for how autonomous berthing technology could be deployed across a fleet.
Human factors: Berthing operations involve not just the ship but the people directing it—pilots, tugboat crews, line handlers. Autonomous berthing systems must interact with these human operators, whether by taking direction, providing status updates, or coordinating timing. How autonomous systems integrate into human-machine teams remains an open question in maritime and other transportation domains.
The Broader Trajectory
Autonomous ships are no longer science fiction. Several companies have demonstrated autonomous vessel operations in controlled conditions. Regulatory frameworks are beginning to take shape, with classification societies issuing guidelines for autonomous ship design and maritime authorities exploring regulatory sandboxes.
But "autonomous operation in controlled conditions" and "autonomous berthing in a working port" are separated by a substantial gap. The current study narrows that gap incrementally. It shows that a fundamental technical barrier—accurate modeling of low-speed dynamics—may be lower than previously assumed. A simpler model can capture the essential behavior. A practical identification method—CMA-ES applied to full-scale data—can estimate that model. The foundation is being laid.
What remains is the engineering: building systems that exploit this foundation, testing them rigorously, deploying them safely, and integrating them into the complex human system that is a modern port.
The dream of an autonomous ship that glides into port under its own control remains a dream. But it is a dream that has become, incrementally, more plausible.
Conclusion: The Art of Approximation
Science, at its best, is the art of productive approximation. The universe is arbitrary in its complexity—every physical system involves more variables than we can measure, more interactions than we can capture, more nuance than we can resolve. The question is never whether our models are perfect; they're not, and never will be. The question is whether our approximations are good enough for the purpose at hand.
The researchers behind this study have taken an old problem—how ships behave at low speed—and shown that the approximation can be simpler than we thought. A linear model, identified from data, captures the essential dynamics. The complexity that seemed necessary to describe low-speed maneuvering may be complexity born of theoretical habit rather than genuine necessity. Strip it away, and you still have a useful description of how ships drift and turn as they approach the pier.
This matters not because it solves the autonomous berthing problem—it doesn't—but because it changes the shape of the solution space. With simpler models, you can build simpler controllers. With simpler controllers, you can build more robust, more verifiable, more deployable systems. The path from research result to practical application becomes shorter.
The next time you watch a ship ease into its berth, with a pilot's practiced hand on the controls and a tugboat's thrust compensating for wind and current, consider what it would take to automate that task. You need to sense the environment—wind, current, distance to the dock. You need to plan a trajectory that avoids obstacles and satisfies constraints. You need to control the vessel's propulsion and steering to follow that trajectory. And you need a model—a prediction of how the ship will respond to your control inputs—that is accurate enough to keep the vessel on course.
That model now looks simpler than we thought.
What remains is everything else.