VPP2: 14B Video Prediction Policy Beats Cosmos3-64B by 11 Points on Zero-Shot Manipulation

zhenjun_zhao · x · 2026-10-09

Video Prediction Policy 2 (VPP2): Predict Better, Act Better

Problem: Existing world action models (WAMs) often produce wrong motion predictions in open-ended environments, because base video models aren't optimized for manipulation and naively adding action modules degrades generalization.
Method:
Curated a large, diverse manipulation-video dataset and performed event-level captioned video pretraining;
Post-trained and distilled the video model into a single-step visual planner with fixed horizon;
Added an action module via a mixture-of-transformers (MoT) architecture learning an implicit inverse dynamics model.
Results: VPP2-14B outperforms Cosmos3-64B by 11.0 points, achieving strong zero-shot generalization in both video prediction and action generation.

More from Embodied