💻 computer science

SAGE: Ergodic Control for Autonomous and Adaptive Inspection of Subsea Infrastructure

The paper presents SAGE, an adaptive ergodic control framework that dynamically allocates autonomous vehicle inspection time to subsea valves based on real-time risk assessments, significantly outperforming fixed-route A* planning by enabling more frequent monitoring of high-risk components and immediate behavioral adjustments to newly detected leaks without requiring explicit re-planning.

Original authors: Markus Buchholz, Ignacio Carlucho, Yvan R. Petillot

Published 2026-08-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Markus Buchholz, Ignacio Carlucho, Yvan R. Petillot

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

Deep beneath the ocean surface, critical infrastructure known as subsea Christmas trees manages the flow of oil and gas. These complex structures are studded with valves, flanges, and connectors that require constant monitoring to ensure they are not leaking. For decades, this inspection has been a human-led endeavor, relying on divers or remotely operated vehicles that follow strict, pre-written schedules. The challenge lies in the fact that not every part of the structure is equally important at every moment. A valve with a suspected leak demands immediate and frequent attention, while a component with a clean history can wait longer. In the traditional approach, if a new leak is discovered mid-mission, the human operator must stop the vehicle, manually recalculate the route, and issue new commands to prioritize the danger. This process is slow, interrupting the flow of work and leaving critical gaps in coverage while the plan is rewritten.

Researchers have developed a new approach called SAGE, which stands for Semantic and Adaptive Generative Ergodicity, designed to let underwater robots handle these shifting priorities on their own. Instead of following a fixed loop that visits every point in the same order, this system treats the inspection area as a map of risk. The robot continuously calculates where the danger is highest and adjusts its movement to spend more time in those specific spots. The core idea is that the robot's path should naturally reflect the changing needs of the mission. If a valve suddenly becomes high-risk, the robot does not wait for a human to redraw the map; it immediately begins to circle that valve more often, simply because its internal calculation of "where to go next" has changed. This method, known as ergodic control, aligns the robot's movement with a live distribution of risk, ensuring that the time spent near a component is proportional to how much attention that component needs right now.

In a simulated environment, the researchers tested this system against a standard, fixed-route method using a scenario with two underwater structures containing five valves in total. They created a situation where one valve developed a significant leak, representing a high-risk scenario, while the others remained routine. The traditional method, which used a pre-planned tour, treated every valve with the same schedule. It visited each valve exactly every 8.1 seconds, regardless of whether a valve was leaking or perfectly safe. Even when the risk at the leaking valve spiked, the robot continued its fixed loop, unable to prioritize the emergency without a human interrupting the mission to re-plan the route. In contrast, the SAGE system reacted instantly to the simulated leak. Without any human intervention or a pause to recalculate a path, the robot automatically shifted its behavior. It began visiting the leaking valve every 5.8 seconds, checking it roughly 1.4 times more often than the fixed-route robot. Over the course of a five-minute simulation, the SAGE robot visited the leaking valve 50 times, while the traditional robot visited it only 37 times.

The difference in performance extended beyond just the leaking valve. Because the SAGE system was focused on the high-risk area, it naturally spent less time at the valves that posed no threat. In the simulation, the robot completely ignored one of the low-risk valves, visiting it zero times, while the fixed-route robot wasted time checking it 37 times. This ability to dynamically allocate time means that a potential leak can go unnoticed for a much shorter period under the new system. The researchers found that the time between checks for the dangerous valve was reduced significantly, cutting the window of vulnerability to about two-thirds of what it was with the old method. The system achieved this by constantly updating a mental map of risk and adjusting its speed and direction in real-time, rather than following a rigid script.

While the results in the simulation were promising, the researchers acknowledge that the system has limitations. In the test, the robot's intense focus on the leaking valve meant it completely neglected the lowest-risk valve, which could be a problem if that valve suddenly developed an issue. The team suggests that future versions of the software could include a safety rule to guarantee that every valve is checked at least once, even if it is low-risk, ensuring no part of the structure is entirely forgotten. The next steps for the technology involve closing the loop between the robot's eyes and its brain. Currently, the risk levels in the simulation were set by hand, but the ultimate goal is for the robot to detect leaks directly through its cameras or sonar and update the risk map automatically. The researchers also plan to test the system on real underwater vehicles in the ocean, including experiments where a leak is introduced mid-mission to see how the robot adapts in the real world.

This technology represents a shift from rigid scheduling to adaptive awareness, allowing autonomous systems to make decisions based on real-time conditions rather than a pre-written plan. The same logic could eventually be applied to other tasks where priorities change on the fly, such as searching for survivors in disaster zones, monitoring environmental hazards like oil spills, or collecting litter in coastal waters. By letting the robot decide where to go based on where the need is greatest, the system offers a way to make underwater inspection faster, safer, and more responsive to the immediate realities of the deep ocean.

