Newest paper · November 2026
Detection and correction are separate problems
An open-world predictive architecture (OWPA) beside an unmodified vision-language-action model, tested with two open VLAs on LIBERO, in nominal scenes and in scenes changed during the episode.
- 1 · WRITE
A language model writes a declared world model of the task as readable rules.
- 2 · PREDICT
A deterministic engine runs it to plan, and to predict the facts each step should produce.
- 3 · MONITOR
After every action chunk it compares prediction and scene, re-instructs a stalled policy and repairs rules that prove wrong.
Preprints.org
Article
An open-world predictive architecture beside a vision-language-action model: declared rules, prediction and repair for robot manipulation
A declared world model notices most failures early and names the fact that did not change. Better instructions do not fix them; only a small fine-tune on the changed scenes moved success.
- The declared model is a reliable observer. It notices most failures early, names the fact that did not change, and costs nothing when the scene is as trained.
- Correction through language does not follow. Better instructions, a repaired or perfect model, or a correct plan leave success unchanged, because the instruction selects among learned behaviours and does not compose new ones.
- A small fine-tune is the only thing that moved success. It uses demonstrations of the changed scenes, which the declared model helps to find.
open-world predictive architecturevision-language-action modelsworld modelsdeclared world modelrobot manipulationLIBEROlarge language modelsruntime verification
Read it on Preprints.org ↗