#China #Beijing #Robotera #VPP2 #RoboDojo

ROBOTERA's VPP2 Secures Top Spot on RoboDojo

In a remarkable feat for artificial intelligence in robotics, ROBOTERA's World Action Model (WAM), VPP2 (Video Prediction Policy 2), has topped the RoboDojo rankings. This achievement is particularly striking as it was accomplished without the use of additional data or the usual agent-based reinforcement self-improvement techniques. Developed through a collaboration with nearly twenty prestigious academic institutions, including contributions from the University of Hong Kong's MMLab, RoboDojo serves as a rigorous benchmark that assesses the capabilities of robots in mastering complex general-purpose manipulation tasks.

Exceptional Performance Metrics

VPP2's performance metrics are impressive, boasting an average success rate of 32.26% and an average score of 39.26. It excelled particularly in categories such as Generalization, Precision, and Memory, outpacing its competitors. Moreover, ROBOTERA made the VPP2 model publicly accessible, releasing it on GitHub and providing further details on their project website, thus encouraging community engagement and innovation.

The Innovation Behind VPP2

ROBOTERA's VPP2 represents a significant advancement in the field of robotics, as it enhances the predictive capabilities of robots regarding how their actions can affect their environments. Unlike traditional video generation models that primarily focus on visual content generation, VPP2 is tailored for understanding object movements and executing precise manipulation commands. This innovative approach combines video prediction with action generation, enabling robots to complete tasks more reliably across a range of objects and environments.

To tackle multifaceted tasks requiring multiple steps, VPP2 can also collaborate with Vision-Language Models (VLM), which decompose high-level instructions into smaller, actionable segments, thereby facilitating complex task execution.

Validation through Diverse Challenges

ROBOTERA's VPP2 model underwent a thorough evaluation consisting of video prediction, instruction following, and robotic manipulation challenges. On the ALOHA platform, it achieved a remarkable average success rate of 58.5% across ten task categories, surpassing leading baselines in nine of these categories. This model also recorded a 45.0% success rate on LIBERO-Pro, a benchmark focusing on robotic manipulation and generalization. With high-level task planning, the average success rates for five task groups notably increased, from 27.6% to an extraordinary 57.6%. These impressive results underscore VPP2’s ability to harmonize visual comprehension, instruction adherence, and physical action, propelling general-purpose robotics toward practical applications.

Progress from Research to Real-World Applications

ROBOTERA's triumph in embodied intelligence with VPP2 marks its fourth benchmark title in 2026, following previous successes at events such as the World Arena, Benjie's Humanoid Olympic Games, and RoboChallenge. In tandem with these accomplishments, the company has also been rolling out humanoid robots in partnership with China Post and SF Express, operational across over ten logistics centers in China. These ongoing efforts underscore ROBOTERA's commitment to advancing general-purpose robotic intelligence while seamlessly transitioning it into practical, real-world implementations.

As it builds upon these significant advancements in world action modeling and real-world deployment, ROBOTERA is dedicated to establishing general-purpose robots as dependable partners in various everyday tasks, setting new standards in the robotics landscape.