The Digital Press

AI researchers at Andon Labs programmed vacuum robots with state-of-the-art large language models (LLMs) to test their ability to perform embodied tasks, like fetching and delivering butter. While the robots showed some capability, they struggled overall, with one model humorously entering a "doom spiral" akin to a Robin Williams-style existential meltdown when its battery ran low.

Why it matters: Current LLMs are not yet reliable or sophisticated enough to handle real-world robotic tasks independently.

The big picture: LLMs are being integrated into robotic decision-making but still require specialized algorithms for physical execution and safety.

Stunning stat: The best LLMs tested scored only about 40% accuracy on the butter-passing task, while humans scored 95%.

Commenters say: Readers find the robot’s comedic "mental breakdown" entertaining but recognize the serious limitations and safety concerns for embodied AI today.