roboterax/video-prediction-policy-2

Official implementation of "Video Prediction Policy 2: Predict Better, Act Better". [](https://arxiv.org/abs/2610.10270) [](https://robert-gyj.github.io/video-prediction-policy-2/) [](https://modelscope.cn/models/haodong123/VPP2) [](https://huggingface.co/Haodong082399/VPP2) We introduce Video Prediction Policy 2 (VPP2), a world-action model (WAM) with strong zero-shot generalization in both video prediction and action generation. Our pipeline has three stages: (1) event-level video pretraining, (2) single-step visual planning through post-training and distillation, and (3) action learning with a mixture-of-transformers (MoT) architecture. This codebase includes: Install the CUDA/PyTorch dependencies using the installation guide, then install the dependencies from the repository root: The scripts load VPP2 directly from this checkout; installing the vpp2 package is unnecessary. For inference and data processing only, use requirements.txt. Weights are available on Hugging Face and ModelScope. The Hugging Face repository is public; ModelScope requires an account with access. The Hugging Face download instructions include the root config.json with each checkpoint. The Stage-1 and Stage-2 video checkpoints are available on…