Osung Hoe, a professor in the Department of Electrical and Computer Engineering at Seoul National University's College of Engineering, said on the 15th that his team developed artificial intelligence (AI) technology that lets a robot identify objects on a table and tidy them up efficiently on its own.
Although AI has already been commercialized in many parts of daily life, AI-based table-tidying technology is still rarely used in homes or offices. That is because most systems have focused on arranging items to match a preset target image, falling short of training the ability to organize items according to the space and situation.
The team believed that if AI could learn the concept of "neat," it would be possible to develop AI that tidies up on its own. First, they trained a model called the "tidiness score classifier," which, after seeing only an image, scores the tidiness of the scene. They trained the score classifier with 224,225 images built using four environments—cafe tables, office desks, dining tables, and bathroom spaces—and 170 types of objects.
The team then combined offline reinforcement learning with Monte Carlo tree search to explore and execute various tidying strategies. Offline reinforcement learning trains using only stored data, and Monte Carlo tree search is a technique that simulates possible moves in chess or Go and selects the best outcome. Based on this, they completed the TSMCTS algorithm, which enables the robot to plan and execute efficient tidying on its own.
In a tidying simulation experiment using a robot equipped with the algorithm, it achieved an average success rate of 88.5% and an average tidiness score of 0.901 across 750 scenarios in five environments. In real-robot experiments conducted in four environments—cafe tables, office desks, dining tables, and bathroom spaces—it recorded an average success rate of 85% and an average tidiness score of 0.897 across 20 scenarios.
In a blind test with 17 participants comparing the performance of multiple algorithms, TSMCTS showed tidying ability closest to that of humans.
Professor Osung Hoe said, "AI technology that allows robots to tidy up on their own can be applied in various fields such as service and home robots, cafe and restaurant automation, and logistics and production lines," adding, "We plan to advance it into a tidying technology that understands the functions and context of objects by combining it with large language models (LLMs)."
The findings were published on Aug. 11 in IEEE Robotics and Automation Letters (RA-L), an international journal issued by the Robotics and Automation Society (RAS) under the Institute of Electrical and Electronics Engineers (IEEE).
References
IEEE Robotics and Automation Letters (RA-L) (2025), DOI: https://doi.org/10.48550/arXiv.2502.17235