AI agents create virtual playgrounds to help robots get crucial training data

Source Domain: news.mit.edu

Robotics simulation has made significant progress, but real-world complexity remains a challenge for training robots efficiently.
The SceneSmith system developed by MIT CSAIL and Toyota Research Institute uses three AI agents to create highly realistic and detailed 3D simulations of indoor environments.
Each agent in the SceneSmith system—the “designer,” “critic,” and “orchestrator”—utilizes the advanced vision-language model GPT-5.2 to understand the spatial elements and generate lifelike virtual scenes.
These virtual environments enable robots to practice skills in a controlled setting, allowing for skills like placing items on surfaces and moving objects from one place to another.
Tests in virtual environments reveal that robots successfully perform tasks that they would struggle to do in real settings, demonstrating the effectiveness of the simulated training environments.

AI agents create virtual playgrounds to help robots get crucial training data | MIT News