Jean-Jacques Dubray · Preprints.org 2026 · 2026
DOI: 10.20944/preprints202611.0557.v1
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Latent world models plan by imagining the result of many action sequences and keeping the one whose predicted features come closest to an encoded goal image. RoboJEPA, the largest such model, needs about a thousand candidates and many rounds of search per step. We measure how much guidance its goal distance gives in an obstacle-reach task, and place an open-world predictive architecture (OWPA) beside the planner: a language model writes the goal and guardrail as facts, a probe reads those facts from the planner's features to form a cost, and a monitor checks that the robot moved as predicted. The image distance guides weakly: it does not fall as the hand approaches the goal, it is dominated by a wrist camera that does not rank outcomes, and it ignores the obstacle the task is about. The language model writes the right goal and guardrail every time, and the monitor catches a robot that does not move. But a linear probe trained on five kitchens reads the facts only in the kitchens it has seen, so it cannot yet steer the planner in a new one. A declared goal states what the task depends on; reading it from a latent world model is the open problem.
Preprints.org 2026 · 0 citations
arXiv (Cornell University) 2026 · 0 citations
arXiv (Cornell University) 2026 · 0 citations
Zenodo (CERN European Organization for Nuclear Research) 2026 · 0 citations
No comments yet — start the discussion below.