CoFL: Continuous Flow Fields for Language-Conditioned Navigation
CoFL is a comprehensive policy that connects bird's-eye view (BEV) observations with language commands to create a continuous flow field for navigation purposes. It redefines navigation as learning a field conditioned on the workspace, acquiring local motion vectors at various BEV points and transforming each scene-instruction annotation into detailed spatial control supervision. By numerically integrating the predicted field, trajectories can be generated from any starting point, facilitating real-time rollout and closed-loop recovery. For extensive training and evaluation, a dataset comprising more than 500,000 BEV image-instruction pairs has been assembled.
Key facts
- CoFL is an end-to-end policy for language-conditioned navigation
- It uses bird's-eye view (BEV) observations and language instructions
- Reformulates navigation as workspace-conditioned field learning
- Learns local motion vectors at arbitrary BEV locations
- Trajectories generated by numerical integration of the predicted field
- Enables real-time rollout and closed-loop recovery
- Dataset of over 500k BEV image-instruction pairs built
- Published on arXiv with ID 2603.02854
Entities
Institutions
- arXiv