AI struggles with Overcooked: Even the best model manages just 26 percent

Researchers have turned the fun couch co-op multiplayer game Overcooked into an AI testing environment called DuoCook, where AI models must demonstrate that they can understand visual information, plan tasks, and coordinate multiple steps.

In the chaotic co-op game Overcooked, 2 to 4 players typically work together in a virtual restaurant to chop ingredients, cook meals, wash dishes, and deliver finished orders on time before customers grow impatient and cancel. Success depends on effective teamwork, with tasks divided among players to work as efficiently and quickly as possible. The result is glorious chaos and great co-op fun.

Instead of human players, researchers put various AI models to the test to see how well artificial intelligence could coordinate and complete these tasks. They tested four vision-language models, AI systems capable of processing both images and text. A single model controls both characters simultaneously.

ⓘ Ziyi Jiao, Heng Zhang et al

ⓘ Ziyi Jiao, Heng Zhang et al

The results are sobering: the strongest model in the test, GPT-6-Astra, completed an average of 26.2 percent of orders. The other three models managed just 8.8 to 11.2 percent. Maps featuring dynamic obstacles and shared equipment proved particularly challenging, as did kitchen areas where ingredients had to be passed between sections.

Adding extra memory didn't reliably help either. The researchers introduced a shared notebook where the agents could record their observations. Claude improved its average score in this experimental setup, while GPT and Gemini actually performed worse. At the same time, API costs increased for all four models.

The narrowly defined task makes the results interesting, but also limits their scope. Even with additional aids such as a grid for determining positions, a legend of kitchen equipment, structured status information, and a list of permitted actions, coordination remains difficult.

This matters in real-world applications because the same challenges can arise with autonomous AI agents outside of games: dividing up tasks, managing resources, tracking intermediate goals, and responding to changing conditions.

However, the study should be treated with some caution. It is a preprint, meaning it has not necessarily undergone the standard peer-review process of a journal or conference. The full study can be read via the link below.

A C64 marked my entry into the world of PCs. I spent my student internship in the repair department of a computer shop and at the end of the day I was allowed to assemble my own 486 PC from “workshop remnants”. As a result of this, I later studied computer science at the Humboldt University in Berlin, with psychology also being added to my studies. After my first job as a research assistant at the university, I went to London for a year and worked for Sega in computer game translation quality assurance. This included working on games such as Sonic & All-Stars Racing Transformed and Company of Heroes. I have been writing for Notebookcheck since 2017.