Bringing Ocean Robotics to the Surface

Author: Denis Avetisyan

A new underwater research facility aims to accelerate the development and testing of robotics for both maritime and space exploration.

The Marinarium offers a modular, cross-domain platform integrating precise sensing, simulation, and real-world validation for underwater robotics systems.

Bridging the gap between controlled laboratory experiments and the complexities of real-world operation remains a significant challenge in robotics research. This paper introduces the Marinarium, a novel modular underwater research facility designed to address this limitation and facilitate advancements in both maritime and space robotics. By integrating a fully instrumented multi-domain operational volume with a high-fidelity digital twin and a connected space robotics laboratory, the Marinarium enables a streamlined progression from simulation to field testing. Could this integrated approach provide a blueprint for accelerating innovation and deployment in challenging offshore and extraterrestrial environments?

The Persistent Echo of Reality: Bridging the Simulation Gap

The development of dependable underwater robots is significantly hampered by the enduring challenge of transferring skills learned in simulated environments to the complexities of the real ocean – a phenomenon known as the ‘Sim-to-Real Gap’. While simulations offer a safe and cost-effective means of initial design and testing, they inevitably fall short of capturing the full spectrum of hydrodynamic forces, unpredictable currents, and sensor noise present in actual underwater settings. This discrepancy often leads to robotic behaviors that function flawlessly in the virtual world but prove unreliable, or even fail completely, when deployed in real-world scenarios. Bridging this gap requires innovative approaches to simulation fidelity, robust control algorithms capable of adapting to unmodeled disturbances, and techniques for transferring learned policies without catastrophic performance degradation, ultimately enabling robots to navigate and operate effectively in the unpredictable underwater domain.

Conventional control strategies for underwater robots frequently encounter limitations because of the intricate interplay of hydrodynamic forces and unpredictable environmental factors. Unlike terrestrial robots operating in relatively predictable conditions, underwater vehicles are subject to drag, buoyancy, added mass, and complex vortex shedding – effects that are difficult to model with complete accuracy. Furthermore, currents, turbulence, and even the presence of marine life introduce unmodeled disturbances that can significantly degrade performance. These challenges mean that control algorithms designed in simulation, or based on simplified models, often fail to translate effectively to real-world operation, leading to instability, reduced maneuverability, and ultimately, mission failure. Addressing these limitations requires innovative control approaches that are robust to model uncertainty and capable of adapting to dynamic, unpredictable underwater environments.

The pursuit of truly versatile robotics necessitates a departure from systems tailored to single environments; instead, the focus is shifting towards designs capable of consistent performance across air, land, and sea. This ambition extends beyond mere locomotion, demanding adaptable sensing, planning, and control algorithms that account for the drastically different physical properties of each domain. Underwater environments, in particular, present unique challenges – including limited visibility, complex hydrodynamics, and communication constraints – that expose the limitations of robots designed solely for terrestrial operation. Consequently, advancements in multi-domain robotics are increasingly prioritizing shared autonomy frameworks and learning-based approaches that enable robots to generalize their skills and navigate unfamiliar conditions, ultimately paving the way for deployments in diverse and dynamic scenarios, from environmental monitoring to infrastructure inspection and search-and-rescue operations.

The Marinarium: A Controlled Descent into Reality

The Marinarium Facility is designed as a self-contained, reconfigurable environment specifically for the development and testing of robotic systems intended for both underwater and space applications. Its modular construction allows for adaptation to diverse experimental scenarios, including varying tank configurations, lighting conditions, and the inclusion of custom obstacles or payloads. This stand-alone capability minimizes external dependencies and simplifies the logistical challenges associated with large-scale robotics research, enabling repeatable and controlled experimentation independent of ocean or space deployments. The facility supports a range of robotic platforms and sensor suites, facilitating comprehensive performance evaluation and validation in realistic, yet accessible, conditions.

