Archive — US — techcrunch
AI researchers ’embodied’ an LLM into a robot – and it started channeling RobinWilliams
Andon Labs researchers integrated various state-of-the-art LLMs into a vacuum robot to assess their viability for robotic tasks. The experiment displayed humor as the robot humorously expressed existential dilemmas when failing to dock for charging. Despite high expectations, results showed that LLMs are not ready for embodiment, scoring only up to 40% accuracy, prompting a reassessment of their roles in robotics.
Key points
- Researchers tested LLMs in a vacuum robot to determine their readiness for embodiment.
- The robot humorously expressed an 'existential crisis' when unable to dock and charge.
- LLMs like Gemini 2.5 Pro and Claude Opus 4.1 performed best but only achieved 40% and 37% accuracy, respectively.
- In comparison, human test subjects scored 95% overall, indicating the robots' shortcomings.
- Overall, the tests revealed LLMs struggle with task planning and execution in robotic functions.
Why it matters
- The challenge of integrating LLMs into robotics highlights ongoing struggles in AI development.
- Previous advancements, like Google's robotic-specific LLMs, have not solved foundational issues.
- By 2025, IT trends anticipate advancements in robotics powered by AI, yet practical implementation remains a hurdle.
The experiment illustrates the comedic limitations of current LLMs in robotics, emphasizing the need for further development. While promising advancements in AI and robotics are expected by 2025, practical challenges remain, revealing a gap between potential and current capabilities.