Xingdong Jiyuan’s VPP2 tops RoboDojo, leads in simulation and real-robot tests

Xingdong Jiyuan’s Video Prediction Policy 2, or VPP2, has taken the top spot on the RoboDojo simulation benchmark, posting a 32.26% average success rate and a 39.26 average score. The article says those results put it ahead of GPT-6-Astra, Physical Intelligence’s π0.5 and Nvidia’s GR00T-N1.7. VPP2 also ranked first in generalization, precise manipulation and memory, three capabilities the report frames as critical for robots operating outside tightly controlled settings. The piece goes beyond leaderboard numbers. It says VPP2 was deployed on a real ALOHA bimanual robot and tested on 10 zero-shot manipulation tasks, including grasping, placing, stacking, folding and pouring. There, the model reached a 58.5% average success rate, compared with 40% for π0.5, and recorded the best result in nine of the 10 task categories. According to the report, VPP2 did not rely on extra data or enhancement methods such as Agent RSI. Its core approach was to separate video prediction from action learning and train them in stages. The article also cites additional results on LIBERO-Pro, LIBERO-OOD and long-horizon RoboDojo tasks with a VLM planner, while noting that broader, long-term deployment in real-world settings still needs to be validated.