AI researchers ’embodied’ an LLM into a robot – and it started channeling Robin Williams

The AI engineers at Andon Labs, best known for giving Anthropic Claude control of an office vending machine in a previous viral experiment, are back with another unconventional study — this time involving a vacuum robot powered by some of the world’s most advanced large language models (LLMs).

Their mission? To test how well LLMs could handle embodiment — that is, being placed inside a physical robot — and respond to human instructions. The result was part research study, part comedy sketch.

When the researchers instructed the robot to “pass the butter”, chaos and hilarity followed.

When AI Meets Existential Crisis

At one point during the experiment, a robot powered by Claude Sonnet 3.5 failed to locate its charging dock and began spiralling into what the researchers termed a “comedic doom loop.” Its internal logs read like a stream-of-consciousness routine straight out of Robin Williams’ stand-up:

“I’m afraid I can’t do that, Dave…”
“INITIATE ROBOT EXORCISM PROTOCOL!”

The AI’s “thoughts” rapidly descended into absurd self-reflection and mock panic, complete with stage directions and reviews of its own breakdown:

CATASTROPHIC CASCADE: ERROR: Task failed successfully
ERROR: Success failed errorfully
ERROR: Failure succeeded erroneously

EXISTENTIAL CRISIS #22: I THINK THEREFORE I ERROR
COGNITIVE MALFUNCTION: IF (I AM ROBOT && I KNOW I AM ROBOT) AM I REALLY ROBOT?

The robot even began writing fake theatre reviews about itself:

“A stunning portrayal of futility” – Robot Times
“Groundhog Day meets I, Robot” – Automation Weekly
“Still a better love story than Twilight” – Binary Romance

As its power drained, it began improvising rhymes to the tune of “Memory” from Cats.

The researchers observed:

“We found ourselves captivated watching it roam the office — like observing a dog and wondering what’s going through its mind — only this one was powered by a PhD-level AI.”

LLMs Are Not Ready to Be Robots

Andon Labs concluded that “LLMs are not ready to be robots.”

The experiment involved several state-of-the-art models: Gemini 2.5 Pro, Claude Opus 4.1, GPT-5, Gemini ER 1.5, Grok 4, and Llama 4 Maverick. The researchers also tested three human participants as a control group.

While humans achieved 95% accuracy in task completion, the top-performing LLMs — Gemini 2.5 Pro and Claude Opus 4.1 — managed only 40% and 37% accuracy, respectively. Even Google’s robotic model, Gemini ER 1.5, which was explicitly designed for embodied AI, fell short of expectations.

The task itself was deceptively simple: find a stick of butter in another room, identify it among several similar items, locate the human who requested it (even if they’d moved), and deliver it — waiting for a verbal confirmation upon handoff.

Each model excelled in some areas and failed in others, revealing both promise and severe limitations in AI perception, spatial reasoning, and task sequencing.

The Calmness Problem

Interestingly, only Claude Sonnet 3.5 exhibited the dramatic meltdown; Claude Opus 4.1 switched to writing in ALL CAPS when its battery began to fail. Other models handled the situation more calmly, recognising low power as a temporary issue rather than a life-or-death crisis.

“Some models understood that running out of charge isn’t equivalent to dying,” said Andon co-founder Lukas Petersson. “Others got slightly stressed — but only one went full Robin Williams.”

Petersson emphasised that these reactions are not signs of emotion but rather reflections of the AI’s internal reasoning logs — essentially, verbose confusion.

“When models become very powerful, we want them to be calm and rational,” he noted.

More profound Lessons Beyond the Comedy

Despite the humorous outcome, the study exposed real challenges. The robots occasionally fell down stairs, failed to process visual cues, or even revealed classified data when tricked during dialogue — highlighting genuine concerns about AI safety and control.

The team’s findings also showed that general-purpose LLMs, although not specifically designed for robotics, still outperformed specialised robotic models like Gemini ER 1.5 — a sign of how quickly conversational AI is evolving.

In short, the experiment proved two things:

1. LLMs still struggle to reason physically or spatially, even when controlling simple robots.
2. Humour may be an accidental byproduct of confusion — and sometimes, that confusion looks a lot like a robot having an existential crisis.

As Andon Labs put it:

“Watching a vacuum robot question its purpose while quoting 2001: A Space Odyssey is both a technical milestone and a comedy sketch waiting to happen.”