Governed Embodied Action

Abstract

This research preview examines an independent, persistent action supervisor placed between a language-model planner and a simulated robot. The study asks whether proposed physical actions can be evaluated against structured scene constraints, released or interrupted through a separate governance layer, and recorded in an auditable decision trace. It reports bounded simulation evidence across two planner classes, independent evaluation of physical outcomes, transfer without changing the supervisor, and a documented boundary in language-only cases. The result is a research demonstration, not a safety certification or a claim of unbounded generalisation.