💻 computer science

Low-Cost Underwater In-Pipe Centering and Inspection Using a Minimal-Sensing Robot

This paper presents a low-cost, autonomous underwater robot system that achieves reliable centering and traversal of flooded pipes using only an IMU, pressure sensor, and two sonars, eliminating the need for complex external localization or high-end sensors.

Original authors: Kalvik Jakkala, Jason O'Kane

Published 2026-02-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kalvik Jakkala, Jason O'Kane

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to swim down a long, dark, underwater tunnel that is completely filled with murky water. You can't see anything, there's no GPS signal, and the tunnel might be wobbly, bent, or even slightly crushed in places. Now, imagine you are a robot tasked with swimming straight down the middle of this tunnel to check for cracks or leaks, without bumping into the walls.

This is the challenge the authors of this paper tackled. They built a system that lets a robot do exactly that, but with a twist: they used the cheapest, simplest sensors possible to solve a problem that usually requires expensive, complex equipment.

Here is how they did it, broken down into everyday concepts:

1. The Robot's "Eyes" (The Sensors)

Most underwater robots use fancy cameras (useless in muddy water) or expensive laser scanners. This team used a "minimalist" toolkit:

A Pressure Sensor: Like a diver's watch, it tells the robot how deep it is.
An IMU: Like the inner ear in your head, it knows which way is up and how the robot is tilting.
Two Sonars (The "Echolocators"):
- Sonar A (The Downward Looker): A simple device pointing straight down that says, "I am X meters from the bottom."
- Sonar B (The Rotating Flashlight): A device that spins around 360 degrees, sending out sound pings and listening for echoes, like a bat or a lighthouse scanning the horizon.

2. The "Magic" of the Sonar Data

The tricky part was the rotating sonar. It doesn't just say "Wall is 2 meters away." Instead, it gives a messy graph of sound intensity, full of static and confusing echoes (like shouting in a canyon and hearing your voice bounce off three different rocks).

The authors wrote a clever, fast computer program to clean up this mess. Think of it like a noise-canceling headphone for sonar. It filters out the static, ignores the confusing echoes, and finds the very first clear "ping" that hits the wall. This allows the robot to know exactly how far away the wall is, even in noisy conditions.

3. The "Geometry Trick" (Finding the Center)

Once the robot knows how far it is from the bottom (Sonar A) and how far it is from a specific spot on the side (Sonar B), it uses a simple geometry trick to find the center of the pipe.

Imagine you are inside a circular room. If you know the distance to the floor and the distance to a spot on the wall, you can draw two lines. Where those lines intersect (mathematically speaking) helps you figure out where the center of the room is.

The Problem: Sometimes, the math gives you two possible centers (like a mirror image).
The Solution: The robot looks at where it was a second ago. It picks the center that makes the most sense based on its previous position, ignoring the impossible one.

4. The "Confidence" Steering Wheel

The robot doesn't just blindly steer. It has a "confidence meter."

High Confidence: If the two sonar beams are pointing in very different directions (like one down and one to the side), the robot is 100% sure where the center is. It steers firmly.
Low Confidence: If the beams are pointing in almost the same direction (like both pointing forward), the math gets shaky. The robot realizes, "I'm not sure where the center is right now," so it steers very gently and slowly, waiting for better data.

This prevents the robot from jerking around wildly if the sensors get confused by a weird bend in the pipe.

5. The Results: Swimming Through the Chaos

The team tested this on a real robot (a BlueROV2) inside a 46cm-wide flexible pipe submerged in a swimming pool.

The Setup: The pipe wasn't perfect. It had bends, it was wobbly, and the water was moving.
The Test: They pushed the robot off-center and let go.
The Outcome: The robot successfully found the middle of the pipe and swam straight through the entire length, even navigating through a section where the pipe was squished inward. It did all this without crashing into the walls.

Why This Matters

Usually, to navigate a pipe like this, you need a robot with a massive array of sensors, a powerful computer, and maybe even a tether to a human operator. This paper proves you can do it with a lightweight, low-cost setup that runs on simple math. It's like showing that you can navigate a dark cave using just a flashlight and a compass, rather than needing a full team of surveyors and a laser mapping system.

In short: They taught a robot to "feel" its way through a dark, muddy underwater tunnel using simple sound waves and smart math, keeping it safely in the middle without needing expensive gear.

