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

Researchers at MIT’s Computer Science & Artificial Intelligence Laboratory (CSAIL) have developed a novel system leveraging generative AI, specifically diffusion models, to co-design robots by optimizing their mechanical structures through iterative simulation and fabrication. This approach allows users to input a 3D model and specify modifiable components, enabling the AI to generate and test numerous design variants before physical production. In a key demonstration, the AI-designed jumping robot outperformed a human-designed counterpart by jumping 41% higher, achieved through innovative curved, drumstick-like linkages that were lighter yet stronger and had greater energy capacity. Additionally, the AI improved the robot’s foot design, enhancing landing stability by 84%, significantly reducing falls. This breakthrough highlights the potential of diffusion models to reveal new insights into structural physics and accelerate hardware design processes, which traditionally lag behind software development due to complexity and manufacturing constraints. By balancing competing objectives such as jump height and landing success, the AI system produced optimized designs that blend human intuition with machine precision.