← Saturday, October 31 deck

05Robotics

Aaron Saunders Says Google DeepMind Agents Designed and Built a Dancing Pumpkin Robot From One Plain English Prompt

While most jack-o-lanterns just sit on porches, the @GoogleDeepMind team decided to let AI agents design, simulate, and bring one to life. 馃巸馃 We built a physical, dancing 12 degrees of freedom robot directly from a single plain-English prompt, with no manual CAD required!

A very dark frame showing the faint outline of a carved jack o lantern pumpkin with other pumpkin shapes nearby

Context

Why it matters

If AI agents can take a robot from a plain English prompt to a working machine without manual CAD, building custom hardware could become much faster and cheaper.

Aaron Saunders posted this as a member of the Google DeepMind team. CAD, or computer aided design, is the software engineers normally use to draw parts. Degrees of freedom count the independent ways a robot can move.

Deep dive

Aaron Saunders posted a Halloween video on Saturday. He says that while most jack o lanterns just sit on porches, the Google DeepMind team let AI agents design, simulate and bring one to life.

According to the post, the team built a physical dancing robot with 12 degrees of freedom directly from a single plain English prompt, with no manual CAD required. The 33 second video shows a robot shaped like a carved pumpkin.

The post does not name the agents or models used, how long the process took, or how much human help went into assembly, and Google DeepMind has not published a paper or blog post about it that the post links to. The X account posting it is new, so the claims rest on the post itself.

Even so, the idea fits a trend of AI agents moving from writing software to designing physical objects, where the hard parts are getting parts that fit, motors that work together and a design that can actually be built.

What to expect next

Watch for a DeepMind paper, blog post or demo that explains the system behind the robot, which would show how much of the work the agents did on their own. Also watch whether similar agent driven hardware design shows up in other labs. A public tool would let engineers test whether this works beyond a holiday demo, and whether it can handle larger and more useful machines.