Learning Underwater Active Perception in Simulation

TL;DR

The paper addresses the challenge of variable underwater visibility by proposing an active perception framework that leverages synthetic data from a physics-informed Blender pipeline, a calibration routine to estimate water-column properties, and a runtime MLP-guided strategy to optimize distance and illumination for high-quality imagery and coverage. The method combines an enhanced underwater image formation model with a water-column calibration and a real-time optimization loop to maximize image contrast while maintaining survey efficiency. Key contributions include (i) physically improved underwater imagery simulation via the Fournier-Forand phase function and diversified water types, (ii) an in-situ monocular water-column estimation method, and (iii) a runtime guidance framework that balances image quality and visual coverage using Nelder-Mead optimization. The findings, demonstrated in simulation, show significant gains in imagery quality and coverage over traditional approaches, enabling more robust autonomous underwater inspections across a broad range of conditions.

Abstract

When employing underwater vehicles for the autonomous inspection of assets, it is crucial to consider and assess the water conditions. Indeed, they have a significant impact on the visibility, which also affects robotic operations. Turbidity can jeopardise the whole mission as it may prevent correct visual documentation of the inspected structures. Previous works have introduced methods to adapt to turbidity and backscattering, however, they also include manoeuvring and setup constraints. We propose a simple yet efficient approach to enable high-quality image acquisition of assets in a broad range of water conditions. This active perception framework includes a multi-layer perceptron (MLP) trained to predict image quality given a distance to a target and artificial light intensity. We generated a large synthetic dataset including ten water types with different levels of turbidity and backscattering. For this, we modified the modelling software Blender to better account for the underwater light propagation properties. We validated the approach in simulation and showed significant improvements in visual coverage and quality of imagery compared to traditional approaches. The project code is available on our project page at https://roboticimaging.org/Projects/ActiveUW/.

Figures (10)

• Figure 1: Overview of the proposed approach, divided into three components. The generation of synthetic data (left) is done in Blender and includes realistic physics and a large variety of water types. The vehicle needs to follow the calibration routine (center) to understand the water column and improve the runtime model (right) to better guide the vehicle and obtain better inspection data.
• Figure 2: Renders with similar absorption characteristics but the scattering coefficient is gradually increasing from top to bottom.
• Figure 3: The runtime model provides guidance suggestions based on the MLP which predicts image contrast given a change in illumination and distance.
• Figure 4: Optimisation over time of the distance and illumination. The vehicle adapts differently according to depth and turbidity.
• Figure 5: The number of feature matching inliers and inlier ratios are compared in two environments, low (\ref{['fig:inliers:1']}) and high (\ref{['fig:inliers:2']}) turbidity. The proposed method shows more consistency and higher numbers on average.