Princeton researchers help a bot tidy up using large language model

At home, however, the place where an item is kept is dependent on various factors ranging from personal taste to cultural backgrounds. For instance, some prefer to put their shirts in a drawer, while others would put them on shelves or hang them up.

While a robot deployed at home can be specifically programmed to carry out these tasks, with the advent of large language models, this step can be conveniently avoided. The researchers at Princeton attempted to train the robot on personal preferences using limited examples of prior interactions and were largely successful.

The team put user preferences through an LLM to generate placement rules for objects using the robot's three primitive functions of picking, placing, and tossing. Based on the rules, the robot was able to separate dark and light clothing into different receptacles and even sort tools, wooden blocks, and fruits into separate shelves.

In a messy room scenario, the robot was successful in sorting plastic utensils from paper bags and cans and even placing them in trashcans or recycling bins as desired. The robotic arm, which is capable of opening and closing drawers, was also successful in placing items inside them, whether they were round or rectangular. It was also able to differentiate storage boxes on the basis of colors and sort items in accordance with the instructions provided.

In real-world scenarios, the robot successfully placed objects with 85 percent accuracy.

The researchers have placed the code for LLM evaluation and real robot implementation on the software repository GitHub. A research paper detailing the work was also put on the pre-print server arxiv. The findings of the experiments have not been peer-reviewed yet.