Native Action-Prior Learning from Videos for World Action Models
NAVA-WAM pretrains robot action policies directly from observation-only videos, then post-trains on labeled demonstrations.
NAVA-WAM introduces native action-prior learning so a robot action policy can be pretrained directly from observation-only videos, without a separate latent-action model. The first stage propagates future-video flow-matching supervision through transition-structured joint attention to train Action-DiT. The second stage post-trains on action-labeled demonstrations with joint video-action flow matching, using asymmetric attention so inference can be action-only. Experiments report gains in in-distribution and out-of-distribution settings, better action-label efficiency, and real-robot generalization.
- NAVA-WAM learns action priors directly from unlabeled videos.
- Stage one pretrains Action-DiT with future-video flow matching.
- Stage two post-trains on action-labeled robot demonstrations.
- Asymmetric attention enables efficient action-only inference.
- It reports stronger in- and out-of-distribution and real-robot results.
From the original article
1 min read at huggingface.co
World action models integrate future visual dynamics with robot action prediction, but their scalability remains limited by the need for action-annotated robot trajectories.
By Zhaochong An, Fei Zhang, Menglin Jia · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.