Tsinghua's embodied model tops global rankings! Breaks through against GPT-6 and Nvidia, without relying on plugins or extra data

What happened: Tsinghua-affiliated robotics company Robot Era said its VPP2 world action model took first place on the RoboDojo simulation benchmark, ahead of GPT-6-Astra, Physical Intelligence's pi0.5 and Nvidia GR00T-N1.7, using only standard datasets without extra data or augmentation. Key numbers: RoboDojo average success rate 32.26% and average score 39.26, both first. Why it matters: The results suggest staged training that decouples video prediction from action learning can improve operation on unfamiliar tasks, and that scale is not the only driver of physical

Tsinghua's embodied model tops global rankings, beating GPT-6 and Nvidia without plugins or extra data.

Tsinghua's embodied model tops global rankings, beating GPT-6 and Nvidia without plugins or extra data. — impact context not separately stored; see What changed.