RESEARCH RECORD / StageNav
METHOD / FINDINGS
Geometrically verified task graphs and state-aligned reinforcement learning help navigation agents assess progress, recognize failure, and recover over long horizons.
StageNav-Bench parses long instructions into geometrically anchored DAGs and evaluates node-level execution and cognitive state alignment. The policy jointly predicts navigation actions and cognitive states (success, failure, or in progress).
59.1% navigation success on R2R-CE. Unseen-validation continuous navigation; this is the paper's reported evaluation result. (Table II, p. 6.)
A training-free navigation framework that combines semantic maps, graph-based target verification, and geometric frontier exploration for object, instance, and text goals.
