Zhiyuan GO-3 Nano Preview Tops Embodied AI Benchmark RoboDojo, Outperforming GPT-6 Astra
On October 27, 2026, RoboDojo, an embodied AI benchmark led by the University of Hong Kong, released its latest ranking. Zhiyuan's GO-3 Nano Preview ranked first with an average score of 61.90 and an average success rate of 56.42%, outperforming GPT-6 Astra. It also led the sub-dimensions of fine manipulation, long-horizon execution and open semantic understanding, with scores of 78.65, 68.21 and 61.71 respectively.
The result adds a new data point to the ongoing debate over whether specialized embodied models are still necessary as general-purpose models expand their capabilities. OpenAI released GPT-6 Astra in September 2026, and the model has since been shown to understand instructions, break down complex tasks and execute steps when connected to robots. Wang Xingxing, founder of Unitree, recently demonstrated a Unitree robot using GPT-6 to draw and play Gomoku, saying the model could handle part of robot task planning without additional training. However, a multi-task evaluation by Galaxy General found that Astra performed well in navigation and object search but still showed limitations in complex contact, dexterous manipulation and highly dynamic motion.
In a research note released on October 26, 2026, CITIC Securities said that general models entering embodied task domains would become the next investment direction for the AI industry in 2027, and that general and specialized models are more likely to form a division-of-labor and collaborative landscape in the short term.
