MIT Researchers Use Generative AI to Design Robots That Jump Higher and Land More Safely

MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) has developed a new approach leveraging generative artificial intelligence (GenAI) to optimize robotic designs, specifically focusing on enhancing jumping and landing capabilities. This innovative method allows users to draft a 3D model of a robot and specify which parts they want the AI to modify. By providing the dimensions and desired performance metrics, the GenAI model generates and evaluates multiple design iterations in simulation, ultimately producing an optimized design that can be 3D printed and tested in the real world. The researchers began by using an initial embedding vector—a numerical representation capturing high-level design features—to sample 500 potential robot designs. From these, they selected the top 12 based on performance in simulations and used them to further refine the embedding vector. This iterative process was repeated five times, each time improving the design quality. Initially, the best designs resembled irregular blobs, but the researchers then scaled them to fit their 3D model. The resulting AI-generated design enabled the robot to jump roughly 2 feet, a 41% improvement over a manually designed version. A key aspect of the AI-generated design was the curvature of the linkages, which resembled thick drumsticks. This unique shape allowed the robot to store more energy before jumping, enhancing its height without compromising structural integrity. Co-lead author Byungchul Kim, a postdoc at CSAIL, highlighted that the diffusion model's ability to find unconventional solutions provided valuable insights into the underlying physics of the robot. Next, the researchers focused on optimizing the robot's landing stability. They repeated the simulation and refinement process, eventually selecting the best-performing design for the robot's feet. The AI-designed foot led to an 84% improvement in landing stability compared to the baseline design, significantly reducing the frequency of falls. This balance between jumping height and landing stability was achieved by representing both goals as numerical data. The AI model was trained to find the optimal intersection point between the two, leading to a 3D structure that excelled in both aspects. The implications of this research extend beyond just the jumping robot. Diffusion models can potentially revolutionize the design process for various types of robots, from manufacturing to household applications, by automating the generation and testing of design modifications. This could save engineers considerable time and resources, streamlining the development of efficient and robust robotic systems. The research was supported by the National Science Foundation's Emerging Frontiers in Research and Innovation program, the Singapore-MIT Alliance for Research and Technology’s Mens, Manus and Machina program, and the Gwangju Institute of Science and Technology (GIST)-CSAIL Collaboration. The findings were presented at the 2025 International Conference on Robotics and Automation. Industry insiders are excited about the potential of GenAI in robotics. The ability to swiftly generate and test innovative designs could accelerate the development of new technologies, making it easier to create robots that perform complex tasks with greater efficiency and reliability. Companies like Boston Dynamics and iRobot may find significant value in integrating such AI tools into their design processes. MIT’s CSAIL, founded in 1963, is one of the world's premier research institutions in the field of computer science and artificial intelligence. Known for groundbreaking work in machine learning, robotics, and human-computer interaction, CSAIL continues to push the boundaries of what AI can achieve, particularly in practical applications. This latest research exemplifies the lab's commitment to advancing the intersection of AI and engineering, with the potential to transform how robots are designed and built in the future.