SuperCLUE has officially released its latest Embodied AI Brain benchmark, designed to evaluate the actual performance of major AI models as the "brains" for robots.
OpenAI's GPT-6 Astra took first place globally. In China, Alibaba and ZTE tied for the top spot on the domestic leaderboard.
An Embodied AI Brain is an AI system specifically designed for robots. It must not only understand visual inputs and language but also plan actions, direct robots to grasp objects and navigate, and handle various unexpected situations in real-world environments, going far beyond the capabilities of standard chat-based large language models.
This benchmark includes numerous real-world robot tests, not just virtual simulations. It evaluates multiple dimensions such as environmental perception, object recognition, long-horizon task planning, and physical world prediction, scoring mainstream domestic and international models under a unified standard.
GPT ultimately achieved a significantly higher comprehensive score, demonstrating clear advantages in long-form logical reasoning and complex task decomposition, securing the top spot globally.
Among domestic entries, Alibaba DAMO Academy and ZTE achieved identical scores, tying for first place in China. Alibaba's model excels in spatiotemporal memory, allowing robots to recall unfinished tasks after interruptions; ZTE's version focuses more on edge deployment and hardware compatibility, making it better suited for running on physical robot devices.
The leaderboard also highlights gaps: top domestic contenders still trail GPT in overall scores. When faced with highly complex, multi-step chained tasks, domestic models occasionally make judgment errors, indicating room for improvement in generalization capabilities when handling unfamiliar environments.
With more teams, including tech giants and startups, investing heavily in Embodied AI Brains, the market is flooded with diverse benchmarks. Different test sets often lead to inconsistent rankings. This benchmark stands out by incorporating extensive real-world robot testing rather than relying solely on virtual simulations, offering greater reference value.
As humanoid robots are still in their early stages, the capabilities of Embodied AI Brains will directly determine how much practical work robots can perform in the future. As companies continue to iterate on their models, the gap between domestic and international solutions is likely to narrow further.
