ProcTHOR

ProcTHOR is a framework and dataset introduced in 2022 by Allen Institute for AI (AI2). It enables procedural generation of arbitrarily large datasets of diverse, interactive, customizable, and performant virtual environments for training and evaluating embodied AI agents. A sample of 10,000 generated environments is provided with the initial release. The generated environments cover a wide variety of room types (kitchens, bedrooms, living rooms, bathrooms, offices) with diverse furniture arrangements, object placements, and architectural styles. Environments are fully interactive with physics-enabled objects and support navigation, interaction, and manipulation tasks. ProcTHOR builds on AI2's experience with AI2-THOR and represents a major step toward scaling Embodied AI training through programmatic environment generation. With 450+ GitHub stars, it is actively used by the embodied AI research community for large-scale agent training.