VLN-CE

CoFL-S: Language-conditioned flow fields for low-level robot navigation in VLN

CoFL-S is a new vision-language-action framework that predicts language-conditioned flow fields over a robot's local visible sector to generate continuous navigation trajectories. The authors address an underexplored gap in Vision-Language Navigation (VLN) by providing frame-level local supervision with sub-instruction alignment and dense flow-field targets. A new continuous-time Habitat benchmark is introduced to enable decomposition-independent closed-loop evaluation across planner frequencies. CoFL-S outperforms action-token and action-chunk baselines in simulation and demonstrates zero-shot transfer to real-world deployment.