October 1, 2026
Maxwell leads robotic arm benchmark with a 91.9% task success rate in Meta-World
The robot control model took first place on Evo-Studio based on results submitted by its developers.
As of Oct 1, Maxwell leads the Meta-World leaderboard with a score of 91.9%. Here, the robot operates in a simulation rather than a physical room.
What was tested. Meta-World tests object manipulation: the robot grasps and moves objects, opens doors and drawers. The test includes 50 tasks.
On the Evo-Studio leaderboard as of Oct 1, Maxwell ranks ahead of FabriVLA and SUREFlow. All three results were submitted by the developers.
| Model | Meta-World success rate | | --- | --- | | Maxwell | 91.9% | | FabriVLA | 90.0% | | SUREFlow | 88.3% |
On Sep 14, the Maxwell team reported testing on 50 MT50 tasks: the model successfully completed 2,305 out of 2,500 trajectories. In the Hard category, it completed 287 out of 300, or 95.67%.
Beyond Meta-World. In LIBERO, the team tested the model on 4 suites comprising 40 tasks and 1,693 trajectories. Success rates for individual tasks reached 99.1%.
According to a Sep 14 statement from Wu Bing and Wang Shaojiang's team, Maxwell performs more than 200 tasks. The model performs the demonstrated operations without additional fine-tuning.
Primary sources: [article about Maxwell](https://www.reddit.com/r/ArtificialInteligence/comments/1wusp1h/maxwell_smart_new_chinese_embodied_model_tops), [project page linked by the developers](https://wwb00l.github.io/maxwell/).
Maxwell has 1 billion parameters, can run on a single GPU and is designed to operate on edge devices.