Robotic intelligence for subsea cable inspection

Everyday life around the world depends on a global network of over 500 subsea cables that silently carry 95% of all international data [1]. These cables stretch more than 1.4 million kilometres across the ocean floor and are critical to the functioning of the internet, financial systems and government communications. But every year, 200 to 300 faults are reported. Most are caused by fishing gear and ship anchors, though abrasion and geological activity also contribute. All can disrupt essential services.

For example, in October 2022, both subsea cables to the UK’s Shetland Islands were damaged, leaving homes and businesses cut off, unable to access the internet or make financial transactions for several days. While various strategies aim to improve the resilience of this infrastructure, robotic systems can play a role through regular inspection and condition monitoring. However, this comes with significant technical challenges.

Challenges for robotic inspection

Near the shore, in water depths up to 2,000 metres, subsea cables are typically buried to protect them from anchors and fishing gear. Although cable routes are planned to avoid hazards, rocky seafloors – where burial is not possible – are sometimes unavoidable. Additionally, underwater currents can shift sediments over time, exposing cables that were previously buried. In areas where such risks are foreseen, alternative protective measures, such as stronger armouring or rock placement, are used. Beyond 2,000 metres, burial becomes too expensive and the risk from human activity is lower. Here, cables are laid directly on the seabed and are thinner – typically under 25mm in diameter.

Physically inspecting exposed cable segments is difficult for several reasons. First, there is uncertainty about the cable’s exact location. During installation, ocean currents can push cables off course by 5–10% of the water depth. Over time, further movement may occur due to seabed currents, landslides or accidental contact. These cables are also thin and hard to spot, and effective inspection requires sub-centimetre resolution. This is further complicated by the high cost of offshore surveys and the lack of GNSS underwater, which makes navigation challenging. Even with state-of-the-art sensor-aided inertial navigation, position errors typically build up at about 1% of the distance travelled. Acoustic tracking systems from the surface also have an error margin of about 1% of the vehicle’s depth.

Putting this into perspective, even in relatively shallow waters of 2,000 metres depth, the positional uncertainty of an undisturbed cable can be up to 200 metres. An underwater vehicle covering 100km in a day might experience position errors of around 20 metres with continuous surface tracking, or as much as one kilometre using only its onboard sensors. And these are best-case figures – uncertainties only increase with depth.

The core challenge for autonomous cable inspection is that the combined uncertainty in both cable routes and vehicle navigation exceed the narrow field of view of the high-resolution sensors needed for cable detection and inspection. This makes the traditional pre-programmed waypoint-following approaches commonly used by autonomous underwater vehicles (AUVs) ineffective. Additionally, the large volumes of high-resolution data collected during inspection make manual analysis labour-intensive and prone to error.

Addressing challenges with intelligent autonomy

To tackle these challenges, engineers at the University of Southampton developed real-time methods that allow AUVs to autonomously detect and track cables and report key findings remotely – without the need for any human intervention. Field trials using the camera-equipped Smarty200 AUV demonstrated the system’s effectiveness (Figure 2). The system integrates three core capabilities:

Cable detection: Smarty200 uses machine learning to detect cables in real time from camera images [2]. The AUV flies between 1.5 to 3m above the seafloor, using acoustic sensors and its vertical thruster to maintain altitude. The system captures millimetre-resolution, strobe-illuminated images and laser-scanned micro-bathymetry to map the cable and surrounding terrain in 3D (Figure 3). The system can detect cables across a variety of seabed types, automatically generating waypoints relative to the cable path to maintain tracking.

Map-guided navigation: The AUV starts with a rough map of the cable path, using it to automatically generate zig-zag search patterns that cover the estimated positional uncertainty of the map. When it detects the cable, it updates an internal probabilistic map that reflects the most likely cable path based on its observations. If the AUV loses its cable tracking – due to burial or detection and trajectory errors – it uses the updated map to resume searching around the most likely area. Search boundaries are generated using physics-based ‘catenary models’, which estimate the plausible region where the cable might be if disturbed, while maintaining consistency with the accumulated observations.

