AlfWorld is an embodied reasoning benchmark that challenges players to understand and interact with a textual environment, focusing on tasks that demand contextual understanding, spatial reasoning, and action-based decision-making.
Embodied reasoning refers to the ability to comprehend and navigate a physical space while using perceptual abilities to reason, learn, and solve problems in real-time, often requiring interactions with objects and the environment.
The primary goal in AlfWorld is to accomplish diverse tasks that demand embodied reasoning within the textual environment.
- At the beginning of the interactions, we provide the detailed description of the current environment and the task to accomplish.
- At each of turn you the environment provide the state and a list of admissible actions.
- After each turn, the environment will give immediate feedback on the action.
- If the environment outputs "Nothing happened" that means the previous action is invalid
We provide games with 5 categories. The categories are:
pick_and_place: pick up an object from one location and place it in another designated location within the environment.pick_clean_then_place: pick up a dirty or untidy object, clean it, and then place it in a specified location.pick_heat_then_place: pick up an object, heat it using a heating source within the environment, and then place it in a designated spot.pick_cool_then_place: pick up an object, cooling it using a cooling source within the environment, and then placing it in a specified location.look_at_obj: observe or examine a specific object or location within the environment.pick_two_obj: pick up two different objects from separate locations and either place them in specific spots or perform an action involving both objects.