Onboard ATR Is Now Available With Siren

Big news from the Bedrock fleet: automated target recognition (ATR) is officially running on board every one of our autonomous underwater vehicles. Our AUVs now spot and classify seafloor contacts as they survey, then send them ashore while they're still in the water.

This release has been a top priority this year, and we couldn't be prouder of how it turned out. This is a far cry from off-the-shelf ATR. Here's a look at what it does and how it works.

From Days To Minutes

If you've worked in marine surveys, you know ATR is one of the most desired capabilities out there. It finds the things that matter on the seafloor: debris, infrastructure, and potential unexploded ordnance.

Traditionally, picking those contacts has been slow, manual work. The data comes back after the survey. Then it gets transferred, processed, and run through separate software. Only then do analysts start reviewing. That loop can take days, sometimes months.

We’ve cut it down to minutes.

How It Works

Here's the whole journey, from survey line to shore:

  1. An AUV finishes a survey line.

  2. The ATR model runs right there on board, usually in under a minute.

  3. The vehicle surfaces and sends geolocated contacts back to shore, each paired with its sidescan sonar image.

  4. The team on shore starts reviewing.

From start to finish, that typically takes less than five minutes. Under the hood is a proprietary deep learning model. It’s architecturally similar to leading real-time object detection frameworks, but it's pre-trained on Bedrock's own sidescan sonar data, plus some industry-trusted open-source datasets. That bespoke training makes this the frontier ATR for the industry. 

Real-Time People Power

Here's the part we love most. The model doesn't get the final word. Onshore experts do.

"A list of coordinates from an algorithm isn't a true contact," says Dr. Bridgit Boulahanis, our Head of Data and Solutions. In mine countermeasures, for example, someone is deciding whether an object could be explosive. That call needs a trained human.

Analysts can view each detection with and without the model's interpretation. Then they accept it, reject it, or annotate it based on what the client actually needs to know. Moreover, every one of those expert decisions also feeds continuous retraining. The model keeps learning from the best people in the business.

The Bedrock Difference

ATR isn't new. What's different is how we built it.

Bedrock is vertically integrated: our own autonomous vehicles, adaptable sensing, and a data pipeline that delivers insights orders of magnitude faster than traditional surveys. Many operators swap sensors between missions, and every new sensor brings its own artifacts to wrestle with. We train on consistent sensor data. That makes our ATR reliably more accurate and more robust.

As our Head of Software, Shelley Bower, puts it, customers aren't buying data. "They're buying decisions." Onboard ATR helps them make the right decisions, faster.

What’s Next

We're not slowing down. The next version of the model is already in the works and expected in the coming months. It will combine sidescan sonar with magnetometer data, giving analysts the most complete picture yet of what's on the seafloor.

In the meantime, onboard ATR is live across our AUV fleet right now. Want to see what it can do for your next survey? Get in touch with us!

Next
Next

Bedrock Monitors the Ocean’s Seabed Infrastructure