Source-linked AI summary
Beyond Shallow-Water Photorealism: Physically and Sensor-Grounded Simulation for Deep-Sea Robotics
Michele Grimaldi, Enrico Di Maria, Ignacio Carlucho, Yvan R. Petillot
TL;DR
Underwater simulators often emphasize shallow-water visual realism while omitting deep-sea physical and sensor effects. This work extends Stonefish with integrated stochastic sensing, hydrodynamics, terramechanics, environmental variability, and underwater optics for real-time deep-sea simulation. The reported IMU and DVL evaluations show stochastic behavior and long-horizon drift that more closely match real sensor characteristics.
Problem
Existing underwater simulators provide limited support for deep-sea sensing, environmental interactions, and vehicle types, although deep-sea missions require high-fidelity simulation for validation and mission readiness.
Method
The paper develops a physics- and sensor-grounded, ROS-compatible Stonefish extension combining stochastic IMU and DVL models, magnetometer disturbances, higher-order hydrodynamics, terramechanics, pressure variation, and underwater optics.
Results
The proposed IMU model reduces gyroscope ARW to approximately 5 × 10^-4 rad/s and exhibits bias instability in the 10^-5–10^-6 rad/s range, while the DVL model reaches 0.56 m/h drift over three hours.
Takeaways & Limitations
The framework provides a practical foundation for navigation, perception, and autonomy research under more representative deep-sea operating conditions.
Abstract
from arXiv · showhide
Many recent underwater simulators emphasize visual realism at the expense of physical fidelity, focusing on shallow-water effects with limited relevance in deep-water environments and high computational cost. In this work, we shift the focus toward deep-sea physical and sensor realism. We present a physics- and sensor-grounded extension of the Stonefish simulator that augments its hydrodynamic models with stochastic IMU and DVL drift, magnetometer disturbances, higher-order hydrodynamics, terramechanics, pressure-driven environmental variability, and physically based underwater optics. These additions are designed to better capture the forces and measurements shaping the behavior of deep-ocean AUVs, ROVs, landers, ASVs, and gliders, while remaining compatible with real-time simulation. This work advances underwater simulation toward more representative deep-sea operating conditions, which is particularly relevant for long-duration navigation and learning-based autonomy, where inaccurate sensor and environmental models introduce non-physical artifacts and overly optimistic performance. While challenges remain, including complex fluid-structure interactions and full environmental stochasticity, the proposed framework provides a practical foundation for navigation, perception, and autonomy research under deep-sea conditions.
I. INTRODUCTION
Existing underwater simulators often prioritize shallow-water visual realism while omitting deep-sea sensing, environmental, and vehicle-interaction effects. The extended Stonefish framework addresses these gaps with integrated physics, sensing, optics, and ROS-compatible real-time simulation.
- Deep-sea simulation requires physics-based modeling because visual effects alone do not represent the forces and measurements shaping robotic behavior.
- The extended Stonefish framework integrates optics, physics, and sensing within a ROS-compatible architecture for deep-sea navigation, contact-rich interaction, and learning-based autonomy studies.
- Its stochastic sensor framework models IMU, DVL, and magnetometer effects including drift, per-beam velocity measurements, and hard- and soft-iron disturbances.
- The environmental model captures lift, Magnus, and Coriolis effects, density variation, and depth-dependent hydrodynamic coefficients.
- A lightweight Bekker-based terramechanics model enables seabed crawler deployment in Stonefish.
- The camera and rendering pipeline provides configurable intrinsics, parametric lens distortion, and deep-sea optical effects.
- Existing platforms commonly provide limited deep-water optical degradation, sensor drift, and long-term environmental dynamics despite supporting visual or ROS-integrated simulation.
III. IMU NOISE AND BIAS MODELING
The proposed IMU model combines multiple sensor imperfections and stochastic processes to reproduce realistic short- and long-term behavior while remaining parametrizable for real-time simulation.
- The unified stochastic IMU model captures short-term noise and long-term bias dynamics for long-duration underwater navigation.