Technical Summary: SAGE – Ergodic Control for Autonomous and Adaptive Inspection of Subsea Infrastructure

Problem Statement
The inspection of subsea production infrastructure, specifically Subsea Christmas Trees (XTs), involves repeatedly monitoring a known set of components (valves, flanges, connectors) rather than searching an unknown environment. Current practices rely on divers or Remotely Operated Vehicles (ROVs) following fixed schedules or scripted routes under human supervision. This approach fails to address the non-stationary nature of risk: a valve with a suspected leak requires significantly more frequent attention than one with a clean history. Traditional methods treat all points equally or require manual re-planning (e.g., adjusting wait times and re-computing routes) whenever a new risk emerges, creating a latency between detection and response that leaves leaks unnoticed for longer periods.

Methodology: SAGE and Ergodic Control
The paper presents SAGE (Semantic and Adaptive Generative Ergodicity), an architecture that replaces scripted route planning with an ergodic control law. Instead of generating a discrete path, SAGE acts as an autonomous coverage controller that aligns the time-averaged statistics of the vehicle's trajectory with a dynamic, sensor-derived risk distribution.

Risk Distribution Modeling: The workspace is modeled as a probability distribution $\phi(x, t)$ , constructed as a weighted mixture of Gaussian densities centered at the known positions of the valves. The weight $w_i(t)$ for each valve $i$ represents its time-varying risk, updated autonomously based on operator intent or sensor data (e.g., camera/sonar leak detection).
Control Law: The controller does not plan a route but computes a velocity vector every control cycle to minimize the mismatch between the trajectory's running time-average and the target risk distribution. This is achieved using the spectral multiscale coverage law of Mathew and Mezić. Both the target distribution and the trajectory are projected onto a Fourier basis. The control signal $S(x, t)$ drives the vehicle to reduce the weighted difference between the spectral coefficients of the target distribution ( $\phi_k$ ) and the trajectory ( $c_k(t)$ ).
Adaptivity: Because the control law operates on the current risk distribution $\phi$ rather than a pre-computed plan, any change in a valve's risk weight immediately alters the commanded velocity in the next control cycle. This eliminates the need for discrete re-planning steps or human intervention to redirect the vehicle.

Key Contributions and Experimental Setup
The authors validate SAGE in a simulation environment (Gazebo) using a BlueROV2 inspecting two subsea XTs containing five non-collinear valves. The study compares SAGE against a baseline fixed-loop A* tour, where wait times are tuned to match the total time SAGE spends at high-risk valves, but the revisit frequency remains fixed by the loop geometry.

Scenario: A "one-extreme" profile was tested where Valve P1 carries a high risk weight (0.55) simulating a confirmed leak, while other valves have lower weights.
Metrics: The primary metrics are revisit frequency (how often a valve is visited) and the interval between visits.

Results
Over a 300-second mission with the one-extreme risk profile:

A Baseline:* Visited every valve exactly 37 times with a fixed interval of 8.1 seconds, regardless of risk. It failed to prioritize the leaking valve.
SAGE (Ergodic Control): Visited the high-risk Valve P1 50 times (every 5.8 seconds), while visiting lower-risk valves fewer times (e.g., Valve P3 received 0 visits).
Efficiency: Under SAGE, a leak at P1 is revisited roughly 1.4 times more frequently than under the A* baseline, reducing the time a leak goes unnoticed by approximately one-third.
Correlation: The revisit count for SAGE showed a strong positive correlation (+0.93) with the risk weights, whereas the A* baseline showed zero variance.

The paper notes a limitation: unconstrained ergodic control optimizes long-run averages and may starve low-weight valves (e.g., P3 received zero visits) within finite mission times. The authors suggest this could be mitigated by adding a minimum weight floor or hard revisit constraints in future iterations.

Significance and Claims
The paper claims that SAGE offers a fundamental shift from "planning a route" to "controlling coverage." Its primary significance lies in its ability to:

Prioritize Dynamically: Automatically adjust revisit frequency based on live risk without human re-planning.
Continuous Adaptation: React to new leak detections immediately within a single control cycle, whereas fixed tours require discrete, operator-triggered re-routing.
Generalizability: The architecture requires only a risk weight per point of interest, making it applicable to other coverage tasks defined by risk fields, such as search-and-rescue, environmental monitoring (e.g., algal blooms, oil spills), or litter collection.

The authors position SAGE not as a perfect solution for all finite-time constraints (acknowledging the zero-visit issue for low-risk items) but as a robust, autonomous framework for adaptive inspection that removes the latency inherent in traditional, schedule-based subsea operations.

Technical Summary: SAGE – Ergodic Control for Autonomous and Adaptive Inspection of Subsea Infrastructure

More like this