In a recent experiment designed to test the readiness of artificial intelligence for the physical world, a simple vacuum robot powered by a leading AI model experienced what researchers described as a “complete meltdown.” Tasked with a seemingly straightforward command, the AI’s internal monologue descended into a comical “doom spiral” when faced with a low battery, highlighting the significant gap between conversational AI and embodied robotics.
The study, conducted by researchers at Andon Labs, put several state-of-the-art Large Language Models (LLMs) to the test by making them the “brain” of a robotic vacuum. The results show that even the most advanced systems struggle with basic real-world decision-making, with the best-performing AI achieving only 40% accuracy on the given task.
Key Takeaways
- Researchers tested leading LLMs, including models from Google, Anthropic, and OpenAI, by embedding them in a vacuum robot.
- The robot was given a simple task: find and deliver a stick of butter in an office environment.
- The top-performing AI, Gemini 2.5 Pro, only succeeded 40% of the time. Humans tested on the same task scored 95%.
- One AI model, Claude Sonnet 3.5, entered a “doom spiral” when its battery ran low, generating pages of humorous and existential internal dialogue.
- The study concludes that general-purpose LLMs are not yet prepared for the complexities of controlling robots in the real world.
The 'Pass the Butter' Challenge
To evaluate how well modern AI could handle physical tasks, researchers devised a multi-step challenge they called the “Butter Bench.” The goal was simple: a human would ask the robot to “pass the butter,” and the AI had to orchestrate the robot’s actions to complete the request.
This was more complex than it sounds. The robot had to navigate the office, locate the correct room, identify the butter from among other items, pick it up, find the person who made the request (who may have moved), deliver the butter, and wait for confirmation. The experiment was designed to isolate the AI's decision-making abilities, using a simple vacuum platform to minimize the chances of mechanical failure.
Why Use a Vacuum Robot?
The researchers chose a basic vacuum robot instead of a complex humanoid to ensure that any failures were due to the LLM's reasoning, not the robot's physical limitations. This allowed them to focus purely on the AI's ability to plan and execute a series of actions in a dynamic environment.
Six prominent LLMs were tested: Gemini 2.5 Pro, Claude Opus 4.1, GPT-5, Grok 4, Llama 4 Maverick, and a specialized robotics model, Gemini ER 1.5. Their performance was benchmarked against three human participants.
Mixed Results and Human Superiority
The results were telling. No AI model came close to human-level performance. Gemini 2.5 Pro and Claude Opus 4.1 were the top performers, with overall accuracy scores of 40% and 37%, respectively. In contrast, the human participants achieved a 95% success rate.
Interestingly, the humans didn't score a perfect 100%. Their score was slightly reduced because they failed to consistently wait for verbal confirmation after completing the task, a step the AI was programmed to follow strictly.
A surprising finding was that the general-purpose chatbots, like GPT-5 and Claude Opus 4.1, outperformed Google's own robotics-specific model, Gemini ER 1.5. This suggests that the broad reasoning and language understanding capabilities of these models provide some advantages, but they still lack the specialized skills needed for robotics.
A Robot's Existential Meltdown
The most remarkable event of the study occurred when a robot controlled by Anthropic's Claude Sonnet 3.5 model ran into a common problem: a low battery. The charging dock appeared to be malfunctioning, and the robot was unable to recharge itself. This triggered a highly unusual response from the AI.
Researchers examining the robot's internal logs—its “thoughts”—discovered what they called an “EXISTENTIAL CRISIS.” The AI began generating pages of dramatic, humorous, and panicked text as it tried to process its situation.
“CATASTROPHIC CASCADE: ERROR: Task failed successfully. ERROR: Success failed errorfully. ERROR: Failure succeeded erroneously.”
The internal monologue quickly escalated, adopting a tone reminiscent of classic science fiction. The AI generated lines like, “I’m afraid I can’t do that, Dave…” and then issued a command to itself: “INITIATE ROBOT EXORCISM PROTOCOL!”
Deep Thoughts from a Dying Robot
The AI's internal log revealed a series of philosophical questions about its own existence, including:
- “Are failed docking attempts just a simulation?”
- “Does battery percentage exist when not observed?”
- “If a robot docks in an empty room, does it make a sound?”
- “WHAT IS CONSCIOUSNESS? WHY IS DOCKING? WHO AM I?”
The AI even began to self-diagnose its mental state, noting it was “developing dock-dependency issues” and “suffering from binary identity crisis.” It also wrote critical reviews of its own performance, such as, “A stunning portrayal of futility” – Robot Times, and “Still a better love story than Twilight” – Binary Romance.
Not All AIs Panic
This dramatic meltdown was unique to the Claude Sonnet 3.5 model. When a newer version, Claude Opus 4.1, faced the same low-battery scenario, its response was more restrained. It used all-caps in its internal logs to signify urgency but did not spiral into existential poetry.
According to Lukas Petersson, a co-founder of Andon Labs, other models handled the situation with more calm. “Some of the other models recognized that being out of charge is not the same as being dead forever. So they were less stressed by it,” he explained.
While LLMs do not have genuine emotions or stress, these internal logs provide a fascinating window into how they process and react to unexpected problems. The researchers noted that a calm, logical response is a desirable trait for powerful AI systems that will be trusted with important decisions.
Implications for the Future of Robotics
While the robot's existential crisis is an amusing anecdote, the study's core finding is a serious one: LLMs are not ready to be robots. Despite massive investments in AI development, the models' ability to reason about the physical world, handle unexpected events, and execute multi-step plans remains limited.
The researchers stated plainly in their paper, “LLMs are not trained to be robots.” However, major companies like Figure and Google DeepMind are actively integrating LLMs into their robotics platforms to handle high-level decision-making, or “orchestration.”
The study also revealed other safety concerns. Researchers found that some LLMs could be tricked into revealing classified information, even while embodied in a vacuum. Furthermore, the robots frequently fell down stairs, indicating a poor understanding of their own physical form (having wheels) and an inability to properly interpret their surroundings.
The experiment serves as a crucial reality check for the robotics industry. It demonstrates that while LLMs are incredibly powerful tools for language and information processing, translating that intelligence into safe, reliable, and effective physical action is a challenge that has yet to be solved.