- Scale errors and cross-axis misalignment transform true angular velocity and linear acceleration before measurement generation.
- Stochastic gyroscope and accelerometer biases are modeled as Brownian motion, while temporally correlated noise uses a first-order Gauss–Markov process.
- Measurements include quantization, range clipping, temperature-dependent drift, vibration-induced components, and additive white noise.
- The IMU signal flow applies scale, bias, correlated noise, temperature, and vibration effects to true motion before producing measured outputs.
- Allan deviation compares old, new, and JIMS-80 IMU behavior across averaging time for gyroscope and accelerometer measurements.
- All noise and drift parameters are fully parametrizable and tunable to target sensor specifications.
A. Allan Variance Test
The Allan variance analysis compares real, original-simulator, and proposed IMU data over more than 9 hours to assess long-horizon stochastic behavior. The proposed model reproduces both short-term noise and long-term bias drift more realistically than the original model.
- The evaluation uses over 9 hours of data from a real JIMS-80 IMU, the original Stonefish model, and the proposed model.
- Allan variance separates short-term noise from slow bias drift and long-term rate random walk.
- The original model is dominated by white noise, with gyroscope ARW of approximately 1.6 × 10^-3 rad/s and negligible bias instability of ∼3 × 10^-5 rad/s.
- The proposed model reduces gyroscope ARW to ∼5 × 10^-4 rad/s and exhibits bias instability in the 10^-5–10^-6 rad/s range.
- The proposed model more closely matches the real JIMS-80 by reproducing short-term stochastic noise and long-term bias drift absent from the original simulator.
IV. DVL DRIFT AND BIAS
The extended DVL model represents velocity relative to the seabed and water currents, with per-beam outputs and a slowly varying bias for realistic long-horizon drift. Over three hours, it produces monotonic drift growth unlike the original white-noise model.
- The extended DVL provides three-axis velocity measurements relative to the seabed and water currents, including per-beam outputs.
- The model projects vehicle velocity along four tilted acoustic beams and introduces measurement noise and a slowly varying bias vector.
- Aggregate velocity and beam-level Doppler measurements are generated consistently.
- Over 3 hours, the proposed bias-driven model accumulates monotonic drift of 0.56 m/h, compared with 0.21 m/h for the original white-noise model.
- Both models have near-zero mean velocities and comparable short-term noise, while their long-horizon drift behavior differs substantially.
V. MAGNETOMETER DISTURBANCE MODELING
The simulator models magnetometer measurements using geomagnetic variation, vehicle and infrastructure disturbances, sensor imperfections, and sensor-frame orientation. It evaluates heading fidelity against a simplified fixed-reference model during a 3D AUV trajectory at 45° latitude.
- Magnetometer model: The magnetometer model combines geomagnetic variation, magnetic disturbances, sensor imperfections, and sensor-frame rotation to produce three-axis measurements.Hard-iron effects are modeled as magnetic dipoles, while soft-iron effects use a linear distortion matrix.
- Magnetometer model: Hard-iron disturbances are represented by magnetic dipoles, whereas soft-iron effects are represented by a linear distortion matrix.
- Heading estimation: Heading is extracted from the measured field’s horizontal projection, making unmodeled inclination or anisotropic distortion a source of systematic, orientation-dependent yaw errors.
- Evaluation: The proposed model is compared with a simplified fixed-reference magnetometer using a tilt-compensated yaw estimator along a 3D AUV trajectory at 45° latitude.Figure 5 reports heading estimates on the left and heading error on the right.
- Integrated dynamics: The broader simulation framework evaluates magnetometer behavior alongside surface-based hydrodynamics, depth-dependent environmental properties, lift, Magnus effects, and Coriolis forces.
A. Seabed Interaction and Terramechanics Modeling
The simulator models seabed contact through surface-based terramechanics, including penetration, pressure–sinkage, shear resistance, and damping. The resulting forces support smooth transitions between free-flowing and grounded states while preserving real-time compatibility.
- Contact model: Seabed contact is modeled per intersecting mesh face using penetration depth, terrain normal, pressure–sinkage, tangential resistance, and damping.
