Maxwell smart: new Chinese embodied model tops global AI ranking on physical tasks

The Chinese Academy of Sciences' Institute of Artificial Intelligence for Industries has unveiled an embodied AI model called Maxwell, which has claimed the top spot on the Meta-World benchmark for robot physical tasks. Maxwell achieved a remarkable score of 91.9, the highest ever recorded on the simulation benchmark. Meta-World, developed by researchers from Stanford University, UC Berkeley and other institutions, evaluates robots' performance on 50 everyday physical tasks, including grasping, carrying, opening doors, and using drawers.

The list draws submissions from leading teams worldwide, with FabriVLA from China's Youibot scoring second at 90.0, followed by SureFlow from South Korea's Kyungpook National University with a score of 88.3. These Chinese large models are surpassing their American rivals, as they are now entering the real physical world and learning to act independently, rather than merely thinking and talking.

Maxwell demonstrated its capabilities by performing over 200 tasks without additional fine-tuning, such as moving objects, adjusting positions, and operating switches and containers. It could move a nut onto a target post, open cabinet doors and drawers, and grasp objects to carry them across containers. The model also showed success in LIBERO, a simulation environment emphasizing continuous operations, successfully completing tasks such as turning on a stove, placing a kettle on it, and placing a book into a storage box.

Maxwell is designed for edge deployment, processing data close to the user or data source, rather than through centralized cloud deployment. With 1 billion parameters and single-card inference, it requires relatively low hardware requirements. This achievement is not isolated, as many Chinese companies, including DeepCybo, Unitree Robotics, and Beijing-based DeepCybo, have recently made significant progress in this field.