CoFL: Continuous Flow Fields for Language-Conditioned Navigation

CoFL is an end-to-end policy that maps a bird’s-eye-view (BEV) observation and a language instruction to a continuous flow field for navigation, reformulating navigation as workspace-conditioned field learning rather than start-conditioned trajectory prediction. Trajectories are generated from any start point by numerically integrating the predicted field, enabling simple real-time rollout and closed-loop recovery. We build a dataset of over 500k BEV image–instruction pairs, each procedurally annotated with a flow field and trajectory derived from semantic maps built on Matterport3D and ScanNet. CoFL significantly outperforms modular VLM-based planners and trajectory generation policies, and is deployed zero-shot in real-world experiments.

arXiv preprint (under review)