• Observation-Reasoning Decoupling: the agent perceives continuously while selectively compressing historical information into a memory bank.
• Adaptive Memory Refinement: memory construction is formulated as an optimization process balancing relevance, diversity, and temporal coverage.
• Corrective Finetuning: state-action pair filtering improves robustness against compounding errors.
• Strong Empirical Performance: DecoVLN achieves leading results under fair settings without global priors or multi-sensor inputs.
• Real-World Deployment: the method has been validated beyond simulation with real-world demos.

DecoVLN departs from the conventional paradigm of storing large historical observation sequences and repeatedly moving sampled frames between RAM and VRAM. Instead, it maintains a VRAM-resident memory bank populated by an adaptive refinement mechanism that preserves high-value semantic information during the observation phase and feeds it directly into the VLN model during inference.

DecoVLN achieves the best performance among prior methods under fair comparison settings, without relying on global priors or additional sensor modalities.

• Release the paper link and BibTeX entry
• Release the dataset
• Open-source training and inference code
• Release pretrained model checkpoints
• Add installation and environment setup instructions

@misc{cvpr2026decovln,
      title={DecoVLN: Decoupling Observation, Reasoning, and Correction for Vision-and-Language Navigation}, 
      author={Zihao Xin and Wentong Li and Yixuan Jiang and Bin Wang and Runming Cong and Jie Qin and Shengjun Huang},
      year={2026},
      eprint={2603.13133},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2603.13133}, 
}