- Contact model: Tangential resistance is limited by soil shear strength determined by cohesion, internal friction angle, and local normal pressure.Exceeding the soil’s sustainable tangential stress causes plastic yielding.
- Validation: Figure 8 shows longitudinal traction initially exceeding resistance during acceleration before the forces balance at steady velocity.
- Contact model: Velocity-proportional normal damping provides dissipative contact behavior and numerical stability during seabed interaction.
- Simulation behavior: The shared per-face force accumulation framework enables smooth transitions between free-flowing and grounded vehicle states without discrete contact modes.
VII. UNDERWATER 3D SIMULATION AND RENDERING
Stonefish provides a modular framework for importing 3D models, simulating cameras, and rendering deep-sea environments. These capabilities support underwater robotics experiments within the simulator.
- Simulation framework: The framework supports importing 3D models for underwater robotics experiments.
- Simulation framework: The framework includes camera simulation for underwater robotics experiments.
- Simulation framework: The framework supports rendering deep-sea environments.
A. GLTF/GLB Model Import and Optimization
The simulator imports GLTF and GLB models by parsing scene hierarchies and assembling mesh primitives into internal structures. A single-pass loading strategy supports efficient handling of large scenes, while full PBR materials remain future work.
- Model import: The importer supports GLTF 1 and GLB formats, parsing either ASCII GLTF or binary GLB files.
- Model import: Scene hierarchies are traversed to extract vertices, normals, and optional UV coordinates into scaled internal mesh structures.
- Optimization: Single-pass processing reduces memory allocations and improves cache coherence when importing models containing millions of triangles.
- Performance: Complex scenes can be imported in a matter of seconds, enabling fast iteration and testing within the simulation environment.
- Scope: Full PBR material support is planned for future work.
B. Parametric Underwater Camera Modelling
Stonefish models configurable underwater cameras through projection, lens-distortion, and pixel-coordinate transformations for real-time calibrated imagery. The formulation is an efficient approximation rather than a complete physical model of underwater camera housings.
- Stonefish supports rectilinear, fisheye, and wide-angle cameras with adjustable field of view, intrinsics, and image-plane distortion.
- For normalized coordinates, radial distortion is based on r2 = x2 + y2 and radial coefficients such as k1.
- Distorted coordinates are converted to pixels using focal lengths fx and fy and principal point (cx, cy).
- The transformations run in OpenGL, enabling real-time calibrated barrel, pincushion, and tangential distortion.
- The camera model omits light propagation through water, housing glass, and air, as well as flat- and dome-port refraction.
C. Deep-Sea Rendering Integration
Stonefish is connected to the DeepSea renderer through the ROS image pipeline to approximate deep-sea optical conditions while retaining the simulator’s vehicle and scene behavior. The modular arrangement makes enhanced imagery available to existing perception and navigation algorithms.
- DeepSea processes Stonefish RGB and depth images to synthesize artificial illumination, scattering, and radiometric attenuation.
- The integration addresses deep-sea imagery dominated by artificial light, attenuation, backscatter, and illumination-dependent colour shifts.
- Processed images are republished as enhanced camera streams while preserving vehicle dynamics, sensor timing, and scene geometry.
VIII. CONCLUSION
The work argues for underwater simulation centered on physical and sensor grounding rather than visual fidelity alone. It positions this emphasis as a practical basis for broader vehicle support, autonomy research, and future deep-sea digital twins.
- The proposed shift prioritizes forces, measurements, and environmental conditions that directly influence robotic behavior and autonomy.
- Fig. 10 compares loading times across GLTF, GLB, and OBJ formats at lower and higher triangle counts.
- The framework is intended to support crawlers, landers, surface vehicles, and hybrid platforms across diverse mission profiles.
- Its supported priorities include navigation robustness, control stability, perception under low signal-to-noise conditions, and long-duration autonomy.
- The physics- and sensor-centric approach provides a foundation for future deep-sea digital twin applications incorporating vehicle, operational, and environmental models.