Chinese academy of sciences model tops Meta-World benchmark with 91.9 score

China's Maxwell embodied AI model has set a new state-of-the-art on the Meta-World benchmark, scoring 91.9 on a suite of 50 everyday physical tasks. The model, developed by the Chinese Academy of Sciences' Institute of Artificial Intelligence for Industries, is the first to break the 90-point threshold on the benchmark, which tests robots on actions like grasping objects, opening doors, and using drawers.

What Maxwell achieved

The 1-billion-parameter model runs on single-card inference and completed more than 200 tasks without additional fine-tuning. On the LIBERO benchmark suite, Maxwell reached a 99.1% success rate. The previous Meta-World high was held by FabriVLA at 90 points, followed by South Korea's SUREFlow at 88.3.

FabriVLA, developed by Shenzhen-based robotics company Youibot, now sits in second place. The gap between Chinese and Western models continues to narrow. DeepCybo's PhysBrain scored 72.5 on a suite of 28 benchmarks, putting it just behind OpenAI's GPT-6 Astra at 73.3.

China's broader push into robot intelligence

Unitree Robotics released a 6-billion-parameter foundation model capable of handling more than 60 whole-body and tabletop tasks. The company plans to reinvest nearly half of its IPO proceeds into robot-intelligence research. These developments signal a concentrated effort across Chinese labs and companies to close the embodied AI gap with Western competitors.

For engineers and researchers tracking the field, the Maxwell result matters because it demonstrates that strong generalization across diverse physical tasks is achievable with relatively modest parameter counts. The model's single-card inference requirement also points toward deployment scenarios where compute constraints are tight, such as on-device robot control in construction or industrial settings.

Why this matters for science and research professionals

Maxwell's performance on Meta-World and LIBERO shows that task-general robot policies are moving from research prototypes toward systems that can handle real variability without per-task retraining. For labs and R&D teams in construction, real estate technology, and industrial automation, this reduces the engineering burden of deploying robots across changing environments. Professionals developing AI R&D Engineering Courses can draw direct lessons from the benchmark results about which architectures and training strategies produce the most transferable manipulation skills.