MIT Researchers Combine GenAI and Physics Simulation Engine to Revolutionize Robot Design

Researchers at the Massachusetts Institute of Technology’s Computer Science and Artificial Intelligence Laboratory have integrated generative artificial intelligence tools with a physics simulation engine to enhance robot designs. The outcome is a robot that outperforms its human-engineered counterparts.

Thanks to diffusion-based techniques, GenAI models can generate new designs and evaluate them through simulations. This innovative approach leverages these capabilities to improve human-created robot structures. Users can craft a 3D model of a robot and specify which components they wish to modify using a diffusion model, providing preliminary dimensions. Following this, GenAI conducts brainstorming sessions and tests its suggestions via simulation. Once the system identifies an optimal design, it can be saved and subsequently 3D printed into a functioning robot without requiring further adjustments.

This method enabled the researchers to develop a robot that can jump an average of 0.6 meters, which is 41% higher than a similar machine they had previously built. Both robots appear almost identical, as they are made from a type of plastic known as polylactic acid.

Upon closer inspection, however, the AI-designed robot’s connections are curved and resemble thick drumsticks, while the standard robot’s joints are straight and rectangular.

The researchers began improving their jumping robot by selecting 500 potential designs using an initial embedding vector—a numerical representation capturing high-level features for optimal choice. They narrowed it down to 12 top performers based on simulation results and used them to refine the embedding vector.

This optimization process was repeated five times, gradually steering the AI model toward superior designs. The final version took on a droplet shape, prompting the researchers to encourage the system to scale the design to match their 3D model. Upon creating a mold, they discovered that it significantly enhanced the robot’s jumping capabilities.

One key advantage of using diffusion models is their ability to identify unconventional solutions for robot enhancements.

«We aimed to make our machine jump higher, so we thought we could simply make the links connecting its parts as thin as possible to reduce weight. However, such a delicate structure could easily break with only 3D-printed materials. Our diffusion model proposed a unique shape that allowed the robot to store more energy before jumping without overly thinning its links. This creativity deepened our understanding of the machine’s underlying physics,» the authors remarked.

Following this, the team tasked the system with developing an optimized robotic leg for secure landings. They repeated the optimization process, ultimately selecting the most effective design. The researchers found that the AI-designed machine experienced significantly fewer falls than the base version, showing an 84% improvement in this area.

To create a robot capable of high jumps and stable landings, the researchers sought to balance both objectives. They represented jump height and landing success percentages as numerical values and subsequently trained their system to find a middle ground between the two vectors.

Although the AI-designed robot surpassed the human-engineered version, further improvements are planned. The current iteration utilized 3D printer-compatible materials, but future versions are expected to jump even higher by incorporating lighter materials.

The researchers believe that this project serves as a foundation for developing new robotic designs aided by generative AI. «Imagine using natural language to control a diffusion model in designing a robot capable of lifting a mug or operating a power drill,» they suggested.

In 2022, roboticists from CSAIL also employed AI-based simulations to quickly train a cheetah robot to adapt its walking style to various conditions, enabling it to achieve a new speed record.