Chinese embodied AI model leads Meta-World robot task benchmark
Maxwell, an embodied artificial intelligence model developed by the Chinese Academy of Sciences, has recorded the highest score on the Meta-World benchmark for simulated robot tasks. The model scored 91.9 and was reported to complete more than 200 tasks without additional fine-tuning, while also achieving a 99.1 per cent success rate in another simulation.
Maxwell, an embodied artificial intelligence model developed by the Chinese Academy of Sciences, has taken first place on the Meta-World benchmark for simulated robot tasks. The model, created by the academy’s Institute of Artificial Intelligence for Industries, scored 91.9, the highest result recorded on the benchmark. Shenzhen-based Youibot’s FabriVLA ranked second with 90, while South Korea’s SUREFlow scored 88.3.
Meta-World evaluates robots across 50 everyday physical tasks, including grasping objects, carrying them, opening doors and using drawers. Completing the tasks requires a model to understand spatial relationships, carry out sequences of actions and adjust its movements after contact. According to a release from the institute, Maxwell performed more than 200 tasks without additional fine-tuning.
Demonstrations included moving a nut onto a target post, opening cabinets and drawers with a handle, operating switches and carrying objects between containers. The team also tested Maxwell in LIBERO, a simulation environment focused on continuous operations. The model reportedly connected several actions in scenarios such as turning on a stove and placing a kettle on it, putting a cup in a microwave and storing objects in drawers and boxes.
Its reported success rate was 99.1 per cent. Maxwell is designed for edge deployment, meaning processing can take place close to the user or data source rather than through a central cloud system. It has 1 billion parameters and supports inference on a single card, which the institute said gives it relatively low hardware requirements.
The result comes amid other advances by Chinese robotics and AI companies. Beijing-based DeepCybo’s PhysBrain scored 72.5 across 28 benchmarks, leading 14 open-source evaluations. Unitree Robotics has also open-sourced a 6-billion-parameter model for more than 60 mobile and tabletop tasks.
The supplied material describes simulation results and demonstrations. It does not establish how Maxwell performs on physical robots in uncontrolled real-world settings.
This independently written report is based on information supplied by the named publisher. Vertrix News has not independently verified the source report.