Source-linked AI summary
Advancing Accessible Underwater Robotics: The Mini-Girona I-AUV at RAMI 2025
Taqi Hamoda, Bilal Ahmed, Deborah Ele-Ojo, Thi Tran Ha Bao, Adel Saidani, Mazen Elgabalawy, Alaaeddine Chaarani, Sebastian Realpe, Patryk Cieslak, Pere Ridao, Narcis Palomeras, Nuno Gracias
TL;DR
Underwater robotics development is constrained by high costs, limited accessibility, and platforms that are poorly suited to complex close-proximity tasks. This paper presents the affordable, usable Mini-Girona I-AUV, validates its algorithms through simulation and pool testing, and reports second place overall at RAMI 2025 despite operational challenges.
Problem
High costs and limited accessibility restrict underwater robotics development, while traditional underactuated AUV designs complicate control in complex environments.
Method
The paper presents the Mini-Girona I-AUV’s design, autonomous-operation algorithms, stereo-based perception pipeline, and validation through Stonefish simulation and real-world pool testing.
Results
Second place overall at RAMI 2025, with second-place finishes in the Grand Challenge and Plant Intervention and third place in Plant Inspection.
Takeaways & Limitations
The Mini-Girona demonstrates potential as an accessible platform for underwater robotics research and experimentation in real-world environments.
Takeaways & Limitations
High ambient temperatures caused internal electronics to overheat, while NATO access restrictions limited venue entry to three of eight team members.
Abstract
from arXiv · showhide
The Mini-Girona Intervention Autonomous Underwater Vehicle (I-AUV) represents an advancement in accessible underwater robotics, designed to bridge the gap between costly, specialized research AUVs and basic Remotely Operated Vehicles (ROVs). Developed with a focus on affordability and usability, the Mini-Girona, priced at approximately $50,000, integrates advanced components such as a 5-DOF manipulator arm, stereo vision, and AI-driven processing for autonomous navigation and intervention tasks. This paper presents the design and development of the Mini-Girona, detailing its performance during the RAMI 2025 student competition. Despite challenges such as thermal management issues and restricted team access, the Mini-Girona achieved second place overall, excelling in vision-based perception and intervention tasks. This work highlights the platform's potential as a tool for underwater robotics research and education, fostering innovation in real-world underwater applications.
I. INTRODUCTION
Underwater robotics remains constrained by high platform costs and limited accessibility, motivating more capable affordable systems. The Mini-Girona was developed to address this need and was evaluated through its design, autonomy, and RAMI 2025 performance.
- High-performance intervention AUVs can cost hundreds of thousands of dollars, restricting advanced underwater research to well-funded institutions.
- Recent AUV designs increasingly prioritize agility, hovering, and subsea residency for precise close-proximity tasks over speed and long-distance cruising.
- The Mini-Girona I-AUV was developed to provide an accessible platform for underwater research and experimentation.
- Its design philosophy emphasizes affordability and usability while retaining essential functionality, bridging basic ROVs and specialized research AUVs.
- The publication presents the Mini-Girona’s design, development, autonomous-operation algorithms, and evaluation in the RAMI 2025 competition.
II. THE MINI-GIRONA
The Mini-Girona is a lower-cost research, inspection, and intervention AUV built around a manipulator, stereo perception, onboard processing, and layered autonomy. Its COLA2 architecture organizes control from reactive behaviors through mission-level planning.
- The Mini-Girona costs approximately $50,000, compared with $400,000 for the Girona 500, to broaden underwater robotics access.
- A 5-DOF manipulator, stereo cameras, DVL, imaging sonar, GPS, Jetson AI processing, and an Intel NUC support intervention, perception, navigation, and control.
- COLA2 provides a hierarchical autonomy architecture with Reactive, Execution, and Mission layers.
- Reactive Layer: The Reactive Layer manages real-time sensor and actuator interactions and executes fundamental robot behaviors.
- Execution Layer: The Execution Layer translates high-level mission plans into low-level commands for the Reactive Layer.
- Mission Layer: The Mission Layer defines high-level goals and supports predefined plans, dynamic mission management, and real-time replanning.
III. THE RAMI COMPETITION
Underwater robotics competitions provide practical exposure to environmental and operational complexities that academic curricula may not cover. They also support technology advancement and expertise development.
- Underwater robotics competitions advance technology and develop expertise through practical experience with real-world complexities.
- Variable salinity and turbidity are examples of underwater conditions often not covered in academic curricula.
- SAUC-E, Europe’s first underwater robotics competition, launched in 2006 and drew inspiration from the 1997 US RoboSub.
A. RAMI Competition Structure
RAMI 2025 evaluates AUVs in realistic infrastructure-inspection and maintenance scenarios involving navigation, sensing, detection, and manipulation. Its benchmark missions range from inspection to intervention and combined emergency response.
- RAMI is an international competition focused on autonomous underwater robotics for critical-infrastructure inspection and maintenance.
- RAMI 2025 evaluated autonomous navigation, data acquisition, object detection, and manipulation in a 50×25 m area 4–6 m deep with 1–3 m visibility.
- TBM-1: Pipeline Area Inspection: TBM-1 required waypoint navigation, colored-buoy mapping, gate traversal, and inspection of a designated structure.
- TBM-2: Intervention on Pipeline Structure: TBM-2 required pipeline following, damage-marker detection and reporting, a 90° valve turn, one-minute contact, and ring-pole retrieval.
- TBM-3: Comprehensive I&M Mission: TBM-3 combined TBM-1 and TBM-2 into an autonomous or semi-autonomous inspection, localization, and intervention mission.
