AI shapes autonomous underwater “gliders”

Marine scientists have long admired the remarkable efficiency with which aquatic animals like fish and seals navigate the ocean. Their natural forms are exquisitely optimized for hydrodynamic movement, allowing them to traverse vast distances with minimal energy expenditure. In contrast, autonomous underwater vehicles (AUVs) designed to drift through the ocean and collect invaluable data about marine environments have traditionally been limited to familiar tube or torpedo shapes, which, while hydrodynamic, lack the diverse efficiency seen in nature. Moreover, developing and testing new AUV designs has historically relied on arduous real-world trial-and-error.

A groundbreaking collaboration between researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the University of Wisconsin at Madison is poised to transform this landscape. They propose a novel approach leveraging artificial intelligence (AI) to conveniently explore and optimize uncharted underwater glider designs. Their innovative methodology employs machine learning to virtually test various 3D designs within a physics simulator, subsequently molding them into significantly more hydrodynamic shapes. The resulting optimized models can then be fabricated with remarkable efficiency using 3D printing technology, requiring substantially less energy than traditional hand-made methods.

This pioneering design pipeline, as highlighted by the MIT scientists, holds immense potential for creating new, highly efficient machines. These advanced gliders could empower oceanographers to precisely measure crucial parameters like water temperature and salt levels, gather more detailed insights into ocean currents, and meticulously monitor the profound impacts of climate change on marine ecosystems. To demonstrate this capability, the team successfully produced two distinct gliders, each approximately the size of a boogie board: one resembling an airplane with two wings, and another unique, four-winged object strikingly similar to a flat fish with multiple fins.

Peter Yichen Chen, an MIT CSAIL postdoc and co-lead researcher on the project, emphasized the vast potential of their approach. “We’ve developed a semi-automated process that can help us test unconventional designs that would be very taxing for humans to design,” he stated. “This level of shape diversity hasn’t been explored previously, so most of these designs haven’t been tested in the real world.” This underscores the human-like innovation AI brings to complex engineering challenges.

But how does AI conceive such novel forms? The process begins with the researchers acquiring 3D models of over 20 conventional sea exploration shapes, ranging from submarines to various marine animals like whales, manta rays, and sharks. These models are then enclosed within “deformation cages,” digital frameworks that map out specific articulation points. Researchers manipulate these points to generate an array of new, diverse shapes.

The CSAIL-led team subsequently constructed a comprehensive dataset comprising both conventional and deformed shapes. They then simulated the performance of these designs at various “angles-of-attack,” which dictate the vessel’s tilt as it glides through the water. For instance, a glider might be tested at a -30-degree angle to mimic a deep dive. These diverse shapes and angles of attack served as critical inputs for a neural network, which effectively predicts how efficiently a glider shape will perform at particular angles and then intelligently optimizes it for superior performance.

Giving Gliding Robots a Lift

Central to this innovation is the team’s neural network, which simulates how a specific glider would interact with underwater physics, focusing on its forward movement and the resistive force (drag) against it. The primary objective is to identify the optimal lift-to-drag ratio. This ratio indicates how effectively the glider is supported or “lifted” through the water compared to the forces holding it back. A higher ratio signifies more efficient travel, while a lower one indicates increased resistance and slower movement during its voyage.

Niklas Hagemann, an MIT graduate student in architecture and CSAIL affiliate, and a co-lead author on the paper presented at the International Conference on Robotics and Automation in June, highlighted the significance of this ratio. “Our pipeline modifies glider shapes to find the best lift-to-drag ratio, optimizing its performance underwater,” Hagemann explained. “You can then export the top-performing designs so they can be 3D-printed.”

Going for a Quick Glide

While the AI pipeline provided realistic predictions, the researchers recognized the necessity of validating its accuracy in more lifelike environments. Their initial fabrication involved a scaled-down version of the two-wing design, resembling a paper airplane. This model underwent testing in MIT’s Wright Brothers Wind Tunnel, an indoor facility simulating wind flow. Remarkably, the glider’s predicted lift-to-drag ratio differed by only about 5 percent on average from the values recorded in the actual wind experiments, indicating a minimal disparity between simulation and reality.

Further digital evaluation using a more complex visual physics simulator also corroborated the AI pipeline’s accuracy in predicting glider movement, visualizing their 3D descent paths. For a true real-world assessment, the team moved to underwater trials. They 3D-printed two designs that demonstrated peak performance at specific angles-of-attack: a jet-like device optimized for 9 degrees and the distinctive four-wing vehicle for 30 degrees.

These prototypes were fabricated as hollow shells with small holes to allow for complete submersion, making them lightweight and requiring less material. A central tube-like device housed the essential hardware, including a pump for buoyancy control, a mass shifter to manage the glider’s angle-of-attack, and various electronic components. In pool tests, both AI-driven designs significantly outperformed a handmade torpedo-shaped glider. Their higher lift-to-drag ratios translated into more efficient movement, mirroring the seemingly effortless navigation of marine animals.

While this project marks a significant stride in glider design, the researchers are dedicated to further narrowing the gap between simulated and real-world performance. Future work also includes developing machines capable of reacting to sudden changes in ocean currents, enhancing the gliders’ adaptability. Chen added that the team plans to explore new, particularly thinner, glider designs and aims to accelerate their framework, potentially integrating features for greater customization, maneuverability, or even the creation of miniature autonomous vehicles.

Peter Yichen Chen and Niklas Hagemann co-led this research alongside OpenAI researcher Pingchuan Ma SM ’23, PhD ’25. The paper was co-authored by Wei Wang, a University of Wisconsin at Madison assistant professor and former CSAIL postdoc; John Romanishin ’12, SM ’18, PhD ’23; and two distinguished MIT professors and CSAIL members: lab director Daniela Rus and senior author Wojciech Matusik. Their groundbreaking work received support, in part, from a Defense Advanced Research Projects Agency (DARPA) grant and the MIT-GIST Program.