Marine scientists have long been surprised at how efficiently they swim despite the different shapes of animals like fish and seals. Their bodies are optimized for efficient, fluid-dynamic aquatic navigation, allowing them to deliver minimal energy when traveling long distances.
Self-driving cars can drift through the ocean in a similar way and collect data on the vast underwater environment. However, the shapes of these gliders are less diverse than those found in marine life. This is often similar to a tube or torpedo because it is quite hydrodynamic. Additionally, testing new builds requires a lot of real-world trial and error.
Researchers at MIT's Computer Science and Artificial Intelligence Institute (CSAIL) and University of Wisconsin-Madison suggest that AI can help explore unknown glider designs more conveniently. Their methods use machine learning to test different 3D designs in a physics simulator and shape them into more hydrodynamic shapes. The resulting model can be manufactured via a 3D printer using significantly less energy than handmade energy.
MIT scientists say the design pipeline will allow oceanographers to create new, more efficient machines that will help oceanographers measure water temperature and salt levels, gather more detailed insights into flow, and monitor the impacts of climate change. The team demonstrated this possibility by producing two gliders of the size of a boogie board. It is a unique four-wing object that resembles a plane-like two-wing machine and a flat fish with four fins.
Peter Yichen Chen, MIT CSAIL POSTDOC and co-lead researcher of the project, points out that these designs are just a few of the new shapes his team's approach can create. “We have developed a semi-automated process that helps us test unconventional designs that are highly taxable for human design,” he says. “As this level of shape diversity has not been investigated previously, most of these designs have not been tested in the real world.”
But how did AI come up with these ideas in the first place? First, researchers discovered 3D models of more than 20 traditional ocean exploration shapes, including submarines, whales, mantas and sharks. These models were then surrounded by “deformation cages” and mapped the various joint points that the researchers had pulled to create new shapes.
The CSAIL-led team built a dataset of traditional and deformed shapes before simulating how it would work with different “attack angles.” For example, swimmers recommend diving at a -30 degree angle to get items from the pool.
These diverse shapes and attack angles were used as inputs for neural networks that essentially predicted how efficiently the glider shape would work at a particular angle and optimized as needed.
Glide robot gives a lift
The team's neural network aims to simulate how a particular glider responds to underwater physics and capture the forces that move forward and drag it. Target: Find the best lift-drug ratio that represents how much the glider is rising compared to the amount that the glider is restrained. The higher the ratio, the more efficient the vehicle will travel. The lower it is, the slower the glider will be during the voyage.
The lift-to-drug ratio of an airplane is key. At takeoff, you need to make sure you can maximize the lift and slide well against wind flow, and when you land, you need enough force to stop it completely.
Niklas Hagemann, a MIT graduate student at Architecture and CSAIL affiliate, notes that this ratio is also useful when similar gliding motions are required at sea.
“Our pipeline fixes glider shapes to find the best lift-to-drag ratio and optimizes performance underwater,” says Hageman, who is also co-lead author of a paper presented at the International Conference on Robotics and Automation in June. “Then you can export your top performance designs so that they can be 3D printed.”
I'll go for a simple gliding
Their AI pipeline seemed realistic, but researchers had to ensure that predictions about glider performance were accurate by experimenting in a more realistic environment.
They first built a two-wing design as a scaled vehicle similar to a paper plane. This glider was taken to the Light Brothers Wind Tunnel in Mitt. This is an indoor space with fans simulating the flow of wind. The predicted lift-to-drag ratios for gliders placed at different angles were about 5% higher on average than those recorded in wind experiments. This is the difference between simulation and reality.
Digital assessments, including visual and more complex physics simulators, also supported the notion that the AI pipeline made fairly accurate predictions about how the glider would move. We visualized how these machines descend in 3D.
However, to truly appreciate these gliders in the real world, the team had to see how their devices were carried underwater. They printed two designs that performed best in certain attacks in this test: a 9-degree jet-like device and a 30-degree four-wing vehicle.
Both shapes were manufactured on a 3D printer as hollow shells with small holes that flood when completely submerged. This lightweight design makes the vehicle easier to handle outside of water and requires less material to manufacture. Researchers placed tube-like devices inside these shell covers. This housed a variety of hardware, including pumps that modify the buoyancy of the glider, mass shifters (device that controls the machine's attack angle), and electronic components.
Each design outperformed the handmade torpedo gliders by moving the pool more efficiently. With a higher lift-to-drag and drag ratio than the counterpart, both AI-driven machines have less energy, just like the easy way marine animals navigate the ocean.
As long as the project is an encouraging step for glider design, researchers aim to narrow the gap between simulation and real-world performance. They also hope to develop machines that can respond to sudden changes in flow, and to adapt gliders more to the ocean and ocean.
Chen added that the team is looking to explore new types of shapes, especially thinner glider designs. They intend to make the framework faster and perhaps enhance it with new features that allow for more customization, ease of use, or miniature cars.
Chen and Hagemann are co-led researchers with Openai researchers Pingchuan Ma SM '23 and PhD '25. They wrote a paper with Wei Wang, assistant professor at Madison University and a recent CSAIL Postdoc's University of Wisconsin. John Romanishin '12, SM '18, PhD '23; two MIT professors and CSAIL members: Director Daniela Lu, Laborecectar, and senior author Wojciech Matusik. Their work was supported in part by the Defence Advanced Research Projects Agency (DARPA) grant and the MIT-GIST program.
