#Qwen2.5-VL (1)

MemoryVLN: From Ego-Aware Memory Representation to Cross-Context Memory Reuse in Vision-Language Navigation

MemoryVLN compresses long navigation histories with forgetting-curve sampling, stamps each memory with 2D appearance, frozen VGGT 3D geometry and 6D ego pose, and manages in-episode plus context-level banks with warm start and confidence-gated write-back; Top-K retrieval then residually enhances the current frame, improving success rate by about 2% on R2R-CE and 6.5% on RxR-CE over the strongest data-comparable baselines.