MIT, Toyota Build SceneSmith to Generate 1,300 Virtual Homes for Robot Training
Updated
Updated · Fox News · Jul 23
MIT, Toyota Build SceneSmith to Generate 1,300 Virtual Homes for Robot Training
3 articles · Updated · Fox News · Jul 23
Summary
SceneSmith lets robots practice household tasks in AI-generated 3D rooms before entering real homes or workplaces, aiming to cut the cost, supervision and breakage risks of physical training.
Three GPT-5.2-powered agents build each scene from text prompts, then add usable physics so robots can open cabinets, move objects and test action policies in spaces that behave like real rooms.
MIT and Toyota researchers generated more than 1,300 scenes, with some containing up to six times more items than earlier methods; 96% of objects stayed stable and fewer than 2% of object pairs collided.
In 100 evaluation scenes across four manipulation tasks, an AI evaluator matched human labels 99.7% of the time, while 205 people gave SceneSmith a 92% win rate for realism and 91% for prompt fidelity.
The system still needs several hours to build one detailed room and has limited support for deformable objects, underscoring that virtual training can supplement but not replace real-world safety testing.
As robots train in perfect virtual worlds, how do we prevent them from failing in our unpredictable physical reality?
What is the biggest safety hurdle preventing humanoid robots from finally entering our homes and helping with chores?
How will AI tutors like Sally reshape student learning and the fundamental role of human teachers in education?
SceneSmith Sets New Benchmark: 3–6x Denser, 96% Stable Virtual Worlds for Realistic Robotics Training
Overview
SceneSmith, developed by MIT and the Toyota Research Institute, is a new AI-powered system that made its public debut at the International Conference on Machine Learning. It stands out by generating highly realistic environments and entirely new 3D objects with accurate physical properties like weight and friction. This level of detail is crucial for robots to learn and interact with their surroundings in a way that closely matches the real world. By moving beyond simple simulations, SceneSmith offers a more authentic training ground for robotics, marking a significant step forward in the field.