Ruixiao Yang and co‑authors released a paper on Oct 15, 2026 describing a training‑free framework that builds reusable perception and planning programs from a single practice scene. The system uses a vision‑language model’s rendering ability to check its own actions and refine the program without any solution labels. After refinement, the frozen program solves new instances without further calls to the large model. Tests on fifteen 2D and 3D domains showed an average planning success rate of 99.3%.
Why it matters
It lets developers create visual planning systems with far lower compute cost, making AI‑driven automation more practical for robotics and simulation.
