Design of the Times
MIT engineers showed how GenAI can be used to co-design robots that perform much better than those developed by humans alone.
They may not be so likely to add extra arms or fingers to images as they were a few years back, but even still, generative artificial intelligence (GenAI) algorithms are not exactly the tools that we want designing real-world machines for us. Or are they? According to a trio of engineers at MIT, GenAI is the perfect way to design a robot. Using a novel approach, they have shown that AI-generated robot designs can actually significantly outperform designs crafted by human engineers.
The researchers’ GenAI-driven framework makes use of diffusion models, which are a type of generative algorithm best known for image creation, to design physical robots. Unlike traditional design methods, this system starts with a base 3D model from a human engineer, along with specified areas for modification. The diffusion model then iteratively generates alternative geometries, evaluates them in simulation, and proposes optimal structures that are ready for fabrication without further editing.
To validate their approach, the researchers applied their system to the design of a jumping robot. The AI-enhanced version of the robot achieved a 41% higher jump (an average of about two feet) compared to the baseline human-designed model. The secret to their success was geometry. While the standard design used straight, rectangular linkages, the AI-generated version featured curved, drumstick-shaped connectors that allowed the robot to store and release more energy during jumps, all without compromising structural integrity.
To achieve this goal, the team first generated 500 candidate designs using a latent embedding vector (a numerical representation of high-level structural features), then selected the top performers based on simulation results. This process was repeated five times, each cycle refining the vector to guide the diffusion model toward higher-performance outputs. The final design was scaled to match the robot's physical constraints and fabricated with a 3D printer.
The same process was then applied to improve the robot’s landing capabilities. The researchers optimized the design of its foot to increase stability, achieving a success rate of 88% compared to only 4% for the original, human-designed model. By focusing on both jump height and landing stability, the researchers demonstrated the ability of GenAI to co-design complex systems where multiple objectives must be balanced.
Beyond jumping robots, the team envisions broader applications for their technology in the future. In coming iterations, the system might allow users to simply describe a robot’s task, like gripping a mug or drilling a hole, and let the AI design the structure accordingly. The team is also exploring ways to enhance the system with more motors and additional control logic.
By blending human intuition with the unbounded creative potential of GenAI, this approach could significantly accelerate hardware development, making robot design more accessible and efficient than ever before.