virtualhome

Why we included this project

VirtualHome is a household simulator you drive from Python, aimed at people training embodied agents or reinforcement-learning policies that need to act in realistic homes. You write an activity as a short program of high-level instructions, and the simulator renders it with humanoid avatars that pick up objects, open appliances, and move between rooms, while streaming ground truth like time-stamped actions, semantic segmentation, and depth. That makes it useful for generating training data and for running controlled, reproducible evaluations. It also supports multiple agents at once and ships OpenAI Gym-style environments, so it fits into existing RL training loops without much glue code. If you work on activity understanding, program prediction, or human-AI collaboration, the bundled activity knowledge base is a practical place to start.