Reading 9 papers

World action models

In PROWorld-action models, World models, Vision-language-action models

1. Feb 2026 NVIDIA

   World Action Models Are Zero-Shot Policies (DreamZero)

   Adapts a pretrained video diffusion model to predict future video and actions jointly, and shows it works as a zero-shot robot policy.

   ConnectsThe starting point: a video model that emits actions. Fast-WAM and Faster-WAM then ask how much of the predicted future inference needs. See Fast-WAM, Faster-WAM.

2. Mar 2026

   Fast-WAM

   The main latent WAM: video prediction is kept as a training loss, but the future is skipped at inference. Simple-WAM is built on its codebase.

   ConnectsOne side of the question: predict the future in training, skip it when acting. Faster-WAM argues the other side on robustness. See Faster-WAM, World Action Models Are Zero-Shot Policies (DreamZero).

3. Aug 2026

   Faster-WAM

   Concurrent work that also argues inference-time future conditioning matters for robustness and develops an efficient way to keep it.

   ConnectsConcurrent with Fast-WAM and on the other side: keep the future at inference, made cheap. When to Trust Imagination needs this side, since it reads the predicted future. See Fast-WAM, When to Trust Imagination.

   In PROWorld models

4. Oct 2026 NVIDIA, MIT, HKU, UCSD

   Long-WAM

   Pretrains a causal video model on robot and egocentric video without action labels, then scales how many seconds of visual history the WAM sees. A co-designed runtime keeps it real-time on a Unitree G1 humanoid.

   ConnectsNAVA-WAM pretrains without action labels too, through the attention path instead. See NAVA-WAM, World Action Models Are Zero-Shot Policies (DreamZero).

   In PRORobot learning, World models, World-action models

5. Oct 2026

   ProWAM

   Predicts an ordered set of sparse visual sub-goals, indexed by task progress, and caches them from one pass of the video backbone so replanning only needs the light action denoising.

   ConnectsViGAR also separates planning from execution, with a planner of its own. See ViGAR.

   In PRORobot learning, Vision-language-action models, World models

6. Oct 2026 Meta

   NAVA-WAM

   Pretrains the action expert on video with no action labels by letting the future-video loss flow back through joint attention into the action branch.

   ConnectsLong-WAM gets action-free pretraining from a causal video model and spends it on history. See Long-WAM.

   In PRORobot learning, Vision-language-action models, World models

7. Oct 2026

   ViGAR

   Splits long tasks into a visual subgoal planner and a subgoal executor that share one pretrained world-model representation.

   ConnectsProWAM keeps a single model and caches the sub-goals instead. See ProWAM.

   In PRORobot learning, Vision-language-action models, World models

8. Oct 2026

   UniWAM

   Combines a vision-language reasoner, a video generator, and an action predictor in one Mixture-of-Transformers model, pretrained on robot data, human egocentric video, and VQA data.

   ConnectsThe all-in-one design: reasoning, video generation and action prediction in one model, where DreamZero adapts a video model alone. See World Action Models Are Zero-Shot Policies (DreamZero).

   In PROVision-language-action models, World models, World-action models

9. May 2026

   When to Trust Imagination

   Uses the WAM's own predicted future to decide when the remaining actions can still be trusted, so the robot replans less often.

   ConnectsTurns the predicted future into a replanning signal, which needs a model that still predicts it at inference, the side Faster-WAM takes. See Faster-WAM.