ai2thor

Unlike 2D game-based RL environments, AI2-THOR provides high-fidelity photorealistic households where agents must visually navigate rooms, find and manipulate objects, understand spatial relationships, and complete multi-step household taskschallenging capabilities that are prerequisites for real-world robot deployment.

The simulator provides over 200 room-scale scenes across kitchen, living room, bedroom, and bathroom environments, with hundreds of interactive objects that support physically realistic manipulationobjects can be picked up, placed, opened, filled with liquid, and heated, modeling the affordances that domestic robots must understand.

AI2-THOR's action space supports both navigation (move forward, rotate, look up/down) and object interaction, enabling research on embodied agents that must coordinate visual perception, spatial reasoning, and physical manipulation in service of high-level task goals.

Robotics and embodied AI researchers use AI2-THOR as a benchmark environment for vision-and-language navigation, object goal navigation, and robotic instruction followingtasks where an agent must interpret a natural language instruction and execute a multi-step plan in a realistic visual environment.

The simulator is used in annual competitions (like ALFRED and TEACh) that have advanced the field's understanding of what capabilities embodied agents currently have and where they fall short.

Researchers developing foundation models for robotics use AI2-THOR environments to collect demonstrations and evaluate zero-shot generalization without requiring physical robot hardware.