NAVA-WAM: Direct Action-Prior Learning from Observation-Only Videos
October 1, 2026
NAVA-WAM enables world action models to scale by pretraining action policies directly from observation-only videos via flow-matching. This avoids the need for action-annotated robot trajectories or indirect representation-to-control transfer steps.
HOW THIS AFFECTS YOU
●
builderYou can train more effective robot policies using unlabelled video data.
●
researcherThis bypasses the requirement for explicit latent-action models during pretraining.