What's this about?
Parse, generate and place — a three-stage approach to composable real-to-sim scene modelling. 78 upvotes.
What is real-to-sim and why does it need composability?
Real-to-sim builds a simulation from a real scene, so a robot can practise in a copy of where it will work. A monolithic reconstruction gives you one frozen scene; you cannot move the chair, add a second cup, or vary the lighting. Composable means the scene is built from identified, separable objects you can rearrange — which is what makes it a training environment rather than a photograph. Parse identifies what is there, generate creates the assets, place arranges them.
Why is it worth noticing?
- It is the practical counterpart to Atlas's space-time simulation claim, from the other direction.
- Composability is what turns one captured scene into thousands of training variations, which is the entire value.
- Real-to-sim is the current best answer to the robot-data problem the VLA paper described yesterday.