# World-Action Models

World-action pretraining now learns from task-paired human demonstrations: the model sees one demonstration's video and actions and predicts another demonstration's actions and future frames, separating scene appearance from task interaction. Previously, within-video next-frame prediction let context and target share scene, viewpoint and collection source, so the model could rely on appearance cues instead of transferable interactions. On MimicDroid, cross-video pretraining raises Cosmos Policy's average success from 40.9% to 68.3% and the unseen-layout level from 18.3% to 40.8%, without matched robot demonstrations.

World-action pretraining now learns from task-paired human demonstrations: the model sees one demonstration's video and actions and predicts another demonstration's actions and future frames, separating scene appearance from task interaction. Previously, within-video next-frame prediction let context and target share scene, viewpoint and collection source, so the model could rely on appearance cues instead of transferable interactions. On MimicDroid, cross-video pretraining raises Cosmos Policy's average success from 40.9% to 68.3% and the unseen-layout level from 18.3% to 40.8%, without matched robot demonstrations.

The work introduces AnyStep World Action Model, a framework that performs budget-aligned flow-map distillation from frozen-teacher trajectory intervals and trains a lightweight risk-benefit scheduler to predict teacher-trajectory difficulty and budget-specific student fidelity from a single one-step preview, selecting the smallest denoising budget that meets a fidelity requirement; on RoboTwin 2.0 it reduces average denoising steps by 60.2%, 49.8%, and 85.28% on Motus, FastWAM, and LingBotVA while keeping average success within 0.24 percentage points of full-budget baselines, raises one-step success by 7.07, 12.08, and 8.94 percentage points respectively, and achieves 1.67-6.14x per-call speedups on six real-world manipulation tasks.

The work introduces AnyStep World Action Model, a framework that performs budget-aligned flow-map distillation from frozen-teacher trajectory intervals and trains a lightweight risk-benefit scheduler to predict teacher-trajectory difficulty and budget-specific student fidelity from a single one-step preview, selecting the smallest denoising budget that meets a fidelity requirement; on RoboTwin 2.0 it reduces average denoising steps by 60.2%, 49.8%, and 85.28% on Motus, FastWAM, and LingBotVA while keeping average success within 0.24 percentage points of full-budget baselines, raises one-step success by 7.07, 12.08, and 8.94 percentage points respectively, and achieves 1.67-6.14x per-call speedups on six real-world manipulation tasks.