Event

MOON: Adding a Deviation Recovery Module to Frozen VLN Policies

1 reports · 1 sources

Jingmeng Zhou and other authors propose the MOON framework, which adds two lightweight modules to a frozen vision-and-language navigation (VLN) policy, introducing fewer than 1% additional parameters relative to the original policy. On R2R-CE and RxR-CE, the navigation success rate of the frozen base improves by 5.2 and 6.0 percentage points, respectively. The two modules have clear division of labor: a deviation detector predicts whether the agent has deviated from the route, and once it determines that the agent is off course, it issues an alert and clears the history; the recovery policy is trained with GRPO from the alert state to learn how to continue moving forward, converting more of the successes that originally could be found only through multiple sampling into successes in a single greedy rollout. The authors state that without retraining, MOON also shows improvements in zero-shot object-goal navigation on HM3D-OVON, and the code has been made public.