💻 computer science

Marinarium: A Modular Experimental Facility for Reproducible Maritime and Space-Analog Field Robotics

This paper presents Marinarium, a modular and instrumented experimental facility that bridges the gap between low-fidelity simulation and costly offshore field trials by enabling reproducible, multi-domain robotics research for both maritime and space-analog applications through integrated sensing, a retractable roof, and a digital twin.

Original authors: Ignacio Torroba, David Dorner, Victor Nan Fernandez-Ayala, Mart Kartasev, Joris Verhagen, Elias Krantz, Gregorio Marchesini, Carl Ljung, Pedro Roque, Chelsea Sidrane, Linda Van der Spaa, Nicola De Car

Published 2026-07-17

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Ignacio Torroba, David Dorner, Victor Nan Fernandez-Ayala, Mart Kartasev, Joris Verhagen, Elias Krantz, Gregorio Marchesini, Carl Ljung, Pedro Roque, Chelsea Sidrane, Linda Van der Spaa, Nicola De Carli, Petter Ogren, Christer Fuglesang, Jana Tumova, Dimos V. Dimarogonas, Ivan Stenius

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to teach a robot how to swim or how to float in space. You can't just throw it into the ocean or launch it into the sky immediately; the risks are too high, the costs are astronomical, and the weather is too unpredictable. So, scientists usually try two things: they build a computer simulation (a video game version of reality) or they test in a small, controlled water tank. But there's a problem. Computer games are great, but they often miss the messy, wobbly details of real physics. Small water tanks are safe, but they are often too simple, lacking the sensors and space to test complex teamwork between robots. It's like trying to learn to drive a race car by playing a video game, then suddenly jumping into a tiny bathtub with a toy car. You need a middle ground—a place that feels real enough to teach the robot, but safe enough to let it crash without breaking the bank. This is the challenge of "field robotics," the science of building machines that can operate in the wild, from the deep ocean to the vacuum of space.

This paper introduces a solution called Marinarium, a brand-new, modular research facility built at the Royal Institute of Technology in Stockholm. Think of it as a "robot gym" that bridges the gap between a video game and the real world. The facility is a large, 9-by-5-meter water tank (about the size of a small swimming pool) that sits right next to a control room. But it's not just a pool; it's a high-tech playground. It has a retractable roof, meaning researchers can test robots in the water while they are also exposed to real wind and rain above the surface. It is covered in high-speed cameras (Motion Capture systems) that track every single movement of robots underwater and in the air with pinpoint accuracy. Most importantly, it has a "digital twin"—a perfect, real-time video game copy of the tank that talks directly to the real robots.

The researchers used this facility to prove that it works for four different types of tricky robot problems. First, they taught a robot how to understand its own movement by collecting massive amounts of data in the tank, creating a much better "brain" for the robot than standard computer models. Second, they tested a team of three different robots—a drone flying above, a boat on the surface, and a submersible underwater—working together to meet up, showing that the facility can handle complex teamwork. Third, they used the tank to fix the "glitches" in their computer simulations, teaching the simulation to act more like the real robot by learning from the mistakes made in the water. Finally, they used the underwater robots as stand-ins for spacecraft, testing space navigation software in the water to see if it would work in zero gravity. The paper suggests that this kind of facility makes robot development faster, cheaper, and safer, allowing scientists to solve problems in the lab that would otherwise require dangerous and expensive trips to the ocean or space.

Technical Summary: Marinarium: A Modular Experimental Facility for Reproducible Maritime and Space-Analog Field Robotics

Problem Statement
Field robotics research in maritime and space domains faces a persistent "gap" between low-cost, low-fidelity simulations and high-cost, high-risk offshore or orbital deployments. While open-water testing offers realism, it suffers from poor observability, weather-dependent repeatability, and difficulty in acquiring accurate ground truth for machine learning. Conversely, desktop simulations often fail to capture complex hydrodynamic, acoustic, and electromagnetic interactions, leading to a significant "Sim2Real" gap that hinders the deployment of simulation-trained policies. Existing water tank facilities often lack the necessary instrumentation, multi-domain capabilities (air, surface, underwater), and tight integration with digital twins or space robotics laboratories to support systematic research in system identification, heterogeneous collaboration, and cross-domain validation.

