Onto-EV-WM adds ontology layer to robots, boosting repair accuracy to 94%

Robots operating in the real world need to diagnose failures and repair their own behavior, but existing world models like EV-WM produce feature and event scores that don't say exactly which goal condition failed or how to fix it. In a new arXiv paper (2608.13901), Kailin Wang and colleagues from an academic team present Onto-EV-WM, a symbolic layer added on top of EV-WM rather than a replacement architecture. The layer uses an ontology—a formal TBox describing entity types, predicate signatures, and constraints—paired with semantic grounding that maps simulator-observed states into task-specific ABoxes. Deterministic rules retain each missing predicate and its arguments, and assign a route label for the appropriate correction mechanism. Learned or heuristic proposers stay separate from this symbolic interface, and a bounded protocol controls retries after failed verification.

In evaluation, the complete Onto-EV-WM configuration matches EV-WM's 94% success rate on the aligned PointMaze task but improves mean final-state distance from 0.90573 to 0.61177—a 32% improvement. On LIBERO-Goal, the ontology records failed task conditions as typed data, keeps predicate arguments intact, and reports 93.8% corrected-window success on seed 0 and 94.05±0.30% across four evaluation seeds. On the fixed 10,030-task LIBERO-Plus registry, Onto-EV-WM completes 8,526 tasks (85.00%), with suite-level success rates of 65.98% for LIBERO-10, 91.39% for LIBERO-Goal, and 91.38% for both LIBERO-Object and LIBERO-Spatial. The authors note these are complete configuration results under simulator protocols; they do not measure an ontology-only causal share or test real-robot recovery, leaving hardware validation for future work.