- Prior competition teams demonstrated varied experience, including podium results and object-detection performance in underwater robotics challenges.
IV. METHODOLOGY
The Mini-Girona algorithms were validated first in a physics-based simulator and then in the CIRS pool using components that replicated the RAMI competition environment.
- Validation setup: Stonefish replicated the Mini-Girona, its onboard sensors, and the RAMI competition scenario for pre-deployment algorithm validation.The model included the robot’s onboard sensors and competition-specific environment.
- Validation setup: Real-world validation followed in the CIRS pool with buoys, a pipe, and a custom-built console arranged to mimic the competition environment.The pool tests included the robot’s DVL, stereo cameras, and safety equipment.
- Validation setup: The tested hardware and algorithms were evaluated on scenario components after extensive pool testing.The evaluation followed hardware checks involving the DVL, stereo cameras, and safety equipment.
A. Perception
The perception and localization pipeline combines stereo reconstruction, learned detection and semantic processing, and sensor fusion for underwater scene understanding and tracking.
- Perception: Stereo block matching, bilateral filtering, and DBSCAN produce a denoised point cloud from rectified underwater images.The pipeline preserves structural details while removing sparse noise.
- Perception: YOLO-E v11 performs real-time object detection and segmentation, while PCA localizes valve handles and guides pipe following.PCA identifies the handle and valve center and derives navigational points along a pipe’s principal axis.
- Perception: LLaVA processes cropped YOLO-E v11 regions for precise number and color recognition and enhanced scene understanding.The model is applied after detection and segmentation.
- Localization: The localization framework fuses DVL, pressure, compass, and sonar data with a constant-velocity state model for robust tracking.The state includes position, orientation, and velocities, while environmental features are aligned with vehicle orientation.
- Localization: An Extended Kalman Filter fuses compass, depth, DVL, and inertial measurements while associating sonar features through a statistical distance metric.Sonar range and bearing are converted to Cartesian coordinates, and line features are transformed into the vehicle frame.
- Localization: GPS alignment overlays the sonar-derived acoustic map onto satellite imagery of the CMRE marina.The map was generated from sonar data collected during TBM 3.
C. Autonomous Navigation and Planning
Autonomous navigation combines sonar-based mapping, reactive replanning, camera–sonar fusion, and task-priority control to support obstacle avoidance, inspection, and intervention.
- Autonomous Navigation and Planning: RRT Connect plans waypoint paths on a 2D projection of 3D voxel data aggregated from onboard Miniking sonar.A proportional controller converts localization and target errors into velocity setpoints for COLA2 thruster control.
- Autonomous Navigation and Planning: Newly detected obstacles trigger replanning from the AUV’s current position to avoid collisions during navigation.The reactive update supplements the initial sonar-derived path.
- Autonomous Navigation and Planning: Camera-based point clouds are fused with sonar data for pipeline navigation and 3D reconstruction where sonar cannot reliably detect pipes.The AUV maintains a constant, slow speed and remains oriented toward the pipeline.
- Manipulation and Intervention: Controller switching supports goal-based and reactive behaviors, transitioning between predefined joint motions and dynamic task-priority interventions.Joint trajectory control executes configurations, while joint velocity control handles task-priority execution.
- Manipulation and Intervention: The end-effector task vector represents pose as a function of joint configuration, and task error measures position and orientation discrepancies.Orientation is represented through quaternion components.
- Manipulation and Intervention: The Jacobian maps joint velocities to Cartesian velocities, enabling task-priority control of end-effector motion.This relation supports accurate end-effector guidance.
V. RESULTS
The Mini-Girona achieved second place overall at RAMI 2025, with especially strong vision-based perception and intervention performance reflected in multiple task awards.
- Competition outcomes: Second place overall was achieved at RAMI 2025 despite the vehicle becoming operational only three weeks before the competition.The competition tested inspection and intervention capabilities in a challenging underwater environment.
- Competition outcomes: TBM 3 - Grand Challenge: 2nd place.This was one of the team’s listed competition awards.
- Competition outcomes: TBM 2 - Plant Intervention: 2nd place.The award reflects performance in an intervention task.
- Competition outcomes: TBM 1 - Plant Inspection: 3rd place.The award reflects performance in an inspection task.
A. AUV Performance
During RAMI 2025, the Mini-Girona demonstrated integrated perception, navigation, and intervention capabilities despite early thermal and communication failures.
- A 75×50 m acoustic map supported spatial awareness, navigation, and self-localization in TBM 3.
- Stereo processing generated disparity maps and 3D depth information for navigation, obstacle avoidance, pipe following, and object interaction.
- YOLO-E v11 successfully identified and localized buoys, pipe markers, the valve, and the ring pole.
- The team completed TBM 2 intervention by turning the valve, maintaining pipe-marker contact, and surfacing with the grasped pole.
- High ambient temperatures disrupted STM-board and sonar communication, impairing autonomous navigation and data acquisition during the initial two days.
- Despite these obstacles, the platform’s performance enabled a high overall ranking and multiple awards.
VI. CONCLUSION AND FUTURE WORK
The RAMI 2025 deployment demonstrated the Mini-Girona’s performance and adaptability while identifying perception, autonomy, localization, and sensor fusion as future development priorities.
- Despite its recent operational status and competition challenges, the Mini-Girona secured second place overall and several prestigious awards.
- Future work targets improved deep-learning perception for more reliable 3D reconstruction and full interventional autonomy.
- The team plans to integrate vision into localization for more precise navigation.
- Additional sensor combinations, including monocular cameras with front-facing sonar, will support testing of advanced sensor-fusion algorithms.