Methodology: The Marinarium Facility
To address these limitations, the authors present Marinarium, a modular, standalone experimental facility at the Royal Institute of Technology (KTH). The facility is designed as a cost-effective intermediate testbed bridging simulation and field deployment. Its core components include:

Physical Infrastructure: A $9 \times 5 \times 3$ m water basin with a $9 \times 5 \times 2$ m aerial volume above the surface. It features a fully retractable roof to allow for both sheltered and open-air (real weather) operations.
Instrumentation: The facility is equipped with two Motion Capture (MoCap) systems: a Qualisys system (8 Arqus cameras) for underwater tracking and an Optitrack system (up to 16 Flex 13 cameras) for aerial tracking. This enables full 6-DoF state estimation with ground truth.
Acoustic Optimization: To mitigate multipath effects common in water tanks, the basin floor is covered with 3 tons of 8–16 mm sea gravel, enabling acoustic communication (e.g., 70% transmission rate with Delphis modems) that is otherwise impossible in smooth-walled tanks.
Digital Twin: A digital replica of the facility is implemented in the SMaRCSim simulator, connected to the physical infrastructure via a ROS 2-Unity bridge. This allows for real-time state synchronization and Software-in-the-Loop (SIL) testing.
Space Robotics Integration: The facility is physically adjacent to the KTH space robotics laboratory (housing the ATMOS air-bearing testbed). Both facilities share a ROS 2 domain and utilize identical software stacks (PX4/Pixhawk on BlueROV2 Heavy and ATMOS free-flyers), enabling direct cross-domain comparison.

Key Contributions and Experimental Validation
The paper validates the facility's utility through four representative studies in field robotics:

Data-Driven System Identification (SYSID):
- Method: The authors collected high-fidelity datasets from a BlueROV2 in the tank using MoCap ground truth. They compared four modeling approaches: a Double Integrator (DI), a standard Fossen physics model, a Physics-informed Neural Network (PINc), and a Koopman operator-based model (EDMDc) using Radial Basis Functions (RBF).
- Result: The Koopman EDMDc model achieved the lowest Root Mean Square Error (RMSE) across 1, 10, and 100-step prediction horizons, outperforming both the physics-based baseline and the PINc model. The study demonstrates that the controlled environment enables the collection of dense, low-noise data required for operator-theoretic system identification.
Heterogeneous Multi-Domain Robotics:
- Method: A rendezvous mission was executed involving a heterogeneous fleet: an underwater AUV (SAM), a surface USV (FloatSAM), and an aerial UAV (Crazyflie).
- Result: The mission successfully demonstrated cross-domain communication (WiFi, 5G, and acoustic links) and coordinated behavior where the UAV landed on the USV while the AUV docked underneath. This validates the facility as a viable arena for stress-testing multi-robot fleets before offshore deployment.
Sim2Real Gap Bridging:
- Method: Using the digital twin in SMaRCSim, the authors trained a Multilayer Perceptron (MLP) to learn the residual dynamics ( $\delta$ ) between the simulated vehicle and the real vehicle (ground truth from MoCap). This residual was injected into the simulator to correct the physics model.
- Result: The residual-corrected simulator showed significant improvements in position and linear velocity prediction errors over long time horizons compared to the uncorrected simulation. However, the authors note that angular velocity errors remained significant, and instability occurred during high-acceleration maneuvers, indicating areas for future robustness improvements.
Spacecraft Autonomy Validation:
- Method: Paired experiments were conducted where an underwater BlueROV2 (acting as a neutrally buoyant surrogate) and the planar ATMOS air-bearing testbed executed the same autonomous inspection task defined by Signal Temporal Logic (STL) under an identical control architecture (Nonlinear MPC).
- Result: Both platforms successfully tracked the reference trajectory with similar motion profiles. The underwater surrogate successfully reproduced the behavior of the free-flying spacecraft, validating the use of underwater surrogates for testing spacecraft guidance and control algorithms.

Significance and Claims
The paper claims that Marinarium provides a unique, reproducible, and instrumented environment that narrows the gap between simulation and reality. Its significance lies in:

Democratization: Offering a low-cost, modular alternative to massive, scarce offshore facilities, making rigorous field robotics research accessible to academic groups.
Cross-Domain Synergy: Being the only known facility to combine underwater and above-water MoCap, a retractable roof for weather exposure, a published digital twin workflow, and tight integration with a space robotics laboratory.
Methodological Advancement: Demonstrating that controlled tank environments are essential for data-driven system identification and Sim2Real transfer, which are often impractical or impossible to achieve with open-water testing or simulation alone.

The authors conclude that while the facility does not replace open-water or orbital testing, it serves as a critical intermediate step that reduces development risk, accelerates iteration, and enables the validation of complex autonomy stacks across maritime and space-analog domains.

More like this