The Zhejiang-based company said on Wednesday that its Spirit v1.6 foundation model scored 1,924 points on the RoboArena leaderboard, ahead of Nvidia's Cosmos3-Nano-Policy at 1,881, according to the South China Morning Post.

RoboArena measures how effectively general-purpose robot AI systems convert perception into real-world actions. The benchmark was co-developed by Nvidia and researchers from U.S.'s Stanford University and the University of California, Berkeley.

DreamZero, another Nvidia project introduced in February, ranked third with a score of 1,763. Nvidia is currently the world's most valuable company, with a market capitalization exceeding $5 trillion, according to CompaniesMarketCap.

Employees of a Chinese robotics firm guide humanoid robots through task in kitchen environment. Photo from X

Spirit AI's announcement came two days after Nvidia unveiled Cosmos 3, an open foundation model for physical AI designed to help robots "think before they act."

Physical AI refers to artificial intelligence that enables machines to perceive, understand and interact with the real world. These systems are generally assessed on two core capabilities: policy models, which determine what actions a robot should take based on what it observes, and world models, which predict the outcomes of those actions.

While Spirit AI currently leads the policy category, Chinese companies also occupy the top positions in several other embodied AI benchmarks. WorldScape-0.2, developed by Chinese startup Manifold AI, ranks first on the WorldArena benchmark for world models, ahead of Nvidia's Cosmos-Predict 2.5.

Chinese robotics company AgiBot leads the perception category with its GenieEnvisioner-Sim2.0-2B, a video world simulator for robotic manipulation unveiled last week.

Meanwhile, Chinese startup DexForce tops the data engine category with DSCFuncWorld, a platform designed to optimize robot training data pipelines.

Manifold AI's WorldScape-0.2 also holds the leading position on the WorldScore benchmark, which measures a model's ability to generate interactive virtual worlds from text prompts, outperforming WonderJourney, a joint project by Stanford University and Google.