The VPP2 world action model, independently developed by STAR ERA, a Chinese enterprise specializing in embodied AI, has secured the number one position on the RoboDojo simulation leaderboard—a highly competitive benchmark in the field of embodied AI. It leads globally in both comprehensive average success rate and average score. VPP2 also demonstrates exceptional performance across three critical dimensions: generalization ability, fine manipulation, and memory capacity. Remarkably, without utilizing extra data or external enhancement methods, VPP2 outperforms leading embodied models worldwide, including GPT-6-Astra and NVIDIA GR00T-N1.7, relying solely on standard datasets.

VPP2 adopts a three-stage training approach that separates video prediction from action learning, progressively enhancing generalization in both video prediction and action execution. During zero-shot operation tests on the real ALOHA dual-arm robot, VPP2 achieved an impressive average success rate of 58.5%, outperforming the leading model π0.5, which scored 40%, and significantly surpassing baseline models across multiple generalization tests.

The technical strategy of integrating general cognitive models with world action models, pioneered by STAR ERA, is set to drive the evolution of new paradigms in physical AI. Currently, relevant technologies have been implemented in collaborative logistics scenarios, and the VPP2 model has been made openly accessible for further research and development.