| Definition | CookingZoo is a customizable cooking simulation environment for testing AI in culinary tasks. |
| No of environments | 14 |
CookingZoo is an innovative cooking simulation environment designed to enhance the generalization capabilities of AI agents through diverse cooking scenarios. Built on visual assets from previous works, it supports multi-agent interactions using the PettingZoo and gymnasium libraries, allowing both AI-controlled and human-controlled agents to participate in cooking tasks.
The environment boasts a rich library of existing recipes and offers the flexibility to create custom recipes and levels. It includes different observation spaces, such as feature vectors for numerical analysis and symbolic representations for reasoning about the objects in the environment. With the capability to handle multiple agents—ranging from cooperating teams to competitive scenarios—CookingZoo facilitates a wide variety of learning and interaction dynamics.
Key features of CookingZoo include customizable level designs, variable agent counts, and various observation space configurations. The environment can be easily extended by adding new objects and recipes, making it adaptable to various research needs. Additionally, it adheres to the updated gym conventions, ensuring compatibility with major reinforcement learning frameworks like RLLib and StableBaselines3.
| Example observations | recipes completion state |
| Example actions | walk (up, down, left, right), interact, execute |
| Example reward | complete recipe |