A new artificial intelligence (AI) model developed by China's Academy of Sciences has set a new standard for robot physical task performance, ranking first on a leading global benchmark. The model, named Maxwell and created by the Institute of Artificial Intelligence for Industries, achieved a top score of 91.9 on the Meta-World benchmark, which measures robot proficiency across 50 everyday physical tasks.
Meta-World, designed by researchers from Stanford University, the University of California Berkeley, and other institutions, evaluates robotic ability in activities such as grasping, carrying, opening doors, and using drawers. The benchmark attracts submissions from numerous major teams worldwide, including Google DeepMind, Carnegie Mellon University, MIT, the University of Cambridge, and several prominent Chinese companies such as Alibaba and Meituan.
Maxwell’s score surpasses all prior results recorded on the simulation platform. The second-place entry, FabriVLA, was developed by Shenzhen-based Youibot and earned a score of 90, while the third-ranked model, SUREFlow from South Korea’s Kyungpook National University, scored 88.3.
The success of Maxwell highlights ongoing advancements in embodied AI, where models go beyond cognition and language to interact physically with the environment. Unlike traditional robotic systems that often operate in fixed industrial settings with preset commands, embodied AI aims to enable autonomous robots to understand and adapt to new physical tasks in real-world conditions.
Completing tasks in the Meta-World environment requires understanding spatial relationships between objects and targets, executing sequences such as approaching, grasping, and moving, and making adjustments based on contact feedback. According to the Chinese Academy of Sciences, Maxwell demonstrated the ability to perform more than 200 tasks without needing additional fine-tuning. Its capabilities include moving objects to precise locations, manipulating handles to open cabinets and drawers, and carrying items across containers in demonstration tests.
The team also evaluated Maxwell in LIBERO, a simulation environment focused on continuous, multi-step operations. Maxwell exhibited proficiency in linking various actions across scenarios, such as turning on a stove and placing a kettle on it, putting cups into microwaves, organizing books into storage boxes, and storing chocolate in drawers. The model achieved a 99.1 percent success rate in this setting.
Maxwell is engineered for edge deployment, meaning it can process data near the user or device instead of relying on centralized cloud computing, which can enhance responsiveness and privacy.
Alongside Maxwell, other Chinese firms have reported breakthroughs in embodied AI. Beijing-based DeepCybo recently introduced PhysBrain, a physical foundation model tested against 28 benchmarks. PhysBrain scored an overall 72.5, leading 14 of the assessments and placing second in 10 others. It was narrowly surpassed by OpenAI’s closed-source GPT-6 Astra model, which outperformed it by a small margin with a score of 73.3.
These developments signal a growing global competition in embodied AI, with Chinese institutions increasingly leading in enabling robots to navigate and manipulate the physical world autonomously.