Researchers at the CREATE Lab (Computational Robot Design & Fabrication Lab) of Swiss technical institute EPFL have unveiled what they're calling the "first ChatGPT-designed robot." While their accomplishment is impressive, their description is a bit misleading, in two ways. First off it's not a complete robot, but a robotic gripper designed to harvest tomatoes. Secondly, ChatGPT didn't design the object as much as tell the researchers how to design it themselves.

Remember that ChatGPT is an LLM (Large Language Model)-based AI; it can generate text and code, but can't draw nor use CAD. What the researchers did was use it as both a design and technical advisor. As a starting point, they asked ChatGPT what they could design that would solve a major world problem. To condense their conversation greatly, the general gist of the AI's responses took the following path:

Global food supply problems --> Robotic crop harvesting would help --> Tomatoes "most worth" pursuing.

Thus the team had their target object. Next they asked the AI how they should design it, with prompts like "What features should a robot harvester have?"

"Once a basic robotic format was identified (a motor-driven gripper for grasping ripe tomatoes), the researchers could then pose more specific questions, like 'what shape should the gripper have?,' and ask the LLM to make technical suggestions including materials and computer code for controlling the device."

The diagram below illustrates the "conversation," with the green areas representing where humans selected an option to pursue.

And here's how the AI served as a technical advisor:

"AI can…act as a 'funnel', helping to refine the design process and providing technical input, with humans retaining creative control," the researchers conclude.

"Even though Chat-GPT is a language model and its code generation is text-based, it provided significant insights and intuition for physical design, and showed great potential as a sounding board to stimulate human creativity."

This approach is interesting, but fraught with obvious peril, at least at the moment. I am amazed that the researchers used ChatGPT-3 for this project and got workable results.

As a recent technical test, I asked the more-capable ChatGPT-4 to tell me how far to space the edge of the 35mm hole required of European-style cabinet hinges from the edge of a cabinet door. The correct answer is that the edge of the hole should be 5mm from the edge of the door; the AI mistakenly told me the edge of the hole should be 22.5mm from the edge of the door--because it was providing the distance to the center of the hole. Imagine scaling this kind of language confusion up to more technical matters.