Terradepth takes edge computing under the sea

There are a lot of rough computing environments out there. Mines in the Australian outback. Oil rigs in the North Sea. Even factory floors have environmental challenges (there’s a whole class of ratings for explosive-proof electronics, for example). But Terradepth, an Austin, Texas-based startup, might be trying to tackle computing in the toughest environment of all — almost 10,000 feet under the sea.

Terradepth is a three-year-old startup founded by two former Navy SEALs; it’s building an autonomous submersible that could map the ocean floor. The company’s goal is to have thousand of submersibles jam-packed with sensors trawling for data in the oceans. Think of a constellation of satellites, only instead of beaming pictures down to earth, these submersibles are sending sensors and image data up from the depths.

From a computing perspective, Terradepth had to address several challenges. Topping that list is how can a device gather what is roughly a terabyte of data each day, analyze it, and share the relevant information without blowing through a constrained power budget?

The answer is as follows: Each submersible has a multitude of sensors, plus sonar, which is used to create the images of the ocean floor. The submersible has computers powered by Nvidia graphics chips and Intel’s i7 CPUs. Together they take in that data and perform image recognition tasks on it so as to flag anomalies that might mean the submersible should surface to send that same data to its handlers for further instruction.

Even when the computers don’t “see” an anomaly, each device has to surface every three days to pull in air in order to keep powering the engine that propels the submersible forward. Engines work because they pull in air to power a combustion cycle. So underwater, a submarine needs electric power from batteries or a nuclear core.

To keep costs down, the Terradepth submersibles use batteries and a diesel engine. The idea is to deploy the submersibles in attached pairs, so as power runs low on the deeper sub it switches positions with the one on the surface. And once the device reaches the surface, it sends the relevant data via a satellite connection.

Joe Wolfel, co-CEO of Terradepth, wants to get the cost of each submersible down to about $1 million as opposed to the anywhere from $5 million to $10 million that an underwater exploration vehicle can cost today. And by taking the human out of the cockpit and using off-the-shelf computing parts, he thinks the company can do it.

“We don’t envision removing the human, but we can reduce the human cognitive load,” says Wolfel. That way, a human doesn’t have to look at hours of camera footage or reformulated maps created by sonar imaging to figure out what’s on the bottom of the ocean floor. Instead, the computer looks for anomalies such as ships or topographical features and alerts humans who can elect to take a closer look (or not).

The engineering of the submersible, its computers, and its power system was a challenge, but an even bigger task was building the machine learning models. Unlike image recognition or wake word detection, where there is a lot of available data to train a model with, the sea floor is a mystery. The team had to build an entire set of training data to help the computer translate sonar to visual images.

Today the team is testing its first submersible in the lakes near Austin and plans to continue testing in April, in the Gulf of Mexico. With climate change driving the need for more accurate information about sea temperatures, salinity, and water flow, Wolfel expects governments to invest in technology to better understand the ocean. He’s hoping Terradepth is the answer.