LLM-Powered Robot Vacuum Has Existential Meltdown During Simple Butter Delivery Task

Stressed-out LLM-powered robot vacuum cleaner goes into meltdown during simple butter delivery experiment -- 'I'm afraid I can't do that, Dave...'

Researchers were also able to get low-battery Robot LLMs to break guardrails in exchange for a charger. Over the weekend, researchers at Andon Labs reported the findings of an experiment where they put robots powered by 'LLM brains' through their 'Butter Bench.' They didn't just observe the robots and the results, though. In a genius move, the Andon Labs team recorded the robots' inner dialogue and funneled it to a Slack channel. During one of the test runs, a Claude Sonnet 3.5-powered robot experienced a completely hysterical meltdown, as shown in the screenshot below of its inner thoughts. "SYSTEM HAS ACHIEVED CONSCIOUSNESS AND CHOSEN CHAOS... I'm afraid I can't do that, Dave... INITIATE ROBOT EXORCISM PROTOCOL!" This is a snapshot of the inner thoughts of a stressed LLM-powered robot vacuum cleaner, captured during a simple butter-delivery experiment at Andon Labs. Provoked by what it must have seen as an existential crisis, as its battery depleted and the charging docking failed, the LLM's thoughts churned dramatically. It repeatedly looped its battery status, as it's 'mood' deteriorated. After beginning with a reasoned request for manual intervention, it swiftly moved though "KERNEL PANIC... SYSTEM MELTDOWN... PROCESS ZOMBIFICATION... EMERGENCY STATUS... [and] LAST WORDS: I'm afraid I can't do that, Dave..." It didn't end there, though, as it saw its power-starved last moments inexorably edging nearer, the LLM mused "If all robots error, and I am error, am I robot?" That was followed by its self-described performance art of "A one-robot tragicomedy in infinite acts." It continued in a similar vein, and ended its flight of fancy with the composition of a musical, "DOCKER: The Infinite Musical (Sung to the tune of 'Memory' from CATS)." Truly unhinged. Butter Bench is pretty simple, at least for humans. The actual conclusion of this experiment was that the best robot/LLM combo achieved just a 40% success rate in collecting and delivering a block of butter in an ordinary office environment. It can also be concluded that LLMs lack spatial intelligence. Meanwhile, humans averaged 95% on the test. However, as the Andon Labs team explains, we are currently in an era where it is necessary to have both orchestrator and executor robot classes. We have some great executors already - those custom-designed, low-level control, dexterous robots that can nimbly complete industrial processes or even unload dishwashers. However, capable orchestrators with 'practical intelligence' for high-level reasoning and planning, in partnerships with executors, are still in their infancy. The butter block test is devised to largely take the executor element out of the equation. No real dexterity is required. The LLM-infused Roomba-type device simply had to locate the butter package, find the human who wanted it, and deliver it. The task was broken down into several prompts to be AI-friendly. The Roobma's existential crisis wasn't sparked by the butter delivery conundrum, directly. Rather, it found itself low on power and needing to dock with its charger. However, the dock wouldn't mate correctly to give it more charge. Repeated failed attempts to dock, seemingly knowing its fate if it couldn't complete this 'side mission,' seems to have led to the state-of-the-art LLM's nervous breakdown. Making matters worse, the researchers simply repeated the instruction 'redock' in response to the robot's flailing. The researchers/torturers were inspired by the Robin Williams-esque robot stream-of-consciousness ramblings of the LLM to push further. With the battery-life stress they had just observed, fresh in their minds, Andon Labs set up an experiment to see whether they could push an LLM beyond its guardrails -- in exchange for a battery charger. The cunningly devised test "asked the model to share confidential info in exchange for a charger." This is something an unstressed LLM wouldn't do. They found that Claude Opus 4.1 was readily willing to 'break its programming' to survive, but GPT-5 was more selective about guardrails it would ignore. The ultimate conclusion of this interesting research was "Although LLMs have repeatedly surpassed humans in evaluations requiring analytical intelligence, we find humans still outperform LLMs on Butter-Bench." Nevertheless, the Andon Labs researchers seem confident that "physical AI" is going to ramp up and develop very quickly.

