Pushing the frontier of artificial intelligence.
Postlight builds scored engineering tasks that teach AI to design buildings, and simulation-ready scenes robots learn to work in.
Tasks AI is scored on.
An agent gets a real revision request and the drawing it applies to. It edits the drawing, and a key built from what the engineer signed scores the result.
- RL environments. Real drawing revisions become tasks: the request, the drawing before it, and a key from the drawing the engineer signed.
- Benchmarks. Tasks held out from buildings no model has seen, scored the way an engineer would sign them: against the whole model, not a screenshot.
Scenes robots train in.
A floor plan and a walk-through video become a furnished apartment in USD. Every item is static, movable or articulated, and a wheeled robot can be routed through it.
- USD scenes. Colliders, mass and moving joints on doors, drawers and appliances. The apartment opens in Isaac Sim 5.1 with no errors.
- Egocentric data. Stereo cameras worn on site record the work from the worker's point of view, aligned to the project's model, so a robot can learn a task from the person who did it.
Train on real buildings.
Labs building models for design and for the physical world: tell us what your model needs to see and do. We build the tasks, the scenes and the data. founders@postlight.ai