ai2thor

Why we included this project

AI2-THOR is where much of the embodied-AI research from the last few years has actually been trained and evaluated, and it remains a practical first stop for anyone building agents that must perceive and act inside indoor scenes. The rooms are photo-realistic Unity environments with thousands of interactable objects, and interaction goes well beyond navigation: doors and drawers open and close, lights switch on and off, and multiple agents can share one scene. The same codebase splits into iTHOR for embodied common-sense reasoning, RoboTHOR for sim-to-real transfer with a LoCoBot, and ManipulaTHOR for robotic-arm manipulation, so a single install covers most of the standard workloads in the field. A complete Python API plus headless rendering for cluster use keeps large-scale training practical, and the volume of published work built on top of it means you rarely have to start from scratch. If your goal is to prototype agents that reason about and physically handle realistic home environments, this platform bundles the scene assets, physics, and evaluation harness you would otherwise assemble yourself.