Marine researchers have frequently expressed awe at how creatures like fish and seals propel themselves so effectively, even with varying forms. Their physiques are tailored for swift, hydrodynamic movement in water, allowing them to use minimal energy while covering substantial distances.
Autonomous craft can glide across the ocean in a comparable manner, gathering information about vast underwater realms. Nonetheless, the configurations of these gliding machines are less varied than those in marine organisms—typical designs often take the form of tubes or torpedoes, as they are quite hydrodynamic too. Moreover, testing new prototypes necessitates considerable real-world trial and error.
Scholars from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) along with the University of Wisconsin at Madison suggest that AI could facilitate the exploration of innovative glider designs more efficiently. Their technique employs machine learning to evaluate various 3D forms within a physics simulator, subsequently refining them into more hydrodynamic configurations. The resulting model can be produced using a 3D printer with significantly reduced energy compared to handcrafted versions.
The MIT scientists assert that this design process could yield new, more effective machines that aid oceanographers in measuring water temperature and salinity, accumulating deeper insights about currents, and observing the consequences of climate change. The team showcased this potential by creating two gliders approximately the dimensions of a boogie board: one featuring two wings akin to an airplane, and another unique four-winged model resembling a flat fish with four fins.
Peter Yichen Chen, a postdoctoral researcher at MIT CSAIL and co-lead author of the study, emphasizes that these models are just a selection of the innovative shapes that his team’s method can produce. “We’ve established a semi-automated workflow that enables us to assess unconventional designs that would be exceedingly taxing for humans to generate,” he notes. “This degree of shape variety hasn’t been previously explored, so most of these designs remain untested in the field.”
But how did AI originate these concepts in the first place? Initially, the researchers sourced 3D representations of over 20 typical marine exploration shapes, including submarines, whales, manta rays, and sharks. They then encased these models within “deformation cages” that outline various articulation points, which the researchers manipulated to create new silhouettes.
The CSAIL-led group established a dataset of conventional and altered shapes before simulating their performances at various “angles-of-attack”—the tilt direction a vessel assumes while gliding through water. For instance, a swimmer might choose to dive at a -30 degree angle to retrieve an object from a pool.
These varied shapes and angles of attack served as inputs for a neural network that effectively predicts how efficiently a glider’s form will perform at specific angles and optimizes it as necessary.
Elevating gliding robots
The team’s neural network simulates how a particular glider would respond to underwater dynamics, aiming to capture its forward motion and the resistance acting against it. The objective is to determine the optimal lift-to-drag ratio, which indicates how much a glider is elevated compared to the drag it encounters. The greater the ratio, the more efficiently the vehicle moves; conversely, a lower ratio results in slower travel during its journey.
Lift-to-drag ratios are crucial for aviation: during takeoff, it’s essential to maximize lift for smooth gliding against wind currents, while landing necessitates enough force to halt safely.
Niklas Hagemann, an architecture graduate student at MIT and CSAIL affiliate, argues that this ratio is equally beneficial if one desires similar gliding motion in aquatic environments.
“Our workflow adapts glider forms to discover the optimal lift-to-drag ratio, enhancing its underwater efficiency,” explains Hagemann, who is also a co-lead author of a study that was presented at the International Conference on Robotics and Automation in June. “You can then export the most efficient designs for 3D printing.”
Aiming for a quick glide
While their AI workflow appeared plausible, the researchers needed to validate its predictions regarding glider performance through experiments in more realistic conditions.
They first crafted their two-wing design as a smaller vehicle mimicking a paper airplane. This glider was tested in MIT’s Wright Brothers Wind Tunnel, an indoor facility equipped with fans to simulate wind currents. Positioned at various angles, the glider’s estimated lift-to-drag ratio was only around 5 percent higher on average than those recorded during wind experiments—a minor discrepancy between simulation and reality.
A digital assessment utilizing a visual, more intricate physics simulator also corroborated that the AI workflow yielded fairly precise forecasts about how the gliders would perform. It illustrated their descent in 3D.
To comprehensively assess these gliders in real-world conditions, however, the team needed to analyze how their devices would behave underwater. They printed two designs that performed optimally at specified angles-of-attack for this evaluation: one jet-like model at 9 degrees and the four-wing design at 30 degrees.
Both forms were constructed in a 3D printer as hollow shells with small openings that flood when fully submerged. This lightweight construction eases handling above water and minimizes material usage during fabrication. The researchers embedded a tube-like device within these shell covers, which contained a variety of hardware, including a pump to regulate the glider’s buoyancy, a mass shifter (a tool governing the angle-of-attack), and electronic components.
Each design surpassed a manually crafted torpedo-shaped glider by moving more effectively across a pool. With higher lift-to-drag ratios than their conventional counterpart, both AI-designed machines expended less energy, akin to the effortless navigation of marine animals across the oceans.
While this project marks an exciting advancement in glider design, the researchers aim to close the gap between simulation and actual performance. They are also eager to develop machines capable of reacting to sudden changes in currents, enhancing the gliders’ adaptability to various aquatic environments.
Chen adds that the team seeks to investigate new types of shapes, particularly slimmer glider designs. They intend to expedite their framework, potentially enhancing it with new features that facilitate more customization, agility, or even the development of miniature vehicles.
Chen and Hagemann co-led this research project along with OpenAI researcher Pingchuan Ma SM ’23, PhD ’25. They co-authored the paper with Wei Wang, an assistant professor at the University of Wisconsin at Madison and recent CSAIL postdoc; John Romanishin ’12, SM ’18, PhD ’23; and two professors from MIT affiliated with CSAIL: lab director Daniela Rus and senior author Wojciech Matusik. Their work received partial support from a Defense Advanced Research Projects Agency (DARPA) grant and the MIT-GIST Program.