Liangjiang company tops RoboDojo benchmark for embodied AI

Overview of the DM0.5 architecture. [Photo provided to english.liangjiang.gov.cn]

RoboDojo, a leading benchmark for evaluating embodied AI systems, has released its latest rankings, with DM0.5, a general-purpose embodied foundation model independently developed by Dexmal (Chongqing) Intelligent Technology Co, a company based in Liangjiang New Area, taking the top spot overall.

RoboDojo is a comprehensive evaluation framework for embodied intelligence, developed by the original team behind the RoboTwin benchmark series and jointly launched by the University of Hong Kong's Multimedia Lab, the University of California, Berkeley, Tsinghua University, and nearly 20 other leading academic institutions worldwide.

In the latest rankings, DM0.5 placed first with an overall score of 24.90 and an average success rate of 19.34 percent.

For pretraining, DM0.5 is equipped with a 4-billion-parameter "smart brain" responsible for visual perception, understanding instructions, and decision-making, alongside a 680-million-parameter "action coach" focused on translating decisions into physical actions. Together, they form a vision-language-action (VLA) architecture that has undergone systematic upgrades in areas including historical context, embodied reasoning, action supervision, and data quality.

DM0.5 is now fully open source, with its model weights, training framework, and complete downstream workflow available through GitHub and Hugging Face, allowing developers to download, reproduce, and further develop the model directly.