Vision-based underwater robots can be useful in inspecting and exploring confined spaces where traditional sensors and preplanned paths cannot be followed. Sensor noise and situational change can cause significant uncertainty in environmental representation. Thus, this paper explores how to represent mapping inconsistency in vision-based sensing and incorporate depth estimation confidence into the mapping framework. The scene depth and the confidence are estimated using the RAFT-Stereo model and are integrated into a voxel-based mapping framework, Voxblox. Improvements in the existing Voxblox weight calculation and update mechanism are also proposed. Finally, a qualitative analysis of the proposed method is performed in a confined pool and in a pier in the Trondheim fjord. Experiments using an underwater robot demonstrated the change in uncertainty in the visualization.
Indoor experiments were conducted at NTNU's MC-lab in a 40 m × 6.45 m × 1.5 m water tank. A custom BlueROV2 Heavy with stereo camera and IMU was manually piloted along varied trajectories, covering over half the tank while collecting synchronized data.
The field experiments were conducted at a pier in the Trondheim Fjord, at a depth of approximately 7 meters. During the trials, the robot descended to a depth of about 0.7 meters and traveled a distance of 20 to 25 meters over a period of 12 minutes.
A modified HoloOcean simulator was used to explore the OpenWater environment, with the HoveringAUV agent controlled via a PID controller for position, velocity, and attitude tracking across 4 DOFs. Confidence was emulated instead of using RAFT-Stereo to enable faster processing in simulation. Our modified Voxblox was used for visualization voxel confidence and, confidence density was computed as total confidence divided by surface area within a volume.
Here are the key libraries and frameworks used in this project.
Voxblox a volumetric mapping library based mainly on Truncated Signed Distance Fields (TSDFs).
RAFT-Stereo a deep architecture for rectified stereo based on the optical flow network RAFT.
ReAqROVIO is a modified implementation of ROVIO, which enables Visual-Inertial Odometry (VIO) underwater without the need of camera calibration in the water.
GBPlanner2, is a Graph-based Exploration Planner for Subterranean Environments.
This work was conducted as part of the Erasmus Mundus Joint Master Degree in Marine and Maritime Intelligent Robotics (EMJMD MIR). The first author received funding from the European Union under the Erasmus+ Programme.
@article{unpublished,
author = {Bhowmik, Abhimanyu and Singh, Mohit and Sannigrahi, Madhushree and Ludvigsen, Martin and Alexis, Kostas},
title = {Uncertainty Aware Mapping for Vision-Based Underwater Robots},
workshop = {Field Robotics at ICRA},
year = {2025},
}