Researchers at MIT's Computer Science and Artificial Intelligence Laboratory, working with the Toyota Research Institute, have developed an AI system called SceneSmith that generates detailed virtual homes to help robots learn everyday tasks before attempting them in the real world. This matters because household chores that humans perform without thinking, such as putting away a mug, require robots to navigate clutter, avoid obstacles and identify the correct location, making such tasks surprisingly difficult to master through real-world training alone.

SceneSmith uses AI agents to build 3D indoor environments from simple text prompts, creating varied rooms and scenarios in which robots can rehearse actions and identify flawed strategies safely. By exposing weaknesses in a robot's planning within these simulated spaces, the system aims to reduce the time and hands-on supervision currently needed to prepare robots for real homes and workplaces.