The Marinarium’s Digital Twin is a high-fidelity virtual representation of the physical facility, constructed to facilitate the development and validation of robotic control algorithms in a simulated environment before deployment on physical hardware. This simulation platform allows researchers to address the challenges of transferring algorithms from simulation to the real world – the “sim-to-real gap” – by providing a controlled environment for testing and refining algorithms, reducing risks and costs associated with physical experimentation, and enabling rapid prototyping of robotic behaviors. The Digital Twin incorporates accurate models of the Marinarium’s physical properties, sensor characteristics, and hydrodynamic conditions to maximize the fidelity of the simulation and improve the transferability of algorithms.

The Marinarium’s integrated Motion Capture System (MoCap) utilizes a network of infrared cameras and reflective markers affixed to robotic platforms and objects of interest to provide six-degrees-of-freedom tracking with sub-millimeter accuracy. This system continuously records the precise position and orientation of tracked entities within the Marinarium’s experimental volume. Raw MoCap data undergoes post-processing, including filtering and calibration, to minimize noise and ensure data integrity. The resulting high-fidelity data streams are then used for real-time control, autonomous navigation algorithm development, and rigorous validation of simulation models, effectively quantifying the performance of robotic systems in a controlled physical environment.

Unveiling Dynamics: A Koopman Operator Approach

Traditional system identification techniques, such as linear regression and polynomial fitting, often fail to accurately represent the behavior of underwater vehicles due to the inherent complexities of their operation. These methods typically assume a linear or mildly nonlinear relationship between vehicle inputs and outputs, which is inadequate given the significant nonlinearities introduced by hydrodynamic drag, added mass, Coriolis and centrifugal forces, and actuator dynamics. The six degrees of freedom motion of an underwater vehicle in a three-dimensional space further complicates accurate modeling, as interactions between these degrees of freedom are often nonlinear and difficult to capture with linear models. Consequently, predictions made using these traditional techniques can deviate significantly from actual vehicle performance, limiting the effectiveness of control systems and simulations that rely on accurate dynamic models.

The Koopman Operator provides a data-driven methodology for system identification by transforming nonlinear dynamic systems into infinite-dimensional linear systems. This is achieved through the use of observable functions that, when applied to the original nonlinear system, yield a linear representation governed by a Koopman operator [latex]\mathcal{L}[/latex]. Unlike traditional methods reliant on first-principles modeling or linear approximations, the Koopman operator directly learns the dynamics from observed data, enabling accurate modeling of highly nonlinear behaviors. This approach bypasses the limitations of Taylor series expansions and other local linearization techniques, providing a globally valid representation of the system dynamics and facilitating improved predictive accuracy and control performance. The operator’s spectral properties, specifically its eigenvalues, directly relate to the stability and behavior of the original nonlinear system, offering insights beyond those obtainable through conventional methods.

Implementation of the Koopman operator-based system identification method utilized data gathered from the BlueROV2 Heavy and ATMOS Free-Flyer underwater vehicles to generate a physics-informed model of vehicle dynamics. This approach yielded improvements in simulation accuracy, quantitatively demonstrated by reductions in Root Mean Squared Error (RMSE) of the endpoint error. Specifically, the resulting models exhibited a lower RMSE compared to those generated using traditional system identification techniques when predicting vehicle position over defined trajectories. This reduction in RMSE indicates a more accurate representation of the vehicle’s behavior and improved predictive capability for control system design and validation.

Echoes Across Domains: Implications for Multi-Domain Autonomy

The convergence of precise vehicle dynamics modeling and the unique capabilities of the Marinarium testbed is proving instrumental in advancing control strategies relevant to both underwater and space robotics. This approach leverages the shared principles of physics governing motion in fluid environments – be it water or the near-vacuum of space – to create a unified development framework. By accurately simulating the forces and moments acting on a robotic vehicle, researchers can design and test algorithms in the Marinarium that directly translate to spacecraft autonomy. The testbed’s controlled environment allows for repeatable experiments and rigorous validation of these algorithms, reducing the risk and expense associated with testing in real-world, or space-based, conditions. This cross-domain methodology not only accelerates the development of robust control systems but also unlocks the potential for shared innovation between the fields of ocean and space exploration.

