Video Prediction Policy 2: Predict Better, Act Better

2026-10-07

AI Summary

This work addresses the limited video prediction accuracy of existing world action models in open environments, which leads to erroneous robotic execution. We propose VPP2-14B, a video prediction policy model that enhances representations through continued pre-training on large-scale manipulation videos and event-level annotations. A Mixture-of-Transformers (MoT) architecture is designed to effectively decouple video generation from action control by integrating single-step visual planner distillation with an implicit inverse dynamics module. The resulting model exhibits strong zero-shot generalization capabilities, improving instruction-following rates by 11.0% and surpassing baseline performance by 18.5% in success rate on real-world ALOHA tasks, while achieving leading results across multiple benchmarks.

Abstract

World action models (WAMs) have emerged as an important class of generalist robot policies, aiming to transfer video prediction priors to action learning. However, we find that existing WAMs frequently produce incorrect motion predictions in open-ended environment, leading to erroneous actions. We attribute this limitation to two factors: (1) base video models are not optimized for manipulation, and (2) naively incorporating action components into video models can substantially degrade their generalization capabilities. We introduce Video Prediction Policy 2 (VPP2), a WAM that enables strong zero-shot generalization in both video prediction and action generation. First, we curate a large-scale, diverse dataset of manipulation videos to continue pretraining the base video foundation model. We annotate video clips with detailed captions and perform \textit{event-level} video pretraining to promote generalization across open-ended manipulation tasks. Second, we post-train and distill the video model into a single-step visual planner with fixed prediction horizon. Finally, we introduce action module via a mixture-of-transformers (MoT) architecture to learn implicit inverse dynamics model. Experiments demonstrate three key results: (1) VPP2-14B outperforms Cosmos3-64B by 11.0\% points in video prediction instruction-following success rate on open-ended tasks; (2) VPP2 surpasses the strongest baseline by 18.5\% points in success rate on real-world zero-shot ALOHA manipulation tasks; and (3) following benchmark-specific post-training, VPP2 achieves the highest success rates among evaluated methods on the challenging LIBERO-Pro, LIBERO-OOD, and RoboDojo benchmarks.

Problem

• World Action Models
• Video Prediction
• Robot Manipulation
• Zero-shot Generalization
• Motion Prediction

Innovation

• World Action Models
• Event-level Pretraining
• Mixture-of-Transformers
• Zero-shot Generalization
• Visual Planner