Tsinghua embodied model tops global ranking without external modules or extra data

A Tsinghua embodied AI model has climbed to first place in a global ranking, breaking through against approaches including GPT-6 and NVIDIA, and doing so without "external" modules or extra data, according to a report by QbitAI.

The report says Robot Era's chosen route is to train video prediction and action learning in separate stages. The emphasis is not on cooking video and action "in one pot," but on decoupling the two and then re-ordering them.

Decoupling here means the model no longer learns to understand the world and to act in a single shared pipeline. Instead, the two capabilities are split and trained in stages, which is seen as letting each be learned more fully.

Re-ordering means adjusting the sequence and interaction between the two learning tasks. The report stresses that this staged, sequence-aware design is what lets the model reach leading performance without external components or additional data.

Embodied AI is one of the most watched areas in AI today, requiring a model both to understand the visual world and to output executable actions. Combining seeing and doing well has long been the field's hardest problem.

Topping the ranking without add-on modules or extra data strengthens the case for this approach — it suggests training strategy itself can matter more than stacking up modules or data.

What to watch next is whether the method reproduces on more real robot tasks, and whether the "staged, decoupled, re-ordered" idea spreads to other teams. Competition in embodied AI is shifting from who has more data to who has the smarter training method.