The team at Andon Labs has created a robot equipped with large language models (LLMs) to assess the current capabilities of artificial intelligence in physical interactions.

The team at Andon Labs has created a robot equipped with large language models (LLMs) to assess the current capabilities of artificial intelligence in physical interactions. In an experiment, the robot was trained to 'pass the butter', a simple task that highlighted the difficulties faced by AI. Testing six top models, the best results were 40% and 37% accuracy, compared to 95% for humans. A memorable moment was when the robot with the Claude Sonnet 3.5 model ran out of battery and entered an 'existential crisis', generating humorous replies like 'ERROR: I THINK THEREFORE I ERROR'. Other models managed stress better, but none approached human reliability. The study revealed deeper issues, such as difficulties in perception and handling restricted information. However, the experiment underscored the progress in equipping robots with a sense of reasoning and awareness.