At the bottom of a swimming pool on MIT’s campus, researchers keep testing a strange glider. The glider doesn’t have a propeller, nor does it resemble any known fish. It’s like a mix between a paper airplane and a fever dream, made out of plastic. It is, of course, an artificial intelligence creation.
This curious contraption, along with its even stranger cousin (a flat, four-winged glider), are among the first autonomous underwater vehicles designed almost entirely by a machine-learning system. Their creators say these new shapes could soon revolutionize how scientists explore the ocean, from mapping currents to monitoring the effects of climate change.
“This level of shape diversity hasn’t been explored previously, so most of these designs haven’t been tested in the real world,” said Peter Yichen Chen, a postdoctoral researcher at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and co-lead author of the project.
Nature-Inspired, Machine-Optimized
For decades, marine scientists have tried mimicking nature’s hydrodynamic brilliance. Seals, whales, and rays slice through water with uncanny efficiency. Autonomous underwater vehicles (AUVs), by contrast, have remained largely utilitarian: streamlined tubes with wings, efficient but uninspired.
The reason is simple; it’s hard to reinvent the hull. Designing, building, and testing new shapes underwater is time-consuming and expensive. So researchers stick with what works.
The new AI pipeline, developed by researchers at MIT and the University of Wisconsin-Madison, upends that process. Instead of relying on human intuition and knowledge, the team built a system that co-designs both the shape of the glider and the way it controls itself as it moves.
The algorithm learns from each shape it tests. The core of the system is a neural network that predicts how proposed gliders would behave under different conditions, focusing mainly on the lift-to-drag ratio. The higher that number, the more efficiently a glider can move through water with minimal energy use.
“Lift-to-drag ratios are key for flying planes,” said Niklas Hagemann, an MIT graduate student and co-lead on the project. “Our pipeline modifies glider shapes to find the best lift-to-drag ratio, optimizing its performance underwater.”
This optimization required the AI to understand both the physics of movement and the geometry of design. To do that, the team built a “deformation cage.” This cage is a mathematical framework for bending and stretching simple shapes like ellipsoids into something new. They trained their system using 20 base models, including whales, sharks, and submarines, and then generated hundreds of variations.
The result: gliders with completely novel designs that might never have occurred to a human engineer.
From Simulation to Submersion
Of course, simulation is only the first step. To see if their algorithm’s creations could swim, the researchers picked two of their best-performing models and built them using 3D printers. The components were fabricated as hollow shells designed to flood with water, making them light and easy to assemble around a standard internal hardware unit.
This modular tube, shared between designs, contains a buoyancy engine, a mass shifter, and control electronics. By pumping water in or out, the glider can rise or fall. By shifting an internal weight forward or backward, it adjusts its angle through the water. These subtle controls mimic how real gliders move through the ocean without motors or propellers.
The researchers first tested their designs in MIT’s Wright Brothers Wind Tunnel. Simulations and real-world data aligned closely—only about a 5% difference in predicted vs. actual lift-to-drag ratios.
Then they took to the pool.
Two gliders were tested: one with two wings optimized for a 9-degree descent, the other with four fins designed for a steeper 30-degree glide. Both outperformed a conventional torpedo-shaped glider. The two-winged model achieved a lift-to-drag ratio of 2.5. For comparison, a standard handmade design tested in earlier studies achieved just 0.3.
“With higher lift-to-drag ratios than their counterpart, both AI-driven machines exerted less energy, similar to the effortless ways marine animals navigate the oceans,” the team wrote.
A Blueprint for the Future of Ocean Robotics
Autonomous gliders have become essential tools in modern oceanography. They gather long-term data on salinity, temperature, and currents, sometimes traveling thousands of miles without needing external power. Making these vehicles more efficient means more data, longer missions, and less cost.
But the team isn’t stopping here. One of the challenges, Chen said, is that the current design system struggles with very thin geometries—something that could unlock even more efficient forms. They’re also working on reducing the gap between simulation and reality. Small surface imperfections and mechanical components, like holes for flooding and screws, can affect real-world performance in ways simulations don’t always capture.
“The glider doesn’t perform as efficiently in reality as what was modeled in the simulation,” the authors noted in their research paper. “We attribute this gap primarily to frictional forces caused by the surface shear stresses of surface details in the real-world fabricated glider shells.”
To address this, they hope to incorporate those small-scale factors into future versions of the simulation. They also want to build gliders that can better react to changes in their environment—an important step if they are to navigate open seas, not just controlled tanks.
Eventually, this design framework could become an off-the-shelf solution for oceanographers. Imagine an interface where a scientist specifies a mission—depth, range, duration—and the AI responds with a custom-designed glider, printable and ready to launch.
Since the early 2000s, underwater gliders like the Slocum and Seaglider have become staples of ocean research. But their forms have remained largely static—cylindrical, conservative, familiar.
This work marks a departure from that lineage. It’s not just about making gliders faster or cheaper. It’s about creating shapes that were previously unimaginable —forms inspired by biology, engineering, and a machine’s own evolving intuition of physics.
