Learning Underwater Active Perception in Simulation

1. Problem Statement

Underwater autonomous inspection missions face significant challenges due to variable water conditions (turbidity, backscattering, absorption) that degrade image quality.

• The Core Issue: Poor visibility leads to low-contrast images, making visual documentation unreliable and hindering downstream tasks like mapping or object recognition.
• Limitations of Current Methods:
  • Traditional active perception often focuses solely on minimizing distance to targets, which reduces visual coverage and increases mission time.
  • Existing quality-aware methods often require complex hardware setups (independently movable cameras and lights) or rely on prior knowledge of the environment.
  • Standard rendering engines (e.g., Blender, Unreal) use simplified physics (Henyey-Greenstein phase functions) that fail to accurately simulate underwater light scattering, making synthetic training data unrealistic.

2. Methodology

The authors propose a learning-based active perception framework that operates in three integrated stages: Simulation, Calibration, and Runtime Optimization.

A. Enhanced Underwater Simulation (Blender Upgrade)

To generate realistic training data, the authors modified the Blender rendering engine:

• Physics Replacement: Replaced the standard Henyey-Greenstein (HG) scattering phase function with the Fournier-Forand (FF) function. The FF function more accurately models light scattering at small and large angles, specifically capturing backscattering and haze.
• Ocean Modeling: Integrated Jerlov water types (10 distinct types ranging from coastal to oceanic) to simulate varying absorption and scattering coefficients.
• Noise Modeling: Added thermal and signal-dependent Poisson noise to mimic real sensor characteristics.
• Dataset Generation: Created a synthetic dataset of 138,240 entries covering various water types, depths, distances, and light intensities.

B. In-Situ Water Column Calibration

Before the main inspection, the vehicle performs a calibration routine to estimate local water properties:

1. Depth Profile: The vehicle observes the water surface at varying depths (lights off) to measure light attenuation ($K_z$).
2. Distance Profile: The vehicle moves toward a target at varying distances with lights on and off to measure contrast changes ($K_c$) and backscattering properties.

• Technique: Instead of using whole-image statistics (which can be biased by ambient light), the method uses patch-based standard deviation to calculate contrast more accurately.

C. Runtime Active Perception Framework

The core runtime system uses a Multi-Layer Perceptron (MLP) to guide the vehicle:

• Inputs (20 total):
  • 9 Calibration parameters (channel-wise attenuation and contrast profiles).
  • 9 Current state parameters (depth, current distance, current light intensity, image mean, and image standard deviation).
  • 2 Optimization variables (proposed distance offset $\Delta d$ and light intensity offset $\Delta l$).
• Outputs: Predicted channel-wise image contrast.
• Optimization Loop:
  • The system uses an Nelder-Mead optimization algorithm to find the optimal $(\Delta d, \Delta l)$ pair.
  • Objective Function: A weighted sum balancing Predicted Contrast ($m_c$) and Visual Coverage ($m_d$).
  • Note: A distance regulator is added to prevent the model from simply driving the robot infinitely close to the target (which maximizes contrast but minimizes coverage).

3. Key Contributions

1. Blender Physics Upgrade: Successfully integrated the Fournier-Forand phase function into Blender (now part of the official 4.3 release), enabling physically accurate underwater light propagation simulation.
2. In-Situ Calibration Routine: A novel method to estimate water column properties (attenuation and backscatter) using only a monocular camera and adjustable artificial lighting, eliminating the need for expensive external sensors.
3. Active Perception Framework: A learning-based system that provides real-time guidance on distance and light intensity to maximize image quality while maintaining efficient visual coverage across diverse water conditions.

4. Experimental Results

The approach was validated in simulation against traditional static patterns (fixed distance/light) and other active perception methods (Gradient-based, NIQE-based).

• Model Accuracy:
  • The MLP trained with calibration parameters achieved a Mean Absolute Error (MAE) of 0.01 on the test set.
  • Without calibration, the MAE rose to 0.05, demonstrating the critical importance of in-situ water property estimation.
• Feature Extraction & Mapping:
  • In low turbidity, the proposed method achieved a 42% inlier ratio for feature matching, outperforming the Gradient (30%) and NIQE (30%) methods.
  • In high turbidity, it maintained a 22% inlier ratio, significantly higher than the Gradient (17%) and 2m pattern (8%) methods.
• Visual Coverage:
  • The proposed method achieved a balance between quality and area. In low turbidity, it covered 35.9 $m^2$, comparable to the 2m pattern (39.4 $m^2$) but with significantly higher image quality.
  • In high turbidity, it covered 16.5 $m^2$, whereas the 2m pattern failed to extract usable features entirely (19.7 $m^2$ but with poor quality).
• Stability: Unlike other methods that fluctuate wildly based on scene content, the proposed approach showed stable parameter adjustments over depth and time.

5. Significance and Impact

• Operational Efficiency: The framework allows AUVs to autonomously adapt to local water conditions, reducing the risk of mission failure due to poor data quality and minimizing the need for human intervention.
• Cost Reduction: By enabling high-quality data collection in turbid waters (previously considered prohibitive), it reduces the need for repeated missions.
• Scalability: The use of synthetic data and a lightweight MLP allows for real-time deployment on underwater vehicles without heavy computational overhead.
• Open Source: The authors have released the modified Blender tools and project resources, fostering further research in underwater robotics simulation.

Limitations: The current model assumes relatively planar assets (e.g., ship hulls) and may struggle with complex 3D structures where a single distance metric is insufficient. It also requires the vehicle to have high maneuverability for the calibration phase. Future work aims to address deep-water extrapolation and complex asset geometries.