FETA: Feedback-Efficient Test-Time Adaptation for Navigation Agents
October 23, 2026
FETA improves vision-and-language navigation by optimizing how human feedback is utilized during test-time adaptation. Unlike existing methods that apply uniform trajectory-level supervision, FETA localizes informative actions to prevent error propagation and cascading failures in unseen environments.
HOW THIS AFFECTS YOU
●
builderYou can implement more efficient feedback loops for agents operating in dynamic, unseen environments.
●
researcherThis method addresses the error propagation issues inherent in unsupervised entropy minimization during navigation tasks.