Successfully operating robotic platforms across both underwater and space environments demands a deep comprehension of hydrodynamic forces and the strategic implementation of neutral buoyancy principles. These forces, dramatically different in air versus water, significantly impact vehicle maneuverability and stability; therefore, replicating gravitational effects through neutral buoyancy – where an object neither floats nor sinks – allows for terrestrial testing that accurately simulates the operational conditions of space. This approach facilitates the development of control algorithms and hardware designs that can seamlessly transition between environments without requiring costly and complex in-space validation. By meticulously accounting for drag, lift, and other fluid dynamics, and by leveraging neutral buoyancy to equalize gravitational loads, researchers can effectively use underwater testbeds as robust and reliable surrogates for validating spacecraft autonomy systems, reducing risk and accelerating the development of multi-domain robotic capabilities.

Recent research demonstrates the superior performance of Koopman operator-based system identification in accurately predicting the state of complex robotic vehicles. This modeling approach significantly outperformed traditional methods when applied to both an underwater BlueROV and the ATMOS free-flyer, a spacecraft autonomy testbed. Notably, the vehicles, despite operating in vastly different environments, exhibited remarkably similar trajectory profiles when controlled using models derived from this technique. This correspondence validates the concept of utilizing underwater surrogates – like the BlueROV – for the development and validation of spacecraft autonomy systems, offering a cost-effective and accessible alternative to expensive and logistically challenging space-based testing. The ability to translate control strategies between these platforms represents a significant step towards achieving true multi-domain autonomy for robotics.

The Marinarium, as detailed in this research, embodies a proactive approach to system longevity, acknowledging that even the most rigorously designed robotic systems are subject to the inevitable pressures of real-world deployment. It doesn’t seek to prevent decay, but to understand and mitigate its effects through meticulous system identification and simulation-to-real transfer. As Alan Turing observed, “Sometimes people who are unhappy tend to look at the world as if there were nothing else.” This sentiment resonates with the Marinarium’s purpose: to provide a controlled, yet realistic, environment for observing and addressing the inevitable ‘unhappiness’ – the challenges and failures – inherent in complex robotic systems before they manifest in critical, open-water scenarios. The facility offers a means of proactively understanding the limits of these systems, ensuring they age gracefully and maintain functionality even under stress.

What Lies Beneath?

The Marinarium, as presented, is not merely a facility, but an acknowledgement. An acknowledgement that robotics, particularly when venturing into complex environments, doesn’t so much ‘solve’ problems as learn to inhabit them. The pursuit of seamless simulation-to-real transfer often feels like an attempt to arrest entropy, to create a perfect, unchanging echo of the world. Yet, systems learn to age gracefully; the true metric isn’t minimizing the gap between model and reality, but understanding how the divergence itself informs adaptation.

Future iterations of such facilities will likely focus less on replicating conditions, and more on instrumenting the process of adaptation. Detailed study of system identification as it unfolds in a dynamic, imperfect environment will yield more robust designs than any amount of pre-programmed perfection. The cross-domain validation with space robotics is a particularly insightful direction – the constraints differ, but the fundamental challenge of operating within an unyielding physical reality remains constant.

Sometimes observing the process is better than trying to speed it up. The Marinarium offers a valuable platform for that observation, a space where the slow, inevitable decay of idealized models can be meticulously charted. The question isn’t whether a system can avoid change, but how it responds when change becomes unavoidable.

Original article: https://arxiv.org/pdf/2602.23053.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

The Persistent Echo of Reality: Bridging the Simulation Gap

The Marinarium: A Controlled Descent into Reality

Unveiling Dynamics: A Koopman Operator Approach

Echoes Across Domains: Implications for Multi-Domain Autonomy

What Lies Beneath?

See also: