Physix FrontierPhysix Frontier · News Briefing Card (QbitAI · Oct 9, 2026)
Tsinghua-backed Robot Era's VPP2 Tops RoboDojo Sim Benchmark
KEY FACTS
- Robot Era's self-developed world action model VPP2 has topped the RoboDojo simulation leaderboard.
- VPP2 achieved a 32.26% overall average success rate and an average score of 39.26 points, ranking first in both.
- VPP2 ranked first in three dimensions: generalization, fine manipulation, and memory.
- VPP2 used no additional data or Agent RSI augmentation, training solely on the standard dataset.
- In zero-shot tests on the real ALOHA robot, VPP2 achieved an average success rate of 58.5%, higher than π0.5's 40%.
KEY DATA
32.26%RoboDojo overall average success rate
39.26 pointsRoboDojo overall average score
22.48%GPT-6-Astra average success rate
58.5%ALOHA real-robot zero-shot average success rate
PHYSIX OBSERVATION
VPP2 proves that you don't need to pile on data and bolt on augmentation—optimizing the training paradigm alone can take the top spot. It decouples video prediction from action learning, first letting the model understand physical change before learning actions, and its real-robot zero-shot performance leads as well. This sends a signal to the embodied AI industry: scale is not the only way forward, and there is still ample room for engineering innovation in data pipelines and training strategies.
Source: QbitAI report