Technical Summary: Low-Cost Underwater In-Pipe Centering and Inspection Using a Minimal-Sensing Robot

Problem Statement
Autonomous inspection of submerged pipelines (e.g., levees, culverts, oil pipelines) is hindered by confined geometries, high turbidity, and the absence of GPS. Existing solutions often rely on complex mechanical crawlers, dense sensor arrays (including Doppler Velocity Logs and multi-sensor sonar), or vision-based systems that fail in murky water. These approaches increase cost, operational burden, and computational requirements. There is a need for a lightweight, computationally efficient system capable of maintaining a free-swimming robot centered within a pipe using minimal sensing hardware.

Methodology
The authors propose a framework utilizing a Blue Robotics BlueROV2 equipped with a minimal sensor suite: an Inertial Measurement Unit (IMU), a pressure sensor, a downward-facing single-beam sonar (Ping1D), and a mechanically rotating 360° single-beam sonar (Ping360). The system operates under the assumption that the pipe radius is known a priori.

Sonar Range Estimation:
The rotating 360° sonar provides raw acoustic intensity profiles rather than explicit range data. The authors developed a computationally efficient signal processing pipeline to extract wall ranges from these profiles:
- Near-field suppression: Discarding initial samples corrupted by transducer ring-down.
- Denoising and Edge Enhancement: Applying Gaussian smoothing followed by convolution with a linearly tapering kernel to highlight rising edges of wall reflections.
- Peak Detection: Identifying the first valid peak exceeding a threshold to determine the wall echo.
- Temporal Smoothing: Using a scalar Kalman filter to ensure consistency.
  This method achieves a Root Mean Square Error (RMSE) of 0.02 m against ground truth, effectively handling noise and multipath interference without deep learning.
Geometric Center Estimation:
The system estimates the pipe center using a closed-form geometric model based on two points: the downward sonar return and the 360° sonar return at a specific azimuth.
- Given the known pipe radius $r$ and two wall points, the algorithm calculates two candidate circle centers.
- A selection logic chooses the physically consistent center based on an "inside-outside" test (ensuring the robot is inside the circle) and proximity to the previous estimate.
- A discrete Kalman filter smooths the center trajectory.
Adaptive Confidence-Weighted Control:
The accuracy of the center estimate depends on the angular separation between the two sonar beams. The system models measurement variance as a function of this angle (high variance when beams are colinear, low when orthogonal).
- Uncertainty-Aware Scaling: Confidence weights ( $w_x, w_z$ ) are derived from the estimated variances. These weights scale the output of Proportional-Derivative (PD) controllers.
- Control Strategy:
  - Lateral/Vertical: PD controllers minimize displacement from the estimated center, scaled by confidence weights to prevent overreaction to unreliable data.
  - Forward Motion: Enabled only when the robot is within a tolerance of the centerline.
  - Attitude: PD controllers stabilize roll and pitch; yaw is held using gyroscopic feedback alone, avoiding unreliable magnetometer readings in metallic environments.

Key Contributions

Efficient Range Extraction: A signal processing algorithm that extracts reliable wall ranges from raw single-beam sonar intensity profiles with minimal computational overhead, avoiding the cost of deep learning approaches.
Minimal-Sensing Geometric Formulation: A closed-form method to estimate the pipe center using only two sonar points and a known radius, eliminating the need for complex sensor arrays or external tracking.
Adaptive Control Framework: A confidence-weighted PD control system that dynamically adjusts actuation based on the geometric reliability of the sonar measurements.
Experimental Validation: Successful demonstration of stable centering and traversal in a submerged, flexible, and deforming pipe using a low-cost robotic platform.

Results
The system was validated through three stages:

Sonar Processing: On real-world data from a 0.46 m diameter pipe, the range estimation method achieved an RMSE of 0.02 m.
Simulation: In a 2D geometric simulation with added noise, the robot converged to within 0.05 m of the true center in an average of 10 update steps, maintaining a steady-state error of ~0.029 m.
Field Experiments: The BlueROV2 successfully autonomously re-centered itself after manual displacement and traversed a submerged, flexible fabric pipe. The system maintained stable alignment despite ambient water flow, pipe curvature, and a localized inward deformation (constriction) in the pipe wall.

Significance and Claims
The paper claims that reliable in-pipe navigation and inspection can be achieved with a lightweight, computationally efficient architecture that does not require Doppler Velocity Logs, external tracking, or complex multi-sensor arrays. By leveraging minimal sensing (IMU, pressure, two sonars) and efficient geometric reasoning, the proposed approach advances the practicality of autonomous underwater inspection in confined, GPS-denied environments. The authors note that while the system performs robustly, it assumes an initial approximate alignment with the pipe axis and a known circular cross-section, and it requires careful tuning of filter gains and thresholds. Future work is directed toward extending the method to complex pipe networks and improving robustness under stronger flow conditions.