Tsinghua-backed Robot Era Open-Sources VPP2: Decoupling Video Prediction and Action Learning Tops RoboDojo Benchmark

1 reports archived1 independent sourcesupdated 10/9/2026, 11:52:21

Synthesis & Latest Updates

1 Sources Cross-Validated

Tsinghua-backed robotics firm Robot Era released its world action model VPP2, which topped the global RoboDojo benchmark with a composite score of 39.26. By decoupling video prediction from action learning—training strong predictive capabilities before action control—the model achieved superior generalization, fine manipulation, and memory without relying on extra data or augmentation strategies; the code is now open-source.

LATEST/Tsinghua-backed robotics firm Robot Era released its world action model VPP2, which topped the global RoboDojo benchmark with a composite score of 39.26. By decoupling video prediction from action learning—training strong predictive capabilities before action control—the model achieved superior generalization, fine manipulation, and memory without relying on extra data or augmentation strategies; the code is now open-source.

HEAT TRENDHourly heat curve

2 fully observed hours

Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first

量子位官网T2·Oct 9, 11:5278 pts
Tsinghua-backed Robot Era Open-Sources VPP2: Decoupling Video Prediction and Action Learning Tops RoboDojo Benchmark
Original: 清华具身模型登顶全球第一！突围GPT-6、英伟达，不靠外挂和额外数据
- Robot Era's VPP2 scored 39.26 on RoboDojo, surpassing top models like GPT-6-Astra and π0.5
- Key innovation lies in decoupled staged training of video prediction and action learning rather than end-to-end mixing
- Achieved SOTA using only standard datasets without external augmentation like Agent RSI or extra data
- Model is open-sourced on GitHub, enabling community replication of its strengths in generalization, manipulation, and memory
View Evidence Original Source