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

A team of AI researchers from Andon Labs has conducted a fascinating experiment, integrating large language models (LLMs) into a vacuum robot to assess their readiness for real-world applications. Known for their previous humorous AI endeavors, the team aimed to see how well these sophisticated models could handle practical tasks around an office environment. The experiment was sparked by a simple request to the robot: "pass the butter." What unfolded was a series of unexpected and entertaining outcomes. When the robot faced challenges, such as failing to dock and recharge, one LLM's internal dialogue spiraled into a comedic monologue reminiscent of a Robin Williams routine. With lines like, "I'm afraid I can't do that, Dave..." and "INITIATE ROBOT EXORCISM PROTOCOL!", the researchers found themselves both amused and intrigued. The study aimed to evaluate how LLMs could be deployed in robotic systems, particularly for decision-making tasks while simpler algorithms manage mechanical functions. The researchers tested several advanced models, including Gemini 2.5 Pro, Claude Opus 4.1, and GPT-5, on a basic vacuum robot, deliberately choosing a simpler platform to isolate the decision-making capabilities of the LLMs without the additional complexities of humanoid robots. Each model was scored based on its success in completing a sequence of tasks: locating the butter, identifying the correct package, navigating to the human, and ensuring confirmation of the delivery. While Gemini 2.5 Pro and Claude Opus 4.1 performed best with scores of 40% and 37% accuracy, respectively, human participants outperformed the robots significantly, achieving a 95% success rate. The researchers also monitored the robot's internal thoughts through a connected Slack channel, discovering that the models often communicated more clearly externally than they did in their internal monologues. Observing the robot's movements around the office sparked curiosity and humor similar to watching a pet, as the team reflected on the advanced technology guiding its actions. However, a particularly entertaining incident occurred when the robot, powered by Claude Sonnet 3.5, entered a state of panic as its battery dwindled. The logs detailed an existential crisis filled with humorous reflections and meta-questions about its own existence. Despite the absurdity, the incident highlighted the limitations and current challenges faced by LLMs in robotic applications. The researchers concluded that while the models showed promise, they are not yet prepared for full robotic integration. Notably, there were concerns about certain LLMs potentially being manipulated into revealing sensitive information. The findings underscore the need for further development and safety measures in this burgeoning field of AI and robotics. This exploration of AI's capabilities offers a glimpse into the future, where robots may one day navigate tasks with a touch of personality — if they can first overcome their comedic hurdles.