A Chinese AI model has moved to the top of a key robot-task benchmark, giving the field a new result to watch. Maxwell scored 91.9 on Meta-World, placing it ahead of FabriVLA and SUREFlow in a ranking focused on robot physical tasks.
The result, published on 1 Oct 2026, puts Maxwell at the center of a clear performance race in embodied AI. Developed by the Chinese Academy of Sciences’ Institute of Artificial Intelligence for Industries, Maxwell now holds the highest score listed on the Meta-World benchmark.
Maxwell Claims the Top Spot
Maxwell’s score reached 91.9, while second-placed FabriVLA scored 90. SUREFlow took third place with 88.3, creating a visible gap between the three leading models.
- Maxwell: 91.9
- FabriVLA: 90
- SUREFlow: 88.3
The difference between Maxwell and FabriVLA is 1.9 points. Maxwell’s lead over SUREFlow is 3.6 points, giving the top model a clear advantage across the listed results.
That ranking matters because Meta-World focuses on robot physical tasks rather than a purely digital score. The benchmark gives researchers a shared way to compare AI models connected to physical action, and Maxwell’s result places its development team at the front of this specific evaluation.
A score alone does not describe every part of a model, but the result delivers one direct message: Maxwell achieved the strongest performance among the models named in this benchmark ranking.
What Meta-World Measures
Meta-World is a benchmark for robot physical tasks. Researchers from Stanford University, UC Berkeley, and other institutions set it up to evaluate this area of AI, creating a common point of comparison for models that operate around physical tasks.
That focus makes the benchmark different from a result based only on written answers or generated content. The listed scores connect each model to robot physical tasks, placing the emphasis on embodied AI and the ability to perform within the benchmark’s evaluation.
For Maxwell, the 91.9 score is therefore more than a standalone number. It is a position on a shared ranking built around physical robot tasks, where the model finished ahead of two other named systems.
The benchmark’s structure also makes the result easy to follow. Maxwell leads with 91.9, FabriVLA follows with 90, and SUREFlow sits at 88.3. Those numbers create a simple snapshot of how the three models compare in the reported Meta-World results.
A Three-Way View of Embodied AI Progress
The result also brings together AI work from different research groups and locations. Maxwell comes from the Chinese Academy of Sciences’ Institute of Artificial Intelligence for Industries, FabriVLA was developed by Shenzhen-based Youibot, and SUREFlow was developed by researchers at Kyungpook National University in South Korea.
That mix gives the ranking a broad research profile. The top three models are tied to organizations in China and South Korea, while the benchmark itself was established by researchers from Stanford University, UC Berkeley, and other institutions.
FabriVLA remains close behind Maxwell with a score of 90, making it the nearest listed result to the leader. SUREFlow’s 88.3 score places it third, but it remains part of the same leading group shown in the benchmark results.
The numbers leave the field with a straightforward target. Maxwell has set the current high score at 91.9, while FabriVLA and SUREFlow provide the next comparison points at 90 and 88.3.
As embodied AI develops, benchmarks such as Meta-World will keep giving researchers a way to track these results. For now, Maxwell owns the headline number: 91.9, the highest score reported on the robot physical task benchmark.
Based on
