There’s something endlessly mesmerizing about watching sea life in motion. The way a seal darts through frigid waves or a manta ray glides with such effortless grace can hold anyone’s attention. For centuries, these elegant feats of movement have fueled the curiosity of scientists, eager to understand—and perhaps someday replicate—nature’s mastery underwater.
Human solutions, though, have been far less creative. Most underwater vehicles—those so-called AUVs, or Autonomous Underwater Vehicles—have stuck to a familiar playbook: build them like torpedoes. That long, smooth, tube-like shape slices through water efficiently and is straightforward to design and build. But side-by-side with the creativity of real ocean dwellers, these manmade machines seem a bit, well, uninspired. Part of the problem is simply logistics: trying out radically different ideas is both expensive and time-consuming, so new approaches have usually taken a back seat.
Now, a team from MIT’s CSAIL and the University of Wisconsin at Madison has set out to flip that script. By putting artificial intelligence and 3D modeling to work, they’ve found a new way forward in underwater vehicle design—one that’s more in tune with the creativity of the natural world. Their approach uses AI not just to automate the hard math and simulations, but to “imagine” aquatic craft that traditional engineering might overlook. Even better, the vehicles that emerge from this process can be cheaper and easier to manufacture than their conventionally designed cousins.
Here’s how it works: Led by researcher Peter Yichen Chen, the group started by assembling digital models of all kinds of underwater shapes—some borrowed from manmade designs, others straight from ocean life like whales and manta rays. They put each one inside a flexible digital “deformation cage,” tweaking and stretching the shapes in countless ways, and then ran them through a virtual ocean to see how they’d perform at different angles, searching for new glider designs that might be just right.
But evaluating every possibility by hand just isn’t practical. Instead, the researchers trained a neural network—a kind of digital prediction engine—to learn how each shape’s subtle tweaks changed its hydrodynamic performance. Their main target was the lift-to-drag ratio, a crucial number that describes how easily a vessel can push through water compared to how much it’s dragged backward. In essence: higher ratios mean you travel further, faster, on less power—mirroring what nature has already optimized in its best aquatic travelers.
Putting the AI’s predictions to the test, the team built a prototype glider based on one of their new designs. In wind tunnel experiments, its lift-to-drag numbers matched the AI’s forecasts almost perfectly—off by only about 5%. Two of the most promising glider designs were even 3D printed and fitted with control systems so they could be maneuvered in real water. In pool trials, these new models left the classic torpedo shapes in their wakes, proving the value of this fresh, AI-guided approach.
There’s still plenty of room for refinement: the team is now looking for ways to further narrow the gap between simulation and reality, and to make their gliders even thinner and more agile. One day, underwater robots that constantly adapt their shape in tune with shifting ocean currents may become the norm. The hope is that such technology could spawn a new class of underwater vehicles, tailor-made for missions like tracking climate change, exploring hidden ocean worlds, or safeguarding sensitive marine environments.
Many hands shaped this project: along with Peter Yichen Chen, there’s Pingchuan Ma of OpenAI, Wei Wang from the University of Wisconsin at Madison, and MIT professors Daniela Rus and Wojciech Matusik. The work was supported by DARPA and the MIT-GIST Program. If you’re curious to read more, the original coverage is available over at MIT News.
This website uses cookies.