Closed-loop validation of hazard-identification benefits

Establish through closed-loop navigation experiments whether Geo-Anchored Fine-Tuning (GAFT) reduces robot immobilization, recovery frequency, or mission time during autonomous off-road navigation.

Background

The reported experiments evaluate offline RGB hazard identification using repeated leave-one-scenario-out testing on two forest hazard morphologies. They do not test whether improved hazard predictions produce better outcomes when integrated into an active navigation policy. The authors therefore leave unresolved whether GAFT delivers operational benefits in closed-loop deployment, specifically by reducing immobilization events, the frequency of human recoveries, or total mission time.