AI2THOR

AI2-THOR is an open-source interactive 3D environment designed to support research in embodied artificial intelligence. It provides photorealistic household scenes implemented in the Unity game engine, where an autonomous agent can perceive, navigate, and interact with objects in a controllable, reproducible setting.

The platform exposes a Python API that allows programmatic control of a virtual agent. Researchers can issue actions such as moving and turning, picking up and placing objects, opening or closing containers, and interacting with appliances. The environment supplies sensor data including RGB images, depth maps, and object-level segmentation or semantic labels to support perception and reasoning tasks. Scenes can include diverse rooms and layouts with various movable and manipulable objects, physics-based interactions, and time-variant elements.

AI2-THOR is widely used for studies in embodied AI tasks, such as navigation to goals, object manipulation and assembly, and visual question answering within a grounded, interactive context. It supports experiments requiring sequential decision making, multi-step plans, and realistic affordances for objects (for example, opening a refrigerator or grasping items). Researchers often integrate AI2-THOR with reinforcement learning frameworks and evaluation protocols to benchmark agent performance across tasks and scenes.

The project emphasizes accessibility and reproducibility, offering a modular architecture, documentation, and examples to facilitate experimentation. It serves as a common platform for comparing embodied AI approaches and for creating new tasks that require coordinated perception, planning, and action in realistic 3D environments.