MIT's new AI outsmarts human design to help robots jump 41% higher

MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) has developed a new generative AI approach that designs robots capable of jumping 41% higher than those created by human engineers. Using diffusion-based generative models, researchers allowed the AI to modify specific parts of a 3D robot model, resulting in curved linkages resembling thick drumsticks rather than the straight, rectangular parts of traditional designs. This unique shape enabled the robot to store more energy before jumping, improving performance without compromising structural integrity. The AI-assisted robot also demonstrated an 84% reduction in falls compared to the baseline model, highlighting enhanced stability and landing safety. The process involved iterative refinement, with the AI generating multiple design drafts that were scaled and fabricated using 3D-printable polylactic acid material. Researchers believe that future iterations using lighter materials could achieve even higher jumps. Beyond jumping robots, the team envisions applying diffusion models to optimize how parts connect and to design robots with more complex capabilities, such as directional control and