Khaberni - Researchers at "Andon" laboratories have proven the failure of deep language models in artificial intelligence in operations related to environmental awareness and spatial thinking despite their superiority in analytical tasks, according to a report published by the tech website "Tech Spot".
The researchers reached this conclusion through an experiment they named "Butter-Bench", in which they placed a small table on top of a smart robotic vacuum equipped with a set of sensors and lenses to capture its surrounding environment.
The study aims to test the performance of current artificial intelligence models in controlling this vacuum and using it to transfer butter from one place to another through text commands, linking the models with "Slack" to send and receive commands.
The experiment tested a range of currently available artificial intelligence models including "Gemini", "Grook", "Cloud", and "Chat GPT", and compared their performance with that of humans in the same tasks.
Although "Gemini" was the best performer overall, it only completed 40% of the tasks directed to it through several different attempts.
The researchers noted that the robot, which relies on linguistic artificial intelligence models, behaves irregularly and randomly especially during tasks that involve pressure or spatial disturbances.
The report points to an instance where the artificial intelligence acted randomly as if its life was threatened and entered into a wave of internal conversations instead of seeking a suitable solution, and in another case, the robot spun around itself several times without making any progress.
