Ai2thor

Ai2-THOR is an open-source framework that provides interactive 3D environments for developing and evaluating embodied AI agents. It offers photo-realistic indoor scenes—such as kitchens, living rooms, bedrooms, and offices—where agents can navigate, interact with objects, and perform tasks through low-level actions (e.g., move, look, pick up, open, place). The platform integrates physics-based simulation, enabling realistic object dynamics, collisions, and state changes like opening containers, toggling appliances, or filling and emptying receptacles.

Developers can program agents in Python and use Ai2-THOR as a testbed for reinforcement learning, imitation learning, planning, and computer vision research. The tool supports generation of egocentric RGB, depth, instance segmentation, and optical flow data, making it suitable for training perception models and multi-modal policies. It also includes configurable scene variations, randomized object placements, and domain randomization features to improve robustness and generalization.