SC²-WM: Self-correcting world model boosts VLN navigation robustness

Vision-and-Language Navigation in Continuous Environments (VLN-CE) demands fine-grained decisions under partial observability, but most existing agents execute open-loop and can't detect or fix internal state drift during inference. The new SC²-WM framework from Yao et al. closes that gap by introducing internal feedback for closed-loop decision making. It derives feedback from world-model foresight to perform state-level plan refinement before action execution, effectively catching errors before they impact navigation.

To handle harder scenarios, SC²-WM adds conditional world-aware adaptation, which selectively updates the world model at test time when feedback indicates model capacity is insufficient. This model-level correction keeps the agent robust without full retraining. On standard VLN-CE benchmarks, the method demonstrates improved navigation robustness and generalization. Accepted at ICML 2026, the paper includes publicly available code, making it immediately useful for researchers working on embodied AI and real-world robot navigation.