Video: MIT AI co-designs a jumping robot that outperforms its human-made twin

Generative AI is no longer just a tool for digital creativity—it’s now helping design real-world robots. Researchers at MIT’s Computer Science & Artificial Intelligence Laboratory (CSAIL) have unveiled a new diffusion-model-powered system that lets users sketch a robot and optimize its design with AI. By specifying modifiable parts and dimensions, the AI generates and tests new structures in simulation before fabrication. In a recent test, the AI-designed robot jumped 41 percent higher than a human-designed counterpart, thanks to curved, drumstick-like linkages. According to researchers, the breakthrough marks a significant step toward automating and enhancing robot design, blending human intuition with machine-generated precision.

AI Refines Mechanics

Although robotic design is advancing, hardware development is trailing software because of its intricacy and reliance on multiple disciplines. The sampling-based methods and simulation tools facilitate iteration and co-design, but difficulties persist in incorporating fabrication constraints and traditional manufacturing due to limitations in data and standardization. Researchers at MIT CSAIL have now created a system based on diffusion models that aids in the design and optimization of robots.

The system enables users to provide a 3D model and specify the components they wish the AI to alter. The model proposes and simulates design variations, facilitating the process from concept to 3D-printed prototype. The team set out to enhance the performance of a jumping robot in one test. They began with a preliminary design and, using a guiding embedding vector, sampled 500 variants. From these, they chose the top twelve based on simulation results and employed them to refine the vector. This iterative process was repeated five times, resulting in a final design with curved, drumstick-like linkages.

Despite appearing nearly identical to a baseline model, the AI-enhanced version jumped 41 percent higher. Not only was the AI’s unconventional design lighter, but it also had a greater energy capacity while maintaining its strength. Researchers also used the same method to improve the design of the robot’s foot for landing. The 84 percent improvement in landing stability due to the new foot design has led to a significant reduction in falls.

Redesigning Robot Motion

Researchers at MIT CSAIL used generative AI to find a balance between these competing goals to build a robot capable of high jumps and stable landings. They assigned numerical values to jumping height and landing success rate, then trained a diffusion model to enhance a design that excelled in both aspects. The AI produced a 3D structure that outperformed a manually designed robot by investigating the area between the two goals. The prototype was made with materials suitable for 3D printing, but the team thinks that using lighter, more advanced materials in the future could greatly enhance performance.

The team sees this work as a foundation for broader applications in robotics. “We want to branch out to more flexible goals. Imagine using natural language to guide a diffusion model to draft a robot that can pick up a mug or operate an electric drill,” said Tsun-Hsuan Wang, a PhD student at MIT CSAIL and co-lead author of the research, in a statement. Researchers add that diffusion models could eventually aid in creating articulation mechanisms and improving the connections between components, thereby enhancing robot performance. The group is also looking into the possibility of using extra motors to manage jump direction and enhance landing precision, underscoring AI’s growing importance in the development and performance of robots.