Imagine handing the reins of robot design to a creative partner that not only dreams up new ideas, but can also bring them to life with more power and precision than ever before. That’s precisely what MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) has done by putting generative AI—think of the engine behind tools like DALL·E—front and center in the process of building working robots.
For years, AI was mostly used for media, like generating pictures or videos. But now, sophisticated tools called diffusion models are stepping right into the world of physical engineering. Instead of relying on slow, repetitive trial-and-error, CSAIL’s new approach lets an AI explore thousands of virtual tweaks to a robot’s design, test them all in simulation, and quickly settle on the most promising versions. Only then do the researchers head to the 3D printer, significantly speeding up what used to be a major bottleneck in robotics.
To show off just how game-changing this can be, the MIT team set their sights on a simple, single mission: get a robot to jump higher. They started with a basic 3D model and told the AI to optimize key parts. The AI responded by reimagining aspects of the robot’s “links”—think of them as the bones or arms—and then virtually testing hundreds of design variations. Once the winning blueprint emerged, it was printed in polylactic acid plastic and put to the test in the real world.
Here’s the jaw-dropping part: the AI-enhanced robot launched itself nearly 2 feet into the air—41% higher than a similar bot designed purely by human intuition and engineering know-how. At first glance, you’d be hard-pressed to tell the two bots apart. Both are powered by a motor and a cord-driven spring action, sharing the same basic materials. But instead of straight, rectangular connectors, the AI offered up curved, drumstick-shaped links. This unusual geometry, it turns out, allowed the robot to store and unleash much more energy with each jump, all while keeping those crucial links sturdy enough to survive the landing.
The path to this success wasn’t a simple one. The team used an AI mapping called an “embedding vector” to generate 500 new design ideas, then picked the top performers in simulation to refine the approach again and again. The iterative process paved the way to a unique, almost blob-like structure that dramatically boosted leaping power. Here, the AI didn’t just adjust small details—it offered a fresh perspective on the core physics that make a spring-loaded robot fly.
Of course, leaping higher isn’t much good if every landing ends in a tumble. Recognizing this, CSAIL used the same AI method to invent and test new foot shapes, striving for smoother landings. The results were just as impressive: an 84% reduction in falls compared to their starting design, opening the door to vastly more stable and reliable machines.
What’s especially thrilling is that this is only the beginning. With diffusion models, researchers can envision giving the AI a plain-language goal—like “design a robot that serves coffee” or “tighten a screw with a drill”—and having it invent both the structure and the control system from scratch. The team is now eyeing robots with more motors for better steering and experimenting with lighter materials, which could push performance even further.
The future of robotics might look less like a human engineer hunched over a workbench, and more like a vibrant collaboration between careful human insight and the boundless, sometimes surprising inventiveness of AI.
Read the original story at MIT-Nachrichten.
Diese Website verwendet Cookies.