Morscy scientists have long been surprising how animals such as fish and seals swim so effectively, even though they have different shapes. Their bodies are optimized for efficient, hydrodynamic water navigation so that they can exert minimal energy when traveling over long distances.
Autonomous vehicles can drift through the ocean in a similar way, collecting data on extensive submarines. However, the shapes of these gliding machines are less diverse than in maritime life-projects often resemble tubes or torpedoes, because they are also quite hydrodynamic. In addition, testing new compilations requires many real trials and errors.
Scientists from Mit's Computer Science and Artificial Intelligence Laboratory (CSAIL) and the University of Wisconsin in Madison suggest that AI can help us more conveniently examine glider designs. Their method uses machine learning to test various 3D projects in a physics simulator, and then forms them in more hydrodynamic shapes. The resulting model can be produced using a 3D printer using much less energy than handmade.
MIT scientists say that this design pipeline can create new, more efficient machines that help oceanographers measure the level of water and salt temperature, collect more detailed information on currents and monitor the impact of climate change. The team demonstrated this potential, producing two gliders about the size of the boogie table: a two -story aircraft reminiscent of a plane and a unique, four -time flat -like object with four fins.
Peter Yichen Chen, the myth of Csail Postdoc and a contemporary project researcher, notes that these projects are just some of the novels of shapes that can generate the approach of his team. “We have developed a semi -automatic process that can help us test unconventional projects that would be very taxed for people,” he says. “This level of shape variety was not previously studied, so most of these projects were not tested in the real world.”
But how did AI come up with these ideas? First of all, scientists found 3D models with over 20 conventional shapes of the sea exploration, such as submarines, whales, manta rays and sharks. Then they closed these models in “deformation cages”, which are mapping various articulation points that scientists stopped to create new shapes.
The team managed by CSAIL built a set of conventional data and deformed shapes before simulating how they will work at different “angle of attractions”-a service in which the ship will tilting when it moves on the water. For example, a swimmer may want to dive at an angle of -30 degrees to download the item from the pool.
These diverse shapes and angles of the attack were then used as input data for the neural network, which basically predicts how effectively the shape of the glider will work at specific angles and optimizes it if necessary.
Raising sliding robots
The neuronal network of the band simulates how a given glider would react to underwater physics, aimed at grasping how it moves forward and the strength that pulls. Goal: Find the best raising attitude to vibration, representing how much the glider is held compared to how much it refrigerates. The higher the ratio, the more efficiently the vehicle travels; The more it is, the more the glider will slow down while traveling.
Vibration lifting coefficients are crucial for flying aircraft: during the start you want to maximize the elevator to make sure that it can move well with wind currents, and during landing you need enough strength to drag him to full stop.
Niklas Hagemann, a graduate of the MIT in the field of architecture and the CSAIL association, notes that this attitude is equally useful if you want a similar exercise movement in the ocean.
“Our pipeline modifies the shape of the glider to find the best raising ratio to vibration, optimizing its performance under water,” says Hagemann, who is also the author of the WA factor paper This was presented at an international conference on robotics and automation in June. “Then you can export the highest projects so that they can be printed 3D.”
Selecting quick slip
While their AI pipeline seemed realistic, scientists needed to assure that his glider performance forecasts were accurate, experimenting in more realistic environments.
First, they produced their duplicating structure as a returned vehicle resembling a paper plane. This glider was taken to the aerodynamic tunnel Mit Wright Brothers, rooms with fans simulating wind flow. The expected ratio of the glider to the comma, placed at different angles, was only about 5 percent higher than those registered in wind experiments-a great difference between simulation and reality.
The digital assessment covering a visual, more complex physics simulator also supported the view that the AI pipeline has made quite accurate forecasts about moving gliders. He visualized how these machines would fall into 3D.
However, to really assess these gliders in the real world, the team had to see how their devices would manage under water. They printed two projects that they best made at specific points of attractions for this test: a device similar to a jet at 9 degrees and a four -time vehicle at 30 degrees.
Both shapes have been fabricated in a 3D printer as empty shells with small holes that flood the fully immersed. This lightweight design makes the vehicle easier to use outside the water and requires the production of less material. Scientists placed a device in the tube in those shell covers, in which there was a number of equipment, including a pump to change the swimming pool, gear shift weight (device controlling the machine's viewing angle) and electronic components.
Each project was exceeded by a handmade torpedo glider, moving more around the pool. With higher lifting indicators than their counterpart, both machines based on AI exerted less energy, just like the effort in which sea animals navigate the oceans.
While the project is an encouraging step forward to designing gliders, scientists want to narrow the gap between the simulation and performance in the real world. They also hope to develop machines that can react to sudden changes in currents, thanks to which gliders are more adapted to the sea and oceans.
Chen adds that the band wants to discover new types of shapes, especially thinner glider patterns. They intend to speed up their frames, perhaps strengthening them with new functions that allow greater adaptation, maneuverability and even creating miniature vehicles.
Chen and Hagemann conducted research on this project with the Openai Pingchuan researcher MA SM '23, dr '25. They wrote an article from Wei Wang, the University of Wisconsin in Madison Assistant Professor and the recent Postdoc CSAIL; John Romanishin '12, SM '18, dr '23; and two MIT professors and CSAIL members: laboratory director Daniel Rus and senior author Wojciech Matusik. Their works were partly supported by the Grant Agency for Advanced Defense Research Projects (DARPA) and MIT-GIST program.
