Tsinghua-backed Robot Era's VPP2 Tops RoboDojo Benchmark, Beating GPT-6-Astra and Others

VPP2, the world action model independently developed by Xingdong Era, has topped the RoboDojo simulation leaderboard, with an overall average success rate of 32.26% and an overall average score of 39.26 points, ranking first in both metrics. For comparison, in the post-training evaluation of the RoboDojo simulation benchmark, GPT-6-Astra had an average success rate of 22.48% and an average score of 28.97 points, with VPP2 leading by 9.78 percentage points and 10.29 points respectively.

To understand this result, one must first understand the exam hall. RoboDojo is led by HKU MMLab and co-built by nearly 20 top academic institutions worldwide; its simulation tests include 42 dual-arm manipulation tasks covering five dimensions: generalization, precise manipulation, long-horizon tasks, memory, and open-vocabulary instruction understanding. Its question design deliberately avoids robots' comfort zones: can it still work after the environment changes, can it still grasp accurately when objects are placed crookedly, and does it still remember what happened earlier when a task is half done.

The two metrics have different emphases: success rate looks at whether the robot can complete a task fully, while average score measures how far a task progressed—even if it was not fully completed, it can still reflect staged performance. Both metrics aggregate test results across the five capability dimensions. Broken down, VPP2 ranks first in three dimensions—generalization ability, fine manipulation, and memory ability—and these three happen to correspond to several hurdles robots face when entering the real world.

It must be made clear that RoboDojo is, after all, a simulation environment. In the physical world, friction, deformation, and positional deviation of objects all introduce variables. Whether the generalization ability shown in simulation holds up on real robots is a question that must be answered separately, beyond this report card.

The team deployed VPP2 on a real ALOHA dual-arm robot to test 10 categories of zero-shot manipulation tasks including grasping, placing, stacking, folding, and pouring—that is, without additional fine-tuning for these test tasks, the robot goes straight to work. The result was an average success rate of 58.5%, higher than π0.5's 40%, and it took best-in-class results in 9 of the 10 task categories.

The significance of this real-robot result lies in the fact that it examines whether capabilities from simulation can transfer to a real physical environment. But the boundary is equally clear: the real-robot test was conducted only on the ALOHA dual-arm robot and did not cover other embodiments; the 58.5% average success rate also means a considerable proportion of tasks still went uncompleted.

Two other generalization tests provide supplementary evidence. LIBERO-Pro examines manipulation ability after changes in object positions and task requirements, where VPP2 had an overall success rate of 45.0% while other baseline models reached at most 11.0%; LIBERO-OOD examines compositional generalization, recombining familiar objects, layouts, and task goals into new tasks never seen before, where VPP2 had an overall success rate of 63.9%.

Putting these results together, the judgment that can be read is this: VPP2 shows leadership across multiple evaluation criteria, but these evaluations all come from tests developed or participated in by its own team, with no third-party independent reproduction results yet.