Researchers have published a new paper on Hugging Face titled Native Action-Prior Learning from Videos for World Action Models, proposing a method called NAVA-WAM that aims to overcome a key scalability bottleneck in world action models: their reliance on action-annotated robot trajectories.
According to the paper, observation-only videos carry rich evidence about interaction dynamics, but existing approaches typically use them in one of two indirect ways — either to pretrain visual representations that must later be adapted for control, or to infer latent actions that are subsequently grounded to robot commands. NAVA-WAM instead pretrains the action policy natively from observation-only videos, avoiding both indirect representation-to-control transfer and a separate latent-action model.
The training procedure has two stages. In the first, the model is pretrained on observation-only videos, where future-video flow-matching supervision over visual transitions is propagated through transition-structured joint attention to optimize the Action-DiT and learn action-relevant priors. In the second stage, action-labeled demonstrations are used to post-train the Action-DiT for robot control via joint video–action flow matching, while asymmetric attention decouples the visual branch from iterative action denoising and enables efficient action-only inference.
The authors report that extensive experiments show NAVA-WAM consistently outperforms prior approaches in both in-distribution and out-of-distribution settings, while demonstrating strong action-label efficiency and effective real-robot generalization.
The paper concludes that these results establish native action-prior learning as an effective approach to directly pretrain action policies from observation-only videos, offering a scalable path beyond action-labeled robot data.