How can robots remember lessons from previous navigation tasks?

Vision-language navigation agents struggle when deployed in shifting real-world environments and forget earlier lessons through catastrophic forgetting. IDEA solves this by building a library of reusable "assets" (optimized soft prompts tied to domain coordinates) and connecting new environments to this historical knowledge via convex hull projection. The system runs entirely at test time without retraining, showing consistent gains across REVERIE, R2R, and R2R-CE benchmarks.