1005 Alcyon Dr Bellmawr NJ 08031
+1 484-446-2566
support@bmcpros.com
Diffusion models such as Openai’s Dall-E are becoming increasingly useful in helping brainstorm new designs. Humans can prompt these systems to generate images, create videos, or perfect blueprints, and come back to ideas they have never considered before.
But did you know that generating artificial intelligence (Genai) models is also making progress in creating working robots? Recent diffusion-based approaches have generated structures and systems that control them from scratch. With or without user input, these models can develop new designs and then evaluate them in the simulation before manufacturing.
A new approach from the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) applies this generative knowledge to improve human robot design. Users can draft a 3D model of the robot and specify which parts of the diffusion model they want to see and provide their size in advance. Genai then brainstorms the best shapes for these fields and tests its ideas in the simulation. When the system finds the right design, you can save it and then build a working real-world robot with a 3D printer without making additional tweaks.
The researchers used this method to create a robot that jumped about 2 feet higher on average, 41% higher than similar machines they created themselves. The machines look almost the same: they are all made of a plastic called polylactic acid, and when they initially appear flat, they pop into diamond shapes when the motor pulls onto the wires connected to them. So what is the difference between AI?
A closer look shows that the AI-generated links are curved, similar to thick drum sounds (used by instrumental drummers), while the standard robots’ connecting parts are straight and rectangular.
Better spots
The researchers began sampling 500 potential designs using the initial embedding vector, a numerical representation that captured advanced features to guide the design generated by AI models. From these processes, they selected the first 12 options based on the performance of the simulation and used them to optimize the embed vector.
The process was repeated five times, gradually guiding the AI model to generate better designs. The resulting design resembles a spot, so the researchers prompted their system to expand the draft to fit their 3D model. Then they made the shape and found that it really improved the robot’s jumping ability.
According to co-leader author and Csail postdoc Byungchul Kim, the advantage of using the diffusion model for this task is that they can find unconventional solutions to perfect the robot.
“We wanted to make the machine jump high, so we thought we could make the links tie their parts as thin as possible to make them lightweight,” King said. “But if we only use 3D printing materials, this thin structure could easily break. Our diffusion model came up with a better idea, a unique shape that allows the robot to store more energy before jumping without making the link too thin. This creativity helped us understand the basic physics of the machine.”
The team then tasked its system by drafting optimized feet to ensure it landed safely. They repeated the optimization process and finally chose the best performing design to attach to the bottom of the machine. Kim and his colleagues found that their AI-designed machines had a drop frequency much lower than their baseline, with an amplitude of 84%.
The ability of the diffusion model to upgrade the robot’s jump and landing skills shows that this may be useful for enhancing the design of other machines. For example, companies working in manufacturing or home robots can use similar methods to improve their prototypes, saving time typically used to iterate over these changes.
The balance behind the bounce
To create a robot that can jump high and land stably, the researchers realized they needed to strike a balance between the two goals. They expressed the jump height and landing success rate as numerical data and then trained their system to find a sweet spot between the two embedding vectors, which could help build the optimal 3D structure.
The researchers noted that while the AI-assisted robot performs better than its artificially designed counterpart, it will soon reach new heights. This iteration involves the use of materials compatible with 3D printers, but the use of lighter materials will make future versions higher.
Co-leader writer and MIT CSAIL PhD student Tsun-Hsuan “Johnson” Wang said the project is the starting point for new robot design, and generative AI can help you.
“We want to achieve more flexible goals,” Wang said. “Imagine using a natural language-guided diffusion model to draft a robot that can pick up a cup or operate an electric drill.”
Diffusion models can also help generate pronunciations and conceive how parts are connected, potentially increasing the level of jumping for the robot, King said. The team is also exploring the possibility of adding more motors to control the direction of the machine’s jump and possibly improving its landing stability.
The National Science Foundation supports the work of researchers in emerging areas of the Research and Innovation Program, Singapore-INAPORE-MIT Men’s Research and Technology Program, Manus and Machina Program, and Gwangju Institute of Science and Technology (GIST)-CSAIL collaboration. They presented their work at the 2025 International Conference on Robotics and Automation.