Rapid reporting: High-resolution AUV camera surveys can generate hundreds of gigabytes of data, traditionally requiring physical recovery of the vehicle and downloading of data via a tether. Smarty200 avoids this by using self-supervised machine learning to generate compact observation summaries that can be remotely transmitted. Operators can provide the AUV with feature representations extracted from reference images (called queries) before, during or after a mission, and the AUV computes compact summaries of the images it captures based on their similarity to these [3]. Geotagged image similarity metrics and a subset of the most similar images are compressed and sent over low-bandwidth satellite or acoustic links. In recent trials, a 146GB cable dataset was reduced to just 188kB – a ratio of 1:800,000 – allowing transmission in about 20 minutes via Iridium satellite.

These methods work together to enable robust, fully autonomous cable inspection. The system does not require perfect detections. Instead, it updates a probabilistic cable path map to centre search patterns around the most likely path if the cable is buried or detection fails. Physics-based models keep the AUV within a likely corridor and summarized image reporting provides rapid and flexible over-horizon awareness of relevant observations before the vehicle is recovered.

Field trials

The system was tested in shallow-water field trials using a section of deep-sea cable provided by the British Telecom Group. The test site featured a range of seafloor types to evaluate the system’s robustness, and multiple tracking experiments were performed. Figure 4 shows results from one of these, where the AUV began its mission by searching around the initial cable route map (which was offset from the actual cable) in a zig-zag pattern and made initial detections. Although the AUV temporarily lost track of the cable at several points, by updating its internal cable map to be consistent with its observations, it was able to successfully reacquire the cable each time. When detection errors led the AUV away from the cable, it realized it was no longer detecting the cable and returned to the most likely path according to its updated map. The observation summary (bottom right of Figure 4) was generated in response to a query focused on the repeater (a device used to boost signal strength within the cables). The AUV assessed the similarity of each image to the query, where the outlines indicate the most similar images that it selected and transmitted at the surface. Figure 5 shows some examples of original and compressed versions of images transmitted during the trials. Key features are preserved despite the compression, supporting operator awareness for decision-making.

Looking ahead

The technologies demonstrate a path towards full autonomous inspection of exposed subsea communication cables. Further work is needed to explore compatibility with other high-resolution sensing modalities, including the integration of multiple sensors, and verify robustness over longer and more complex cable routes. While the Smarty200 is a capable demonstrator, real-world inspections will require deeper-diving AUVs with greater range and endurance. Recent multi-week, 1,000km shore-launched seafloor camera surveys – using the National Oceanography Centre’s Autosub Long-Range AUV and the BioCam imaging system – demonstrate that such extended operations are already technically feasible [4].

References

[1] Clare, M., Yeo, I., Bricheno, L., Aksenov, Y., Brown, J., Haigh, I., Wahl, T., Hunt, J., Sams, C., Chaytor, J., Bett, B., & Carter, L.  (2022). Climate change hotspots and implications for the global subsea telecommunications network. Earth-Science Reviews, 237, 104296. https://doi.org/10.1016/j.earscirev.2022.104296

[2] Yamada, T., Prügel-Bennett, A., Williams, S., Pizarro, O., & Thornton, B. (2022). GEOCLR: Georeference Contrastive Learning for efficient Seafloor Image Interpretation. Field Robotics, 2(1), 1134–1155. https://doi.org/10.55417/fr.2022037

[3] Yamada, T., Prügel‐Bennett, A., & Thornton, B. (2020). Learning features from georeferenced seafloor imagery with location guided autoencoders. Journal of Field Robotics, 38(1), 52–67. https://doi.org/10.1002/rob.21961

[4] Bodenmann, A., Jones, D. O. B., Phillips, A. B., Templeton, R., Sherif, R., Fanelli, F., Newborough, D., & Thornton, B. (2025). Remote awareness of image quality for multi-week shore-launched AUV surveys. IEEE Transactions on Field Robotics., 2, 147–164. https://doi.org/10.1109/tfr.2025.3529435