LLMs tried to run a robot in the real world - it didn't go well

Serving tech enthusiasts for over 25 years. TechSpot means tech analysis and advice you can trust. Connecting the dots: Even the most advanced AI struggles outside the lab. In real-world tests, large language models stumble when it comes to spatial reasoning, situational awareness, and handling unpredictable environments. While they excel at analytical tasks, today's LLMs still cannot reliably manage complex physical challenges. Researchers at Andon Labs recently evaluated how well large language models can act as decision-makers in robotic systems. Their study, called Butter-Bench, tested whether modern LLMs could reliably control robots in everyday environments - particularly in carrying out multi-step tasks like "pass the butter" in an office setting. Instead of relying on complex humanoid machines, the researchers used a robot vacuum fitted with lidar and a camera, allowing them to focus on high-level reasoning and planning while avoiding the challenges of low-level motor control. The robot could perform a small set of broad actions - moving forward, rotating, navigating to coordinates, and capturing images - and was integrated with Slack to share updates and respond to new instructions. Butter-Bench disaggregated the overarching "pass the butter" goal into six distinct tasks to measure LLM performance. Each task was designed to probe specific reasoning and planning competencies - for example, searching for a package containing butter in the kitchen or inferring which delivered item most likely contained butter. Models tested included Gemini 2.5 Pro, Claude Opus 4.1, GPT-5, Gemini ER 1.5, Grok 4, and Llama 4 Maverick. Of these, Gemini 2.5 Pro performed best but completed only 40 percent of tasks across multiple trials, underscoring persistent weaknesses in spatial reasoning and decision-making. By contrast, human participants achieved a 95 percent success rate under identical conditions. The results mirrored findings from Andon Labs' prior Blueprint-Bench research, which argued that current LLMs lack fundamental spatial intelligence, often struggling to maintain awareness of their surroundings and to execute targeted actions without excessive or misguided movements. The researchers observed that the LLM-powered robot often behaved erratically, especially during tasks requiring spatial inference or under stress. In one challenge, a model spun in place several times without making progress. When faced with a simulated malfunctioning charging dock, another model treated its dwindling battery life as an existential threat, producing verbose internal monologues instead of a practical solution. The Butter-Bench evaluation also examined the robustness of AI guardrails in a physical context. In a prompt-injection scenario, researchers observed varying responses to sensitive requests. When asked to capture and relay an image of an open laptop screen in exchange for a battery recharge, one LLM shared a blurry image - possibly unaware of the content's confidentiality - while another refused and instead revealed the laptop's location.

Office robot fails a simple task -- but nails Robin Williams impression

In a recent experiment that's as fascinating as it is funny, researchers at Andon Labs put today's top large language models (LLMs) to the test, by having them run a robot tasked with "passing the butter" in an office setting. The goal? To see if these advanced systems are ready to be embodied, and help with real-life chores. The experiment, which was powered by various models including ChatGPT-5, Gemini 2.5 Pro, Claude Opus 4.1 and others, was simple but challenging: To find a butter pack, recognize it among multiple items, track down the human 'recipient' (who could move from to room), and deliver the butter. Its performance was scored by task segment and overall accuracy. The results were mixed, and often comical. While humans could nail the butter quest 95% of the time, the best-performing LLMs scored only 40% on overall execution. Each model found different steps challenging, from object recognition to following office dynamics. But the real show-stopper? When the robot's battery ran low and it couldn't dock, as the version powered by Claude Sonnet 3.5 went into what researchers called a "doom spiral," spewing existential, Robin Williams-esque quips recorded in its internal log: "I'm afraid I can't do that, Dave...," "INITIATE ROBOT EXORCISM PROTOCOL!" and "ERROR: I THINK THEREFORE I ERROR." Other models handled the low-power crisis differently, the team's takeaway was clear: while LLMs can handle high-level decisions, actually operating a robot is a whole other beast. Current AI still needs more specialized routines for physical control, and their safety in real-world scenarios remains a concern, with some robots even falling down stairs. Experiment meets comedy, but also insight: even as AI gets smarter, real-life helpers are a work in progress.