ACE ROBOTICS launches Kairos 3.1 world model with 125ms inference

ACE ROBOTICS is betting that a unified world model — not a patchwork of specialized AI modules — will unlock commercial-scale embodied intelligence across retail, hospitality and outdoor service.

SHANGHAI — ACE ROBOTICS unveiled Kairos 3.1, an action-oriented world model that integrates generative, physical and cognitive intelligence within a single architecture, alongside three commercial solutions at the 2026 World Artificial Intelligence Conference. The company's 8-billion-parameter model achieved an inference latency of 125 milliseconds on Nvidia's Jetson Thor platform at BF16 precision, according to internal testing.

"The digital world, a model error may result in a flawed image or paragraph. In the physical world, an incorrect action can have real consequences," Wang Xiaogang, Chairman of ACE ROBOTICS, said. "Kairos 3.1 is built around a first-principles approach to embodied world models, helping robots act more reliably in complex and uncertain environments."

The model's spatial understanding core, ACE-BRAIN-0.5, achieved state-of-the-art results across 12 public evaluations covering spatial understanding, navigation, manipulation and task-progress assessment among publicly reported models as of July 2026, the company said. Kairos 3.1 uses a hybrid Transformer architecture with a shared mixed-attention mechanism that maps visual observations, language instructions, force and tactile signals, and policy trajectories into a unified latent space. This enables an "understand, reason, execute and reflect" loop — the model can break down long-horizon tasks, simulate physical outcomes, rank candidate actions, execute the selected strategy and evaluate the result for adjustment. In one test, a robot changed from a three-finger to a four-finger grasp after an unsuccessful attempt and completed the task.

Data infrastructure underpins the commercial push

ACE ROBOTICS introduced the "Information-Density Law" for embodied data, classifying interaction data across five levels from L1 to L5. At L5, data incorporates three-dimensional force and tactile signals, failure-recovery trajectories and open-environment variables — what the company says supports stronger generalization and continuous learning.

Ambient Capture Engine 2.0, the company's human-centric data capture system, includes the ACE Sense Glove with 0.01-newton sensitivity and joint-angle error below 2 degrees, synchronizing data from more than 20 heterogeneous sensors with timing error under 1 millisecond. Its ACE-ViDiHand generative 4D hand-motion-capture framework led three public benchmarks — ARCTIC, HOT3D and HOI4D — as of July 2026, recording frame-level accuracy of 0.997 and up to a 4.8-fold improvement in motion smoothness. The company open-sourced ACE Ego Matrix, a data standardization framework that ranked first on the RoboCase and RoboTwin leaderboards, and released ACE-Data-0, an L5 household-interaction dataset.

Three commercial solutions target specific verticals

Xiaoman, the instant-retail fulfillment solution paired with the W1 robot, has a robot-to-payload weight ratio below 2:1 and force-control precision within 1 newton, with a minimum aisle width of 75 centimeters. ACE ROBOTICS said it has deployed the system with Sense MartGo, Kuaikeda and PetroChina convenience stores and plans to reach 1,000 retail locations within a year and approximately 10,000 stores within two years.

Xiaoxin addresses hotel-laundry operations, where the company estimates roughly 80 percent of demand occurs at night. The solution automates collection, loading, unloading, washing, sorting and folding. Xiaotu, built on a "one brain, many bodies" architecture, applies a general-purpose intelligent brain across multiple quadruped robot platforms for outdoor environments and has been deployed at cultural-tourism events including the Begonia Flower Festival in Tianjin.

ACE ROBOTICS also announced strategic collaborations with the Caohejing Development Zone for an embodied AI innovation platform, and with PetroChina, Kuaikeda and SenseTime Shanhui for retail and fulfillment. A World Model Cloud Ecosystem initiative with Baidu AI Cloud, Alibaba Cloud, Huawei Cloud, Tencent Cloud and SenseCore AI Cloud was launched to support world-model computing requirements. The PHYSICAL IQ benchmark, jointly initiated by the Shanghai Artificial Intelligence Association, the Shenzhen Loop Area Institute and the East China branch of the China Academy of Information and Communications Technology with more than 20 universities and industry partners, aims to standardize evaluation of embodied physical intelligence.

ACE ROBOTICS is positioning Kairos 3.1 against a fragmented competitive field where most embodied AI efforts combine separately developed capabilities rather than using a unified architecture. The company's ability to scale from 1,000 to 10,000 retail deployments over two years will test whether its unified-model thesis translates into commercial traction. For investors, the key metric is deployment velocity: ACE ROBOTICS has named customers and a clear timeline, but the embodied AI market remains nascent, with no dominant player yet established. The company's open-source strategy for data frameworks and datasets could accelerate ecosystem adoption, though it also risks commoditizing its data advantage.

This article is for informational purposes only and does not constitute investment advice.