AI2-THOR

The Allen Institute for AI's interactive indoor 3D simulator, widely used for navigation and household tasks.

AI2-THOR is an open-source indoor simulation platform from the Allen Institute for AI (Ai2), with its paper published in 2017, built on the Unity engine. It offers near-photorealistic kitchen, bedroom, bathroom, and living-room scenes that an agent can walk through and interact with — opening a cabinet door, picking up and putting down objects, turning appliances on and off — with objects' open/closed and hot/cold states, among others, changing accordingly. It emphasizes high-level semantic interaction over fine-grained contact physics, so it's mostly used for visual navigation, embodied question answering, instruction following, and task planning. It includes several sub-frameworks: iTHOR (120 rooms with over 2,000 object types), RoboTHOR (apartment scenes with a matching physical test space), ManipulaTHOR (adds robot-arm manipulation), and ProcTHOR (procedurally generates large numbers of houses).

ExampleIn an iTHOR kitchen scene, an agent can execute a sequence like “walk to the fridge,” “open the fridge,” and “pick up the apple,” with the fridge door's open/closed state changing accordingly.

Also called

ai2thor, iTHOR

Related

ProcTHOR (Large-Scale Embodied AI Using Procedural Generation) · ALFRED · Habitat · Object-Goal Navigation · Embodied Question Answering · Allen Institute for AI

Sources

AI2-THOR 官网 (Chinese)
AI2-THOR: An Interactive 3D Environment for Visual AI (arXiv 1712.05474)

As of

2026-09