Dyfusion models, such as Dall-E Openai, are becoming more and more useful in helping brainstorming new designs. People can make these systems to generate image, create a film or improve the plan and return with ideas that they did not consider before.
But do you know that generative models of artificial intelligence (Genai) also make progress in creating robots? Last Diffusion -based approaches generated structures and systems that control them from scratch. At the user's entry or without it, these models can create new projects, and then evaluate them in the simulation before making them.
The new approach of the Computer Science and Artificial Intelligence Laboratory (CSAIL) uses this generative knowledge to improve human robotic projects. Users can develop a 3D robot model and determine which parts would like to see the modification of the diffusion model, ensuring its dimensions earlier. Then the brainstorming genai optimal shape of these areas and tests its ideas in the simulation. When the system finds the right design, you can save and then produce a working, actual robot with a 3D printer, without the requirement of additional corrections.
Scientists applied this approach to creating a robot, which jumps on an average of about 2 feet, i.e. 41 percent higher than a similar machine they created. The machines are almost identical in terms of appearance: both are made of a kind of plastic called polylactic acid, and although at first they seem flat, they grow in the shape of a diamond when the engine stretches to the attached string. So what exactly did AI do?
A closer look reveals that the connections generated by AI are curved and resemble thick drums (they use musical drummers of instruments), while the parts connecting standard work are simple and rectangular.
Better and better stains
Scientists began to improve their jumping robot, trying 500 potential projects using the initial numerical cessation vector, which records high-level features to conduct projects generated by the AI model. Of these, they chose the 12 best options based on simulation performance and used them to optimize the prisoning vector.
This process was repeated five times, gradually conducting the AI model to generate better projects. The resulting project resembled a blob, so scientists prompted their system to scale the sketch to the 3D model. Then they created a shape, stating that it actually improved the ability to jump the robot.
The advantage of using diffusion models for this task, according to the author of the coefficient and CSail Postdoc Byungchul Kim, is that they can find unconventional solutions to improve robots.
“We wanted our machine to be higher, so we thought that we could simply make the links connecting its parts as thin as possible to make them light,” says Kim. “However, such a thin structure can easily crack if we only use 3D printed material. Our diffusion model came up with a better idea, suggesting a unique shape that allowed the robot to store more energy before jumper, not making the links too thin. This creativity helped us learn about the basic physics of the machine.”
The team then submitted to their system to develop an optimized foot to ensure safe landing. They repeated the optimization process, ultimately choosing the best efficient construction to be attached to the bottom of their machine. Kim and his colleagues said that their machine designed by AI fell much less often than the base line, in the amount of 84 % improvement.
The ability of the diffusion model to improve the skills of jumping and landing the robot suggests that it can be useful in improving the way other machines designed. For example, a company working on production robots or households could apply a similar approach to improving its prototypes, saving the time of engineers usually reserved for the iteration of these changes.
Balance behind the reflection
To create a robot that can jump stable and stable, researchers decided that they must maintain a balance between the two goals. They represented both the height of the jumping and the success rate of landing as numerical data, and then trained their system to find a sweet point between the two embedding vectors that could help build the optimal 3D structure.
Scientists note that although this robot supported by AI has surpassed its counterpart designed by man, it can soon achieve even greater new heights. This iteration included the use of materials compatible with a 3D printer, but future versions would jump even higher with lighter materials.
Co-chairman author and PhD student MIT CSAIL, TSUN-HSUAN “Johnson” Wang claims that the project is a jumping point for new robotics projects in which generative artificial intelligence can help.
“We want to branch for more flexible purposes,” says Wang. “Imagine using a natural language to lead the diffusion model to design a robot that can pick up a cup or support an electric drill.”
Kim claims that the diffusion model can also help generate articulation and soles, how parts connect, potentially improving how high the robot would jump. The team also examines the possibility of adding more engines to control, in which direction the machine jumps and may improve landing stability.
Scientists' work was partly supported by Emerging Frontiers in Research and Innovation Program National Science Foundation, a research alliance program and technology, Manus and Machina, and Gwangju Institute of Science and Technology (GIST) -CSail. They presented their work at an international conference on robotics and automation in 2025.
