Event
VPP2 World Action Model Paper Released
1 reports · 1 sources
Yanjiang Guo and other authors submitted a paper to arXiv cs.RO, proposing the world action model Video Prediction Policy 2 (VPP2). The authors state that the model, through event-level pretraining on large-scale manipulation videos, distills the video model into a single-step visual planner with a fixed prediction horizon, and uses a MoT architecture to connect an action module to learn implicit inverse dynamics, thereby achieving strong zero-shot generalization in video prediction and action generation. The authors state that after post-training on each benchmark, VPP2 achieved the highest success rates among the evaluated methods on the three benchmarks LIBERO-Pro, LIBERO-OOD, and RoboDojo. These results are as stated by the authors and await independent verification.
Report timeline
- Video Prediction Policy 2: Event-Level Video Pretraining Improves Manipulation Policy Generalization
The authors propose the world action model Video Prediction Policy 2 (VPP2), which uses event-level pretraining on large-scale manipulation video, distills the video model into a single-step visual planner, and connects an action module via a MoT architecture to learn implicit inverse dynamics.
arXiv cs.RO ·