FlowDec: Temporal Conditional Flow Decorruptor for Robust Continuous Vision-Language Navigation

As Large Models (LMs) become the standard for Vision-and-Language Navigation in Continuous Environments (VLN-CE), their performance remains highly sensitive to real-world visual corruptions such as motion blur, low-light conditions, and sensor noise. The authors investigate how to improve the robustness of these agents to unpredictable visual degradation without requiring retraining of the navigation backbone or sacrificing the real-time latency requirements of robotic deployment.

The authors introduce FlowDec (Temporal Conditional Flow Decorruptor), a framework that treats image restoration as a conditional flow-matching problem. FlowDec consists of two primary innovations:

1. Hybrid Temporal Conditioning: A training strategy that integrates historical frames and action labels to align the generative flow path, ensuring that restored images are temporally consistent with previous observations.
2. Action-Centroid Guided Filtering: An inference-time mechanism that uses the agent's predicted atomic actions (e.g., forward, turn left) to model expected latent differentials. This allows the system to dynamically assess the quality of restored frames and selectively fuse single-frame and temporal-based reconstructions to avoid error accumulation.

FlowDec significantly outperforms state-of-the-art decorruption methods in both navigation accuracy and generation latency. On the R2R-CE and RxR-CE benchmarks, FlowDec improved relative navigation success by 25.33% and 9.38%, respectively, across various corruption types. Furthermore, by utilizing flow matching instead of iterative diffusion sampling, the framework achieves 3×–8× faster inference, making it suitable for online robotic applications.