Turning Adaptation into Assets: Cross-Domain Bridging for Online Vision-Language Navigation

A new framework called IDEA aims to solve a persistent failure mode in vision-and-language navigation agents: catastrophic forgetting when they encounter unfamiliar environments. The approach, detailed in a paper posted to arXiv, treats test-time adaptation not as a series of isolated fixes but as the accumulation of reusable knowledge assets [1][2]. The Inter-Domain BridgE with Historical Assets (IDEA) framework is designed for Vision-and-Language Navigation (VLN) agents operating under non-stationary environment shifts [2]. Standard test-time adaptation methods often treat each online update as a transient, isolated event, which the paper's authors argue leads to negative transfer and the erosion of previously learned skills [1][2]. IDEA instead transforms the adaptation process into the accumulation and composition of assets [1]. The system introduces soft prompts that are optimized using a Fisher-guided weighting scheme to capture transferable knowledge [2]. These optimized prompts are then augmented with domain coordinates to build a dynamic asset library [1]. Leveraging this library, IDEA constructs a cross-domain bridge by projecting the target domain onto the convex hull of historical knowledge [2]. The paper describes this as a complementary loop: the evolving library underpins bridge construction, while the bridge provides a superior initialization that accelerates asset optimization [1][2]. The researchers tested the framework across three standard benchmarks: REVERIE, R2R, and R2R-CE [1]. Results demonstrated consistent superiority over existing methods, showcasing the ability to enable training-free adaptation through asset sharing [2]. The work was submitted to arXiv on 22 May 2026 in the field of robotics [1].